Rally 0.9.3¶

You want to benchmark Elasticsearch? Then Rally is for you. It can help you with the following tasks:

Setup and teardown of an Elasticsearch cluster for benchmarking
Management of benchmark data and specifications even across Elasticsearch versions
Running benchmarks and recording results
Finding performance problems by attaching so-called telemetry devices
Comparing performance results

We have also put considerable effort in Rally to ensure that benchmarking data are reproducible.

In general, Rally works with all versions of Elasticsearch starting from 1.x. Benchmarking with plugins and benchmarking source builds will only work from Elasticsearch 5.0 onwards.

Getting Help or Contributing to Rally¶

Use our Discuss forum to provide feedback or ask questions about Rally. Please see our contribution guide on guidelines for contributors.

Source Code¶

Rally’s source code is available on Github. You can also check the changelog and the roadmap there.

Quickstart¶

Install¶

Install Python 3.4+ including pip3, JDK 8 or 9 and git 1.9+. Then run the following command, optionally prefixed by sudo if necessary:

pip3 install esrally

If you have any trouble or need more detailed instructions, please look in the detailed installation guide.

Configure¶

Just invoke esrally configure.

For more detailed instructions and a detailed walkthrough see the configuration guide.

Run your first race¶

Now we’re ready to run our first race:

esrally --distribution-version=6.0.0

This will download Elasticsearch 6.0.0 and run Rally’s default track - the geonames track - against it. After the race, a summary report is written to the command line::

------------------------------------------------------
    _______             __   _____
   / ____(_)___  ____ _/ /  / ___/_________  ________
  / /_  / / __ \/ __ `/ /   \__ \/ ___/ __ \/ ___/ _ \
 / __/ / / / / / /_/ / /   ___/ / /__/ /_/ / /  /  __/
/_/   /_/_/ /_/\__,_/_/   /____/\___/\____/_/   \___/
------------------------------------------------------

|                         Metric |                 Task |     Value |   Unit |
|-------------------------------:|---------------------:|----------:|-------:|
|                  Indexing time |                      |   28.0997 |    min |
|                     Merge time |                      |   6.84378 |    min |
|                   Refresh time |                      |   3.06045 |    min |
|                     Flush time |                      |  0.106517 |    min |
|            Merge throttle time |                      |   1.28193 |    min |
|               Median CPU usage |                      |     471.6 |      % |
|             Total Young Gen GC |                      |    16.237 |      s |
|               Total Old Gen GC |                      |     1.796 |      s |
|                     Index size |                      |   2.60124 |     GB |
|                Totally written |                      |   11.8144 |     GB |
|         Heap used for segments |                      |   14.7326 |     MB |
|       Heap used for doc values |                      |  0.115917 |     MB |
|            Heap used for terms |                      |   13.3203 |     MB |
|            Heap used for norms |                      | 0.0734253 |     MB |
|           Heap used for points |                      |    0.5793 |     MB |
|    Heap used for stored fields |                      |  0.643608 |     MB |
|                  Segment count |                      |        97 |        |
|                 Min Throughput |         index-append |   31925.2 | docs/s |
|              Median Throughput |         index-append |   39137.5 | docs/s |
|                 Max Throughput |         index-append |   39633.6 | docs/s |
|      50.0th percentile latency |         index-append |   872.513 |     ms |
|      90.0th percentile latency |         index-append |   1457.13 |     ms |
|      99.0th percentile latency |         index-append |   1874.89 |     ms |
|       100th percentile latency |         index-append |   2711.71 |     ms |
| 50.0th percentile service time |         index-append |   872.513 |     ms |
| 90.0th percentile service time |         index-append |   1457.13 |     ms |
| 99.0th percentile service time |         index-append |   1874.89 |     ms |
|  100th percentile service time |         index-append |   2711.71 |     ms |
|                           ...  |                  ... |       ... |    ... |
|                           ...  |                  ... |       ... |    ... |
|                 Min Throughput |     painless_dynamic |   2.53292 |  ops/s |
|              Median Throughput |     painless_dynamic |   2.53813 |  ops/s |
|                 Max Throughput |     painless_dynamic |   2.54401 |  ops/s |
|      50.0th percentile latency |     painless_dynamic |    172208 |     ms |
|      90.0th percentile latency |     painless_dynamic |    310401 |     ms |
|      99.0th percentile latency |     painless_dynamic |    341341 |     ms |
|      99.9th percentile latency |     painless_dynamic |    344404 |     ms |
|       100th percentile latency |     painless_dynamic |    344754 |     ms |
| 50.0th percentile service time |     painless_dynamic |    393.02 |     ms |
| 90.0th percentile service time |     painless_dynamic |   407.579 |     ms |
| 99.0th percentile service time |     painless_dynamic |   430.806 |     ms |
| 99.9th percentile service time |     painless_dynamic |   457.352 |     ms |
|  100th percentile service time |     painless_dynamic |   459.474 |     ms |

----------------------------------
[INFO] SUCCESS (took 2634 seconds)
----------------------------------

Next steps¶

Now you can check how to run benchmarks, get a better understanding how to interpret the numbers in the summary report or start to create your own tracks. Be sure to check also some tips and tricks to help you understand how to solve specific problems in Rally.

Also run esrally --help to see what options are available and keep the command line reference handy for more detailed explanations of each option.

Installation¶

This is the detailed installation guide for Rally. If you are in a hurry you can check the quickstart guide.

Hardware Requirements¶

Use an SSD on the load generator machine. If you run bulk-indexing benchmarks, Rally will read one or more data files from disk. Usually, you will configure multiple clients and each client reads a portion of the data file. To the disk this appears as a random access pattern where spinning disks perform poorly. To avoid an accidental bottleneck on client-side you should therefore use an SSD on each load generator machine.

Prerequisites¶

Rally does not support Windows and is only actively tested on MacOS and Linux. Before installing Rally, please ensure that the following packages are installed.

Python¶

Python 3.4 or better available as python3 on the path. Verify with: python3 --version.
Python3 header files (included in the Python3 development package).
pip3 available on the path. Verify with pip3 --version.

Debian / Ubuntu

sudo apt-get install gcc python3-pip python3-dev

RHEL 6/ CentOS 6

Tested on CentOS release 6.9 (Final).

Note

You will need to enable EPEL before.

sudo yum install -y gcc python34.x86_64 python34-devel.x86_64 python34-setuptools.noarch
# installs pip as it is not available as an OS package
sudo python3 /usr/lib/python3.4/site-packages/easy_install.py pip

RHEL 7 / CentOS 7

Note

You will need to enable EPEL before.

Tested on CentOS Linux release 7.4.1708 (Core).

sudo yum install -y gcc python34.x86_64 python34-devel.x86_64 python34-pip.noarch

Amazon Linux

sudo yum install -y gcc python35-pip.noarch python35-devel.x86_64

MacOS

We recommend that you use Homebrew:

brew install python3

git¶

git 1.9 or better is required. Verify with git --version.

Debian / Ubuntu

sudo apt-get install git

Red Hat / CentOS / Amazon Linux

sudo yum install git

Note

If you use RHEL, please ensure to install a recent version of git via the Red Hat Software Collections.

MacOS

git is already installed on MacOS.

JDK¶

A JDK is required on all machines where you want to launch Elasticsearch. If you use Rally just as a load generator, no JDK is required.

We recommend to use Oracle JDK but you are free to use OpenJDK as well. For details on how to install a JDK, please see your operating system’s documentation pages.

Note

If you have Rally download, install and benchmark a local copy of Elasticsearch (i.e., the default Rally behavior) be sure to configure the Operating System (OS) of your Rally server with the recommended kernel settings

Installing Rally¶

Simply install Rally with pip: pip3 install esrally

Note

Depending on your system setup you may need to prepend this command with sudo.

If you get errors during installation, it is probably due to the installation of psutil which we use to gather system metrics like CPU utilization. Please ensure that you have installed the Python development package as documented in the prerequisites section above.

Non-sudo Install¶

If you don’t want to use sudo when installing Rally, installation is still possible but a little more involved:

Specify the --user option when installing Rally (step 2 above), so the command to be issued is: python3 setup.py develop --user.
Check the output of the install script or lookup the Python documentation on the variable site.USER_BASE to find out where the script is located. On Linux, this is typically ~/.local/bin.

You can now either add ~/.local/bin to your path or invoke Rally via ~/.local/bin/esrally instead of just esrally.

VirtualEnv Install¶

You can also use Virtualenv to install Rally into an isolated Python environment without sudo.

Set up a new virtualenv environment in a directory with virtualenv --python=python3 .
Activate the environment with source /path/to/virtualenv/dir/bin/activate
Install Rally with pip install esrally

Whenever you want to use Rally, run the activation script (step 2 above) first. When you are done, simply execute deactivate in the shell to exit the virtual environment.

Kubernetes Job¶

You can run Rally as a Kubernetes Job via this Helm Chart.

Docker¶

You can run Rally as a docker container too. Follow the instructions here.

Offline Install¶

If you are in a corporate environment where your servers do not have any access to the Internet, you can use Rally’s offline installation package. Follow these steps to install Rally:

Install all prerequisites as documented above.
Download the offline installation package for the latest release and copy it to the target machine(s).
Decompress the installation package with tar -xzf esrally-dist-*.tar.gz.
Run the install script with sudo ./esrally-dist-*/install.sh.

Next Steps¶

After you have installed Rally, you need to configure it. Just run esrally configure or follow the configuration help page for more guidance.

Configuration¶

Rally has to be configured once after installation. If you just run esrally after installing Rally, it will detect that the configuration file is missing and asks you a few questions.

If you want to reconfigure Rally at any later time, just run esrally configure again.

Simple Configuration¶

By default, Rally will run a simpler configuration routine and autodetect as much settings as possible or choose defaults for you. If you need more control you can run Rally with esrally configure --advanced-config.

Rally can build Elasticsearch either from sources or use an official binary distribution. If you have Rally build Elasticsearch from sources, it can only be used to benchmark Elasticsearch 5.0 and above. The reason is that with Elasticsearch 5.0 the build tool was switched from Maven to Gradle. As Rally only supports Gradle, it is limited to Elasticsearch 5.0 and above.

If you want to build Elasticsearch from sources, Gradle needs to be installed prior to running the configuration routine.

Let’s go through an example step by step: First run esrally:

dm@io:~ $ esrally

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/

Running simple configuration. You can run the advanced configuration with:

  esrally configure --advanced-config

* Autodetecting available third-party software
  git    : [OK]
  gradle : [OK]
  JDK    : [OK]

* Setting up benchmark data directory in /Users/dm/.rally/benchmarks

As you can see above, Rally autodetects if git, Gradle and a JDK are installed. If you don’t have Gradle, that’s no problem, you are just not able to build Elasticsearch from sources. Let’s assume you don’t have Gradle installed:

dm@io:~ $ esrally

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/

Running simple configuration. You can run the advanced configuration with:

  esrally configure --advanced-config

* Autodetecting available third-party software
  git    : [OK]
  gradle : [MISSING]
  JDK 8  : [OK]

********************************************************************************
You don't have the required software to benchmark Elasticsearch source builds.

You can still benchmark binary distributions with e.g.:

  esrally --distribution-version=5.0.0
********************************************************************************

As you can see, Rally tells you that you cannot build Elasticsearch from sources but you can still benchmark official binary distributions.

It’s also possible that Rally cannot automatically find your JDK 8 or JDK 9 home directory. In that case, it will ask you later in the configuration process. If you do not provide a JDK home directory, Rally cannot start Elasticsearch on this machine but you can still use it as a load generator to benchmark remote clusters.

After running the initial detection, Rally will try to autodetect your Elasticsearch project directory (either in the current directory or in ../elasticsearch) or will choose a default directory:

* Setting up benchmark data directory in /Users/dm/.rally/benchmarks
* Setting up benchmark source directory in /Users/dm/.rally/benchmarks/src/elasticsearch

If Rally has not found Gradle in the first step, it will not ask you for a source directory and just go on.

Now Rally is done:

Configuration successfully written to /Users/dm/.rally/rally.ini. Happy benchmarking!

To benchmark Elasticsearch with the default benchmark, run:

  esrally

More info about Rally:

* Type esrally --help
* Read the documentation at https://esrally.readthedocs.io/en/latest/
* Ask a question on the forum at https://discuss.elastic.co/c/elasticsearch/rally

Congratulations! Time to run your first benchmark.

Advanced Configuration¶

If you need more control over a few variables or want to store your metrics in a dedicated Elasticsearch metrics store, then you should run the advanced configuration routine. You can invoke it at any time with esrally configure --advanced-config.

Prerequisites¶

When using the advanced configuration, you can choose that Rally stores its metrics not in-memory but in a dedicated Elasticsearch instance. Therefore, you will also need the following software installed:

Elasticsearch: a dedicated Elasticsearch instance which acts as the metrics store for Rally. If you don’t want to set it up yourself you can also use Elastic Cloud.
Optional: Kibana (also included in Elastic Cloud).

Preparation¶

First install Elasticsearch 5.0 or higher. A simple out-of-the-box installation with a single node will suffice. Rally uses this instance to store metrics data. It will setup the necessary indices by itself. The configuration procedure of Rally will you ask for host and port of this cluster.

Note

Rally will choose the port range 39200-39300 (HTTP) and 39300-39400 (transport) for the benchmark cluster, so please ensure that this port range is not used by the metrics store.

Optional but recommended is to install also Kibana. However, note that Kibana will not be auto-configured by Rally.

Configuration Options¶

Rally will ask you a few more things in the advanced setup:

Benchmark data directory: Rally stores all benchmark related data in this directory which can take up to several tens of GB. If you want to use a dedicated partition, you can specify a different data directory here.
Elasticsearch project directory: This is the directory where the Elasticsearch sources are located. If you don’t actively develop on Elasticsearch you can just leave the default but if you want to benchmark local changes you should point Rally to your project directory. Note that Rally will run builds with Gradle in this directory (it runs gradle clean and gradle :distribution:tar:assemble).
JDK root directory: Rally will only ask this if it could not autodetect the JDK home by itself. Just enter the root directory of the JDK you want to use. By default, Rally will choose Java 8 if available and fallback to Java 9.
Metrics store type: You can choose between in-memory which requires no additional setup or elasticsearch which requires that you start a dedicated Elasticsearch instance to store metrics but gives you much more flexibility to analyse results.
Metrics store settings (only for metrics store type elasticsearch): Provide the connection details to the Elasticsearch metrics store. This should be an instance that you use just for Rally but it can be a rather small one. A single node cluster with default setting should do it. When using self-signed certificates on the Elasticsearch metrics store, certificate verification can be turned off by setting the datastore.ssl.verification_mode setting to none. Alternatively you can enter the path to the certificate authority’s signing certificate in datastore.ssl.certificate_authorities. Both settings are optional.
Name for this benchmark environment (only for metrics store type elasticsearch): You can use the same metrics store for multiple environments (e.g. local, continuous integration etc.) so you can separate metrics from different environments by choosing a different name.
whether or not Rally should keep the Elasticsearch benchmark candidate installation including all data by default. This will use lots of disk space so you should wipe ~/.rally/benchmarks/races regularly.

Proxy Configuration¶

Rally downloads all necessary data automatically for you:

Elasticsearch distributions from elastic.co if you specify --distribution-version=SOME_VERSION_NUMBER
Elasticsearch source code from Github if you specify a revision number e.g. --revision=952097b
Track meta-data from Github
Track data from an S3 bucket

Hence, it needs to connect via http(s) to the outside world. If you are behind a corporate proxy you need to configure Rally and git. As many other Unix programs, Rally relies that the HTTP proxy URL is available in the environment variable http_proxy (note that this is in lower-case). Hence, you should add this line to your shell profile, e.g. ~/.bash_profile:

export http_proxy=http://proxy.acme.org:8888/

Afterwards, source the shell profile with source ~/.bash_profile and verify that the proxy URL is correctly set with echo $http_proxy.

Finally, you can set up git:

git config --global http.proxy $http_proxy

For details, please refer to the Git config documentation.

Please verify that the proxy setup for git works correctly by cloning any repository, e.g. the rally-tracks repository:

git clone https://github.com/elastic/rally-tracks.git

If the configuration is correct, git will clone this repository. You can delete the folder rally-tracks after this verification step.

To verify that Rally will connect via the proxy server you can check the log file. If the proxy server is configured successfully, Rally will log the following line on startup:

Rally connects via proxy URL [http://proxy.acme.org:3128/] to the Internet (picked up from the environment variable [http_proxy]).

Note

Rally will use this proxy server only for downloading benchmark-related data. It will not use this proxy for the actual benchmark.

Run a Benchmark: Races¶

Definition¶

A “race” in Rally is the execution of a benchmarking experiment. You can choose different benchmarking scenarios (called tracks) for your benchmarks.

List Tracks¶

Start by finding out which tracks are available:

esrally list tracks

This will show the following list:

Name        Description                                          Documents  Compressed Size    Uncompressed Size    Default Challenge        All Challenges
----------  -------------------------------------------------  -----------  -----------------  -------------------  -----------------------  ---------------------------
geonames    POIs from Geonames                                    11396505  252.4 MB           3.3 GB               append-no-conflicts      append-no-conflicts,appe...
geopoint    Point coordinates from PlanetOSM                      60844404  481.9 MB           2.3 GB               append-no-conflicts      append-no-conflicts,appe...
http_logs   HTTP server log data                                 247249096  1.2 GB             31.1 GB              append-no-conflicts      append-no-conflicts,appe...
nested      StackOverflow Q&A stored as nested docs               11203029  663.1 MB           3.4 GB               nested-search-challenge  nested-search-challenge,...
noaa        Global daily weather measurements from NOAA           33659481  947.3 MB           9.0 GB               append-no-conflicts      append-no-conflicts,appe...
nyc_taxis   Taxi rides in New York in 2015                       165346692  4.5 GB             74.3 GB              append-no-conflicts      append-no-conflicts,appe...
percolator  Percolator benchmark based on AOL queries              2000000  102.7 kB           104.9 MB             append-no-conflicts      append-no-conflicts,appe...
pmc         Full text benchmark with academic papers from PMC       574199  5.5 GB             21.7 GB              append-no-conflicts      append-no-conflicts,appe...

The first two columns show the name and a description of each track. A track also specifies one or more challenges which describe the workload to run.

Starting a Race¶

Note

Do not run Rally as root as Elasticsearch will refuse to start with root privileges.

To start a race you have to define the track and challenge to run. For example:

esrally --distribution-version=6.0.0 --track=geopoint --challenge=append-fast-with-conflicts

Rally will then start racing on this track. If you have never started Rally before, it should look similar to the following output:

dm@io:~ $ esrally --distribution-version=6.0.0 --track=geopoint --challenge=append-fast-with-conflicts

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/

[INFO] Racing on track [geopoint], challenge [append-fast-with-conflicts] and car ['defaults'] with version [6.0.0].
[INFO] Downloading Elasticsearch 6.0.0 ... [OK]
[INFO] Rally will delete the benchmark candidate after the benchmark
[INFO] Downloading data from [http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/geopoint/documents.json.bz2] (482 MB) to [/Users/dm/.rally/benchmarks/data/geopoint/documents.json.bz2] ... [OK]
[INFO] Decompressing track data from [/Users/dm/.rally/benchmarks/data/geopoint/documents.json.bz2] to [/Users/dm/.rally/benchmarks/data/geopoint/documents.json] (resulting size: 2.28 GB) ... [OK]
[INFO] Preparing file offset table for [/Users/dm/.rally/benchmarks/data/geopoint/documents.json] ... [OK]
Running index-update                                                           [  0% done]

Please be patient as it will take a while to run the benchmark.

When the race has finished, Rally will show a summary on the command line:

|                          Metric |         Task |     Value |   Unit |
|--------------------------------:|-------------:|----------:|-------:|
|                   Indexing time |              |   124.712 |    min |
|                      Merge time |              |   21.8604 |    min |
|                    Refresh time |              |   4.49527 |    min |
|             Merge throttle time |              |  0.120433 |    min |
|                Median CPU usage |              |     546.5 |      % |
|              Total Young Gen GC |              |    72.078 |      s |
|                Total Old Gen GC |              |     3.426 |      s |
|                      Index size |              |   2.26661 |     GB |
|                 Totally written |              |    30.083 |     GB |
|          Heap used for segments |              |   10.7148 |     MB |
|        Heap used for doc values |              | 0.0135536 |     MB |
|             Heap used for terms |              |   9.22965 |     MB |
|            Heap used for points |              |   0.78789 |     MB |
|     Heap used for stored fields |              |  0.683708 |     MB |
|                   Segment count |              |       115 |        |
|                  Min Throughput | index-update |   59210.4 | docs/s |
|               Median Throughput | index-update |   65276.2 | docs/s |
|                  Max Throughput | index-update |   76516.6 | docs/s |
|       50.0th percentile latency | index-update |   556.269 |     ms |
|       90.0th percentile latency | index-update |   852.779 |     ms |
|       99.0th percentile latency | index-update |   1854.31 |     ms |
|       99.9th percentile latency | index-update |   2972.96 |     ms |
|      99.99th percentile latency | index-update |   4106.91 |     ms |
|        100th percentile latency | index-update |   4542.84 |     ms |
|  50.0th percentile service time | index-update |   556.269 |     ms |
|  90.0th percentile service time | index-update |   852.779 |     ms |
|  99.0th percentile service time | index-update |   1854.31 |     ms |
|  99.9th percentile service time | index-update |   2972.96 |     ms |
| 99.99th percentile service time | index-update |   4106.91 |     ms |
|   100th percentile service time | index-update |   4542.84 |     ms |
|                  Min Throughput |  force-merge |  0.221067 |  ops/s |
|               Median Throughput |  force-merge |  0.221067 |  ops/s |
|                  Max Throughput |  force-merge |  0.221067 |  ops/s |
|        100th percentile latency |  force-merge |   4523.52 |     ms |
|   100th percentile service time |  force-merge |   4523.52 |     ms |


----------------------------------
[INFO] SUCCESS (took 1624 seconds)
----------------------------------

Note

You can save this report also to a file by using --report-file=/path/to/your/report.md and save it as CSV with --report-format=csv.

What did Rally just do?

It downloaded and started Elasticsearch 6.0.0
It downloaded the relevant data for the geopoint track
It ran the actual benchmark
And finally it reported the results

If you are curious about the operations that Rally has run, please inspect the geopoint track specification or start to write your own tracks. You can also configure Rally to store all data samples in Elasticsearch so you can analyze the results with Kibana. Finally, you may want to change the Elasticsearch configuration.

Compare Results: Tournaments¶

Suppose, we want to analyze the impact of a performance improvement.

First, we need a baseline measurement. For example:

esrally --track=pmc --revision=latest --user-tag="intention:baseline_github_1234"

Above we run the baseline measurement based on the latest source code revision of Elasticsearch. We can use the command line parameter --user-tag to provide a key-value pair to document the intent of a race.

Then we implement our changes and finally we want to run another benchmark to see the performance impact of the change. In that case, we do not want Rally to change our source tree and thus specify the pseudo-revision current:

esrally --track=pmc --revision=current --user-tag="intention:reduce_alloc_1234"

After we’ve run both races, we want to know about the performance impact. With Rally we can analyze differences of two given races easily. First of all, we need to find two races to compare by issuing esrally list races:

dm@io:~ $ esrally list races

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/
Recent races:

Race Timestamp    Track    Track Parameters   Challenge            Car       User Tag
----------------  -------  ------------------ -------------------  --------  ------------------------------
20160518T122341Z  pmc                         append-no-conflicts  defaults  intention:reduce_alloc_1234
20160518T112057Z  pmc                         append-no-conflicts  defaults  intention:baseline_github_1234
20160518T101957Z  pmc                         append-no-conflicts  defaults

We can see that the user tag helps us to recognize races. We want to compare the two most recent races and have to provide the two race timestamps in the next step:

dm@io:~ $ esrally compare --baseline=20160518T112057Z --contender=20160518T112341Z

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/

Comparing baseline
  Race timestamp: 2016-05-18 11:20:57
  Challenge: append-no-conflicts
  Car: defaults

with contender
  Race timestamp: 2016-05-18 12:23:41
  Challenge: append-no-conflicts
  Car: defaults

------------------------------------------------------
    _______             __   _____
   / ____(_)___  ____ _/ /  / ___/_________  ________
  / /_  / / __ \/ __ `/ /   \__ \/ ___/ __ \/ ___/ _ \
 / __/ / / / / / /_/ / /   ___/ / /__/ /_/ / /  /  __/
/_/   /_/_/ /_/\__,_/_/   /____/\___/\____/_/   \___/
------------------------------------------------------
                                                  Metric    Baseline    Contender               Diff
--------------------------------------------------------  ----------  -----------  -----------------
                        Min Indexing Throughput [docs/s]       19501        19118  -383.00000
                     Median Indexing Throughput [docs/s]       20232      19927.5  -304.45833
                        Max Indexing Throughput [docs/s]       21172        20849  -323.00000
                                     Indexing time [min]     55.7989       56.335    +0.53603
                                        Merge time [min]     12.9766      13.3115    +0.33495
                                      Refresh time [min]     5.20067      5.20097    +0.00030
                                        Flush time [min]   0.0648667    0.0681833    +0.00332
                               Merge throttle time [min]    0.796417     0.879267    +0.08285
               Query latency term (50.0 percentile) [ms]     2.10049      2.15421    +0.05372
               Query latency term (90.0 percentile) [ms]     2.77537      2.84168    +0.06630
              Query latency term (100.0 percentile) [ms]     4.52081      5.15368    +0.63287
        Query latency country_agg (50.0 percentile) [ms]     112.049      110.385    -1.66392
        Query latency country_agg (90.0 percentile) [ms]     128.426      124.005    -4.42138
       Query latency country_agg (100.0 percentile) [ms]     155.989      133.797   -22.19185
             Query latency scroll (50.0 percentile) [ms]     16.1226      14.4974    -1.62519
             Query latency scroll (90.0 percentile) [ms]     17.2383      15.4079    -1.83043
            Query latency scroll (100.0 percentile) [ms]     18.8419      18.4241    -0.41784
 Query latency country_agg_cached (50.0 percentile) [ms]     1.70223      1.64502    -0.05721
 Query latency country_agg_cached (90.0 percentile) [ms]     2.34819      2.04318    -0.30500
Query latency country_agg_cached (100.0 percentile) [ms]     3.42547      2.86814    -0.55732
            Query latency default (50.0 percentile) [ms]     5.89058      5.83409    -0.05648
            Query latency default (90.0 percentile) [ms]     6.71282      6.64662    -0.06620
           Query latency default (100.0 percentile) [ms]     7.65307       7.3701    -0.28297
             Query latency phrase (50.0 percentile) [ms]     1.82687      1.83193    +0.00506
             Query latency phrase (90.0 percentile) [ms]     2.63714      2.46286    -0.17428
            Query latency phrase (100.0 percentile) [ms]     5.39892      4.22367    -1.17525
                            Median CPU usage (index) [%]     668.025       679.15   +11.12499
                            Median CPU usage (stats) [%]      143.75        162.4   +18.64999
                           Median CPU usage (search) [%]       223.1        229.2    +6.10000
                                  Total Young Gen GC [s]      39.447       40.456    +1.00900
                                    Total Old Gen GC [s]       7.108        7.703    +0.59500
                                         Index size [GB]     3.25475      3.25098    -0.00377
                                    Totally written [GB]     17.8434      18.3143    +0.47083
                             Heap used for segments [MB]     21.7504      21.5901    -0.16037
                           Heap used for doc values [MB]     0.16436      0.13905    -0.02531
                                Heap used for terms [MB]     20.0293      19.9159    -0.11345
                                Heap used for norms [MB]    0.105469    0.0935669    -0.01190
                               Heap used for points [MB]    0.773487     0.772155    -0.00133
                               Heap used for points [MB]    0.677795     0.669426    -0.00837
                                           Segment count         136          121   -15.00000
                     Indices Stats(90.0 percentile) [ms]     3.16053      3.21023    +0.04969
                     Indices Stats(99.0 percentile) [ms]     5.29526      3.94132    -1.35393
                    Indices Stats(100.0 percentile) [ms]     5.64971      7.02374    +1.37403
                       Nodes Stats(90.0 percentile) [ms]     3.19611      3.15251    -0.04360
                       Nodes Stats(99.0 percentile) [ms]     4.44111      4.87003    +0.42892
                      Nodes Stats(100.0 percentile) [ms]     5.22527      5.66977    +0.44450

Tips and Tricks¶

This section covers various tips and tricks in a recipe-style fashion.

Benchmarking an existing cluster¶

Warning

If you are just getting started with Rally and don’t understand how it works, please do NOT run it against any production or production-like cluster. Besides, benchmarks should be executed in a dedicated environment anyway where no additional traffic skews results.

Note

We assume in this recipe, that Rally is already properly configured.

Consider the following configuration: You have an existing benchmarking cluster, that consists of three Elasticsearch nodes running on 10.5.5.10, 10.5.5.11, 10.5.5.12. You’ve setup the cluster yourself and want to benchmark it with Rally. Rally is installed on 10.5.5.5.

First of all, we need to decide on a track. So, we run esrally list tracks:

Name        Description                                          Documents  Compressed Size    Uncompressed Size    Default Challenge        All Challenges
----------  -------------------------------------------------  -----------  -----------------  -------------------  -----------------------  ---------------------------
geonames    POIs from Geonames                                    11396505  252.4 MB           3.3 GB               append-no-conflicts      append-no-conflicts,appe...
geopoint    Point coordinates from PlanetOSM                      60844404  481.9 MB           2.3 GB               append-no-conflicts      append-no-conflicts,appe...
http_logs   HTTP server log data                                 247249096  1.2 GB             31.1 GB              append-no-conflicts      append-no-conflicts,appe...
nested      StackOverflow Q&A stored as nested docs               11203029  663.1 MB           3.4 GB               nested-search-challenge  nested-search-challenge,...
noaa        Global daily weather measurements from NOAA           33659481  947.3 MB           9.0 GB               append-no-conflicts      append-no-conflicts,appe...
nyc_taxis   Taxi rides in New York in 2015                       165346692  4.5 GB             74.3 GB              append-no-conflicts      append-no-conflicts,appe...
percolator  Percolator benchmark based on AOL queries              2000000  102.7 kB           104.9 MB             append-no-conflicts      append-no-conflicts,appe...
pmc         Full text benchmark with academic papers from PMC       574199  5.5 GB             21.7 GB              append-no-conflicts      append-no-conflicts,appe...

We’re interested in a full text benchmark, so we’ll choose to run pmc. If you have your own data that you want to use for benchmarks, then please create your own track instead; the metrics you’ll gather which be representative and much more useful than some default track.

Next, we need to know which machines to target which is easy as we can see that from the diagram above.

Finally we need to check which pipeline to use. For this case, the benchmark-only pipeline is suitable as we don’t want Rally to provision the cluster for us.

Now we can invoke Rally:

esrally --track=pmc --target-hosts=10.5.5.10:9200,10.5.5.11:9200,10.5.5.12:9200 --pipeline=benchmark-only

If you have X-Pack Security enabled, then you’ll also need to specify another parameter to use https and to pass credentials:

esrally --track=pmc --target-hosts=10.5.5.10:9243,10.5.5.11:9243,10.5.5.12:9243 --pipeline=benchmark-only --client-options="use_ssl:true,verify_certs:true,basic_auth_user:'elastic',basic_auth_password:'changeme'"

Benchmarking a remote cluster¶

Contrary to the previous recipe, you want Rally to provision all cluster nodes.

We will use the following configuration for the example:

You will start Rally on 10.5.5.5. We will call this machine the “benchmark coordinator”.
Your Elasticsearch cluster will consist of two nodes which run on 10.5.5.10 and 10.5.5.11. We will call these machines the “benchmark candidate”s.

Note

All esrallyd nodes form a cluster that communicates via the “benchmark coordinator”. For aesthetic reasons we do not show a direct connection between the “benchmark coordinator” and all nodes.

To run a benchmark for this scenario follow these steps:

Install and configure Rally on all machines. Be sure that the same version is installed on all of them and fully configured.
Start the Rally daemon on each machine. The Rally daemon allows Rally to communicate with all remote machines. On the benchmark coordinator run esrallyd start --node-ip=10.5.5.5 --coordinator-ip=10.5.5.5 and on the benchmark candidate machines run esrallyd start --node-ip=10.5.5.10 --coordinator-ip=10.5.5.5 and esrallyd start --node-ip=10.5.5.11 --coordinator-ip=10.5.5.5 respectively. The --node-ip parameter tells Rally the IP of the machine on which it is running. As some machines have more than one network interface, Rally will not attempt to auto-detect the machine IP. The --coordinator-ip parameter tells Rally the IP of the benchmark coordinator node.
Start the benchmark by invoking Rally as usual on the benchmark coordinator, for example: esrally --distribution-version=5.0.0 --target-hosts=10.5.5.10:9200,10.5.5.11:9200. Rally will derive from the --target-hosts parameter that it should provision the nodes 10.5.5.10 and 10.5.5.11.
After the benchmark has finished you can stop the Rally daemon again. On the benchmark coordinator and on the benchmark candidates run esrallyd stop.

Note

Logs are managed per machine, so all relevant log files and also telemetry output is stored on the benchmark candidates but not on the benchmark coordinator.

Now you might ask yourself what the differences to benchmarks of existing clusters are. In general you should aim to give Rally as much control as possible as benchmark are easier reproducible and you get more metrics. The following table provides some guidance on when to choose which option:

Your requirement	Recommendation
You want to use Rally’s telemetry devices	Use Rally daemon, as it can provision the remote node for you
You want to benchmark a source build of Elasticsearch	Use Rally daemon, as it can build Elasticsearch for you
You want to tweak the cluster configuration yourself	Use Rally daemon with a custom configuration or set up the cluster by yourself and use `--pipeline=benchmark-only`
You need to run a benchmark with plugins	Use Rally daemon if the plugins are supported or set up the cluster by yourself and use `--pipeline=benchmark-only`
You need to run a benchmark against multiple nodes	Use Rally daemon if all nodes can be configured identically. For more complex cases, set up the cluster by yourself and use `--pipeline=benchmark-only`

Rally daemon will be able to cover most of the cases described above in the future so there should be almost no case where you need to use the benchmark-only pipeline.

Distributing the load test driver¶

By default, Rally will generate load on the same machine where you start a benchmark. However, when you are benchmarking larger clusters, a single load test driver machine may not be able to generate sufficient load. In these cases, you should use multiple load driver machines. We will use the following configuration for the example:

You will start Rally on 10.5.5.5. We will call this machine the “benchmark coordinator”.
You will start two load drivers on 10.5.5.6 and 10.5.5.7. Note that one load driver will simulate multiple clients. Rally will simply assign clients to load driver machines in a round-robin fashion.
Your Elasticsearch cluster will consist of three nodes which run on 10.5.5.11, 10.5.5.12 and 10.5.5.13. We will call these machines the “benchmark candidate”. For simplicity, we will assume an externally provisioned cluster but you can also use Rally to setup the cluster for you (see above).

Install and configure Rally on all machines. Be sure that the same version is installed on all of them and fully configured.
Start the Rally daemon on each machine. The Rally daemon allows Rally to communicate with all remote machines. On the benchmark coordinator run esrallyd start --node-ip=10.5.5.5 --coordinator-ip=10.5.5.5 and on the load driver machines run esrallyd start --node-ip=10.5.5.6 --coordinator-ip=10.5.5.5 and esrallyd start --node-ip=10.5.5.7 --coordinator-ip=10.5.5.5 respectively. The --node-ip parameter tells Rally the IP of the machine on which it is running. As some machines have more than one network interface, Rally will not attempt to auto-detect the machine IP. The --coordinator-ip parameter tells Rally the IP of the benchmark coordinator node.
Start the benchmark by invoking Rally on the benchmark coordinator, for example: esrally --pipeline=benchmark-only --load-driver-hosts=10.5.5.6,10.5.5.7 --target-hosts=10.5.5.11:9200,10.5.5.12:9200,10.5.5.13:9200.
After the benchmark has finished you can stop the Rally daemon again. On the benchmark coordinator and on the load driver machines run esrallyd stop.

Note

Rally neither distributes code (i.e. custom runners or parameter sources) nor data automatically. You should place all tracks and their data on all machines in the same directory before starting the benchmark. Alternatively, you can store your track in a custom track repository.

Note

As indicated in the diagram, track data will be downloaded by each load driver machine separately. If you want to avoid that, you can run a benchmark once without distributing the load test driver (i.e. do not specify --load-driver-hosts) and then copy the contents of ~/.rally/benchmarks/data to all load driver machines.

Changing the default track repository¶

Rally supports multiple track repositories. This allows you for example to have a separate company-internal repository for your own tracks that is separate from Rally’s default track repository. However, you always need to define --track-repository=my-custom-repository which can be cumbersome. If you want to avoid that and want Rally to use your own track repository by default you can just replace the default track repository definition in ~./rally/rally.ini. Consider this example:

...
[tracks]
default.url = git@github.com:elastic/rally-tracks.git
teamtrackrepo.url = git@example.org/myteam/my-tracks.git

If teamtrackrepo should be the default track repository, just define it as default.url. E.g.:

...
[tracks]
default.url = git@example.org/myteam/my-tracks.git
old-rally-default.url=git@github.com:elastic/rally-tracks.git

Also don’t forget to rename the folder of your local working copy as Rally will search for a track repository with the name default:

cd ~/.rally/benchmarks/tracks/
mv default old-rally-default
mv teamtrackrepo default

From now on, Rally will treat your repository as default and you need to run Rally with --track-repository=old-rally-default if you want to use the out-of-the-box Rally tracks.

Define Custom Workloads: Tracks¶

Definition¶

A track is the specification of one or more benchmarking scenarios with a specific document corpus.

Note

Please see the track reference for more information on the structure of a track.

Example track¶

Let’s create an example track step by step. We will call this track “tutorial”. The track consists of two components: the data and the actual track specification which describes the workload that Rally should apply. We will store everything in the directory ~/rally-tracks/tutorial but you can choose any other location.

First, we need some data. Geonames provides geo data under a creative commons license. We will download allCountries.zip (around 300MB), extract it and inspect allCountries.txt.

You will note that the file is tab-delimited but we need JSON to bulk-index data with Elasticsearch. So we can use a small script to do the conversion for us:

import json

cols = (("geonameid", "int", True),
        ("name", "string", True),
        ("asciiname", "string", False),
        ("alternatenames", "string", False),
        ("latitude", "double", True),
        ("longitude", "double", True),
        ("feature_class", "string", False),
        ("feature_code", "string", False),
        ("country_code", "string", True),
        ("cc2", "string", False),
        ("admin1_code", "string", False),
        ("admin2_code", "string", False),
        ("admin3_code", "string", False),
        ("admin4_code", "string", False),
        ("population", "long", True),
        ("elevation", "int", False),
        ("dem", "string", False),
        ("timezone", "string", False))


def main():
    with open("allCountries.txt", "rt", encoding="UTF-8") as f:
        for line in f:
            tup = line.strip().split("\t")
            record = {}
            for i in range(len(cols)):
                name, type, include = cols[i]
                if tup[i] != "" and include:
                    if type in ("int", "long"):
                        record[name] = int(tup[i])
                    elif type == "double":
                        record[name] = float(tup[i])
                    elif type == "string":
                        record[name] = tup[i]
            print(json.dumps(record, ensure_ascii=False))


if __name__ == "__main__":
    main()

Store the script as toJSON.py in our tutorial directory (~/rally-tracks/tutorial) and invoke the script with python3 toJSON.py > documents.json.

We also need a mapping file for our documents. Store the following snippet as index.json in the tutorial directory:

{
  "settings": {
    "index.number_of_replicas": 0
  },
  "mappings": {
    "docs": {
      "dynamic": "strict",
      "properties": {
        "geonameid": {
          "type": "long"
        },
        "name": {
          "type": "text"
        },
        "latitude": {
          "type": "double"
        },
        "longitude": {
          "type": "double"
        },
        "country_code": {
          "type": "text"
        },
        "population": {
          "type": "long"
        }
      }
    }
  }
}

For details on the allowed syntax, see the Elasticsearch documentation on mappings and the create index API.

Finally, add a file called track.json in the tutorial directory:

{
  "description": "Tutorial benchmark for Rally",
  "indices": [
    {
      "name": "geonames",
      "auto-managed": false,
      "body": "index.json",
      "types": [ "docs" ]
    }
  ],
  "corpora": [
    {
      "name": "rally-tutorial",
      "documents": [
        {
          "source-file": "documents.json",
          "document-count": 11658903,
          "uncompressed-bytes": 1544799789
        }
      ]
    }
  ],
  "challenges": [
    {
      "name": "index-and-query",
      "default": true,
      "schedule": [
        {
          "operation": {
            "operation-type": "delete-index"
          }
        },
        {
          "operation": {
            "operation-type": "create-index"
          }
        },
        {
          "operation": {
            "operation-type": "cluster-health",
            "request-params": {
              "wait_for_status": "green"
            }
          }
        },
        {
          "operation": {
            "operation-type": "bulk",
            "bulk-size": 5000
          },
          "warmup-time-period": 120,
          "clients": 8
        },
        {
          "operation": {
            "operation-type": "force-merge"
          }
        },
        {
          "operation": {
            "name": "query-match-all",
            "operation-type": "search",
            "body": {
              "query": {
                "match_all": {}
              }
            }
          },
          "clients": 8,
          "warmup-iterations": 1000,
          "iterations": 1000,
          "target-throughput": 100
        }
      ]
    }
  ]
}

A few things to note:

The numbers below the documents property are needed to verify integrity and provide progress reports. Determine the correct document count with wc -l documents.json and the size in bytes with stat -f "%z" documents.json.
You can add as many queries as you want. We use the official Python Elasticsearch client to issue queries.

Note

You can store any supporting scripts along with your track. However, you need to place them in a directory starting with “_”, e.g. “_support”. Rally loads track plugins (see below) from any directory but will ignore directories starting with “_”.

Note

We have defined a JSON schema for tracks which you can use to check how to define your track. You should also check the tracks provided by Rally for inspiration.

When you invoke esrally list tracks --track-path=~/rally-tracks/tutorial, the new track should now appear:

dm@io:~ $ esrally list tracks --track-path=~/rally-tracks/tutorial

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/
Available tracks:

Name        Description                   Documents    Compressed Size  Uncompressed Size  Default Challenge  All Challenges
----------  ----------------------------- -----------  ---------------  -----------------  -----------------  ---------------
tutorial    Tutorial benchmark for Rally      11658903  N/A              1.4 GB            index-and-query    index-and-query

Congratulations, you have created your first track! You can test it with esrally --distribution-version=6.0.0 --track-path=~/rally-tracks/tutorial.

Adding support for test mode¶

When you invoke Rally with --test-mode, it switches to a mode that allows you to check your track very quickly for syntax errors. To achieve that, it will postprocess its internal track representation after loading it:

Iteration-based tasks will run at most one warmup iteration and one measurement iteration.
Time-period-based tasks will run for at most 10 seconds without any warmup.

Rally will postprocess all data file names of a track. So instead of documents.json, Rally will attempt to find documents-1k.json and will assume it contains 1.000 documents. However, you need to prepare these data files otherwise this test mode is not supported.

The preparation is very easy. Just pick 1.000 documents for every data file in your track. We choose the first 1.000 here but it does not matter usually which part you choose: head -n 1000 documents.json > documents-1k.json.

Structuring your track¶

track.json is just the entry point to a track but you can split your track as you see fit. Suppose you want to add more challenges to the track above but you want to keep them in a separate files. Let’s start by storing our challenge in a separate file, e.g in challenges/index-and-query.json. Create the directory and store the following in index-and-query.json:

{
  "name": "index-and-query",
  "default": true,
  "schedule": [
    {
      "operation": {
        "operation-type": "delete-index"
      }
    },
    {
      "operation": {
        "operation-type": "create-index"
      }
    },
    {
      "operation": {
        "operation-type": "cluster-health",
        "request-params": {
          "wait_for_status": "green"
        }
      }
    },
    {
      "operation": {
        "operation-type": "bulk",
        "bulk-size": 5000
      },
      "warmup-time-period": 120,
      "clients": 8
    },
    {
      "operation": {
        "operation-type": "force-merge"
      }
    },
    {
      "operation": {
        "name": "query-match-all",
        "operation-type": "search",
        "body": {
          "query": {
            "match_all": {}
          }
        }
      },
      "clients": 8,
      "warmup-iterations": 1000,
      "iterations": 1000,
      "target-throughput": 100
    }
  ]
}

Now modify track.json so it knows about your new file:

{
  "description": "Tutorial benchmark for Rally",
  "indices": [
    {
      "name": "geonames",
      "auto-managed": false,
      "body": "index.json",
      "types": [ "docs" ]
    }
  ],
  "corpora": [
    {
      "name": "rally-tutorial",
      "documents": [
        {
          "source-file": "documents.json",
          "document-count": 11658903,
          "uncompressed-bytes": 1544799789
        }
      ]
    }
  ],
  "challenges": [
    {% include "challenges/index-and-query.json" %}
  ]
}

We replaced the challenge content with {% include "challenges/index-and-query.json" %} which tells Rally to include the challenge from the provided file. You can use include on arbitrary parts of your track.

If you want to reuse operation definitions across challenges, you can also define them in a separate operations block and just refer to them by name in the corresponding challenge:

{
  "description": "Tutorial benchmark for Rally",
  "indices": [
    {
      "name": "geonames",
      "auto-managed": false,
      "body": "index.json",
      "types": [ "docs" ]
    }
  ],
  "corpora": [
    {
      "name": "rally-tutorial",
      "documents": [
        {
          "source-file": "documents.json",
          "document-count": 11658903,
          "uncompressed-bytes": 1544799789
        }
      ]
    }
  ],
  "operations": [
    {
      "name": "delete",
      "operation-type": "delete-index"
    },
    {
      "name": "create",
      "operation-type": "create-index"
    },
    {
      "name": "wait-for-green",
      "operation-type": "cluster-health",
      "request-params": {
        "wait_for_status": "green"
      }
    },
    {
      "name": "bulk-index",
      "operation-type": "bulk",
      "bulk-size": 5000
    },
    {
      "name": "force-merge",
      "operation-type": "force-merge"
    },
    {
      "name": "query-match-all",
      "operation-type": "search",
      "body": {
        "query": {
          "match_all": {}
        }
      }
    }
  ],
  "challenges": [
    {% include "challenges/index-and-query.json" %}
  ]
}

challenges/index-and-query.json then becomes:

{
  "name": "index-and-query",
  "default": true,
  "schedule": [
    {
      "operation": "delete"
    },
    {
      "operation": "create"
    },
    {
      "operation": "wait-for-green"
    },
    {
      "operation": "bulk-index",
      "warmup-time-period": 120,
      "clients": 8
    },
    {
      "operation": "force-merge"
    },
    {
      "operation": "query-match-all",
      "clients": 8,
      "warmup-iterations": 1000,
      "iterations": 1000,
      "target-throughput": 100
    }
  ]
}

Note how we reference to the operations by their name (e.g. create, bulk-index, force-merge or query-match-all).

If your track consists of multiple challenges, it can be cumbersome to include them all explicitly. Therefore Rally brings a collect helper that collects all related files for you. Let’s adapt our track to use it:

{% import "rally.helpers" as rally %}
{
  "description": "Tutorial benchmark for Rally",
  "indices": [
    {
      "name": "geonames",
      "auto-managed": false,
      "body": "index.json",
      "types": [ "docs" ]
    }
  ],
  "corpora": [
    {
      "name": "rally-tutorial",
      "documents": [
        {
          "source-file": "documents.json",
          "document-count": 11658903,
          "uncompressed-bytes": 1544799789
        }
      ]
    }
  ],
  "operations": [
    {
      "name": "delete",
      "operation-type": "delete-index"
    },
    {
      "name": "create",
      "operation-type": "create-index"
    },
    {
      "name": "wait-for-green",
      "operation-type": "cluster-health",
      "request-params": {
        "wait_for_status": "green"
      }
    },
    {
      "name": "bulk-index",
      "operation-type": "bulk",
      "bulk-size": 5000
    },
    {
      "name": "force-merge",
      "operation-type": "force-merge"
    },
    {
      "name": "query-match-all",
      "operation-type": "search",
      "body": {
        "query": {
          "match_all": {}
        }
      }
    }
  ],
  "challenges": [
    {{ rally.collect(parts="challenges/*.json") }}
  ]
}

We changed two things here. First, we imported helper functions from Rally by adding {% import "rally.helpers" as rally %} in line 1. Second, we used Rally’s collect helper to find and include all JSON files in the “challenges” subdirectory with the statement {{ rally.collect(parts="challenges/*.json") }}. When you add new challenges in this directory, Rally will automatically pick them up.

Note

If you want to check the final result, please check Rally’s log file. Rally will print the fully rendered track there after it has loaded it successfully.

Note

If you define multiple challenges, Rally will run the challenge where default is set to true. If you want to run a different challenge, provide the command line option --challenge=YOUR_CHALLENGE_NAME.

You can even use Jinja2 variables but you need to import the Rally helpers a bit differently then. You also need to declare all variables before the import statement:

{% set clients = 16 %}
{% import "rally.helpers" as rally with context %}

If you use this idiom you can then refer to variables inside your snippets with {{ clients }}.

Advanced topics¶

Template Language¶

Rally uses Jinja2 as template language. This allows you to use Jinja2 expressions in track files.

Extension Points¶

Rally also provides a few extension points to Jinja2:

now: This is a global variable that represents the current date and time when the template is evaluated by Rally.
days_ago(): This is a filter that you can use for date calculations.

You can find an example in the http_logs track:

{
  "name": "range",
    "index": "logs-*",
    "type": "type",
    "body": {
      "query": {
        "range": {
          "@timestamp": {
            "gte": "now-{{'15-05-1998' | days_ago(now)}}d/d",
            "lt": "now/d"
          }
        }
      }
    }
  }
}

The data set that is used in the http_logs track starts on 26-04-1998 but we want to ignore the first few days for this query, so we start on 15-05-1998. The expression {{'15-05-1998' | days_ago(now)}} yields the difference in days between now and the fixed start date and allows us to benchmark time range queries relative to now with a predetermined data set.

Custom parameter sources¶

Note

This is a rather new feature and the API may change! However, the effort to use custom parameter sources is very low.

Warning

Your parameter source is on a performance-critical code-path so please double-check with Rally’s profiling support that you did not introduce any bottlenecks.

Consider the following operation definition:

{
  "name": "term",
  "operation-type": "search",
  "body": {
    "query": {
      "term": {
        "body": "physician"
      }
    }
  }
}

This query is defined statically in the track specification but sometimes you may want to vary parameters, e.g. search also for “mechanic” or “nurse”. In this case, you can write your own “parameter source” with a little bit of Python code.

First, define the name of your parameter source in the operation definition:

{
  "name": "term",
  "operation-type": "search",
  "param-source": "my-custom-term-param-source"
  "professions": ["mechanic", "physician", "nurse"]
}

Rally will recognize the parameter source and looks then for a file track.py in the same directory as the corresponding JSON file. This file contains the implementation of the parameter source:

import random


def random_profession(track, params, **kwargs):
    # choose a suitable index: if there is only one defined for this track
    # choose that one, but let the user always override index and type.
    if len(track.indices) == 1:
        default_index = track.indices[0].name
        if len(track.indices[0].types) == 1:
            default_type = track.indices[0].types[0].name
        else:
            default_type = None
    else:
        default_index = "_all"
        default_type = None

    index_name = params.get("index", default_index)
    type_name = params.get("type", default_type)

    # you must provide all parameters that the runner expects
    return {
        "body": {
            "query": {
                "term": {
                    "body": "%s" % random.choice(params["professions"])
                }
            }
        },
        "index": index_name,
        "type": type_name,
        "use_request_cache": params.get("cache", False)
    }

def register(registry):
    registry.register_param_source("my-custom-term-param-source", random_profession)

The example above shows a simple case that is sufficient if the operation to which your parameter source is applied is idempotent and it does not matter whether two clients execute the same operation.

The function random_profession is the actual parameter source. Rally will bind the name “my-custom-term-param-source” to this function by calling register. register is called by Rally before the track is executed.

The parameter source function needs to declare the parameters track, params and **kwargs. track contains a structured representation of the current track and params contains all parameters that have been defined in the operation definition in track.json. The third parameter is there to ensure a more stable API as Rally evolves. We use it in the example to read the professions to choose.

We also derive an appropriate index and document type from the track’s index definitions but allow the user to override this choice with the index or type parameters as you can see below:

{
  "name": "term",
  "operation-type": "search",
  "param-source": "my-custom-term-param-source"
  "professions": ["mechanic", "physician", "nurse"],
  "index": "employee*",
  "type": "docs"
}

If you need more control, you need to implement a class. The example above, implemented as a class looks as follows:

import random


class TermParamSource:
    def __init__(self, track, params, **kwargs):
        # choose a suitable index: if there is only one defined for this track
        # choose that one, but let the user always override index and type.
        if len(track.indices) == 1:
            default_index = track.indices[0].name
            if len(track.indices[0].types) == 1:
                default_type = track.indices[0].types[0].name
            else:
                default_type = None
        else:
            default_index = "_all"
            default_type = None

        # we can eagerly resolve these parameters already in the constructor...
        self._index_name = params.get("index", default_index)
        self._type_name = params.get("type", default_type)
        self._cache = params.get("cache", False)
        # ... but we need to resolve "profession" lazily on each invocation later
        self._params = params

    def partition(self, partition_index, total_partitions):
        return self

    def size(self):
        return 1

    def params(self):
        # you must provide all parameters that the runner expects
        return {
            "body": {
                "query": {
                    "term": {
                        "body": "%s" % random.choice(self._params["professions"])
                    }
                }
            },
            "index": self._index_name,
            "type": self._type_name,
            "use_request_cache": self._cache
        }


def register(registry):
    registry.register_param_source("my-custom-term-param-source", TermParamSource)

Let’s walk through this code step by step:

Note the method register where you need to bind the name in the track specification to your parameter source implementation class similar to the simple example.
The class TermParamSource is the actual parameter source and needs to fulfill a few requirements:
- It needs to have a constructor with the signature __init__(self, track, params, **kwargs). You don’t need to store these parameters if you don’t need them.
- partition(self, partition_index, total_partitions) is called by Rally to “assign” the parameter source across multiple clients. Typically you can just return self but in certain cases you need to do something more sophisticated. If each clients needs to act differently then you can provide different parameter source instances here.
- size(self): This method is needed to help Rally provide a proper progress indication to users if you use a warmup time period. For bulk indexing, this would return the number of bulks (for a given client). As searches are typically executed with a pre-determined amount of iterations, just return 1 in this case.
- params(self): This method needs to return a dictionary with all parameters that the corresponding “runner” expects. For the standard case, Rally provides most of these parameters as a convenience, but here you need to define all of them yourself. This method will be invoked once for every iteration during the race. We can see that we randomly select a profession from a list which will be then be executed by the corresponding runner.

For cases, where you want to provide a progress indication but cannot calculate size up-front (e.g. when you generate bulk requests on-the fly up to a certain total size), you can implement a property percent_completed which returns a floating point value between 0.0 and 1.0. Rally will query this value before each call to params() and uses it for its progress indication. For the implementation, please notes:

Rally will not check percent_completed, if it can derive progress in any other way.
The value of percent_completed is purely informational and has no influence whatsoever on when Rally considers an operation to be completed.

Note

Be aware that params(self) is called on a performance-critical path so don’t do anything in this method that takes a lot of time (avoid any I/O). For searches, you should usually throttle throughput anyway and there it does not matter that much but if the corresponding operation is run without throughput throttling, please double-check that you did not introduce a bottleneck in the load test driver with your custom parameter source.

In the implementation of custom parameter sources you can access the Python standard API. Using any additional libraries is not supported.

You can also implement your parameter sources and runners in multiple Python files but the main entry point is always track.py. The root package name of your plugin is the name of your track.

Custom runners¶

Warning

Your runner is on a performance-critical code-path so please double-check with Rally’s profiling support that you did not introduce any bottlenecks.

You cannot only define custom parameter sources but also custom runners. Runners execute an operation against Elasticsearch. Out of the box, Rally supports the following operations:

Bulk indexing
Force merge
Searches
Index stats
Nodes stats

If you want to use any other operation, you can define a custom runner. Consider, we want to use the percolate API with an older version of Elasticsearch (note that it has been replaced by the percolate query in Elasticsearch 5.0). To achieve this, we c

In track.json specify an operation with type “percolate” (you can choose this name freely):

{
  "name": "percolator_with_content_google",
  "operation-type": "percolate",
  "body": {
    "doc": {
      "body": "google"
    },
    "track_scores": true
  }
}

Then create a file track.py next to track.json and implement the following two functions:

def percolate(es, params):
    es.percolate(
        index="queries",
        doc_type="content",
        body=params["body"]
    )


def register(registry):
    registry.register_runner("percolate", percolate)

The function percolate is the actual runner and takes the following parameters:

es, which is the Elasticsearch Python client
params which is a dict of parameters provided by its corresponding parameter source. Treat this parameter as read only and do not attempt to write to it.

This function can return either:

Nothing at all. Then Rally will assume that by default 1 and "ops" (see below)
A tuple of weight and a unit, which is usually 1 and "ops". If you run a bulk operation you might return the bulk size here, for example in number of documents or in MB. Then you’d return for example (5000, "docs") Rally will use these values to store throughput metrics.
A dict with arbitrary keys. If the dict contains the key weight it is assumed to be numeric and chosen as weight as defined above. The key unit is treated similarly. All other keys are added to the meta section of the corresponding service time and latency metrics records.

Similar to a parameter source you also need to bind the name of your operation type to the function within register.

If you need more control, you can also implement a runner class. The example above, implemented as a class looks as follows:

class PercolateRunner:
    def __enter__(self):
        return self

    def __call__(self, es, params):
        es.percolate(
            index="queries",
            doc_type="content",
            body=params["body"]
        )

    def __repr__(self, *args, **kwargs):
        return "percolate"

def register(registry):
    registry.register_runner("percolate", PercolateRunner())

The actual runner is implemented in the method __call__ and the same return value conventions apply as for functions. For debugging purposes you should also implement __repr__ and provide a human-readable name for your runner. Finally, you need to register your runner in the register function. Runners also support Python’s context manager interface. Rally uses a new context for each request. Implementing the context manager interface can be handy for cleanup of resources after executing an operation. Rally uses it for example to clear open scrolls.

Note

You need to implement register just once and register all parameter sources and runners there.

Custom schedulers¶

Warning

Your scheduler is on a performance-critical code-path so please double-check with Rally’s profiling support that you did not introduce any bottlenecks.

If you want to rate-limit execution of tasks, you can specify a target-throughput (in operations per second). For example, Rally will attempt to run this term query 20 times per second:

{
  "operation": "term",
  "target-throughput": 20
}

By default, Rally will use a deterministic distribution to determine when to schedule the next operation. This means, that it will execute the term query at 0, 50ms, 100ms, 150ms and so on. Note that the scheduler is aware of the number of clients. Consider this example:

{
  "operation": "term",
  "target-throughput": 20,
  "clients": 4
}

If Rally would not take the number of clients into account and would still issue requests (from each of the four clients) at the same points in time (i.e. 0, 50ms, 100ms, 150ms, …), it would run at a target throughput of 4 * 20 = 80 operations per second. Hence, Rally will automatically reduce the rate at which each client will execute requests. Each client will issue requests at 0, 200ms, 400ms, 600ms, 800ms, 1000ms and so on. Each client issues five requests per second but as there are four of them, we still have a target throughput of 20 operations per second. You should keep this in mind, when writing your own custom schedules.

If you want to create a custom scheduler, create a file track.py next to track.json and implement the following two functions:

import random

def random_schedule(current):
    return current + random.randint(10, 900) / 1000.0


def register(registry):
    registry.register_scheduler("my_random", random_schedule)

You can then use your custom scheduler as follows:

{
  "operation": "term",
  "schedule": "my_random"
}

The function random_schedule returns a floating point number which represents the next point in time when Rally should execute the given operation. This point in time is measured in seconds relative to the beginning of the execution of this task. The parameter current is the last return value of your function and is 0 for the first invocation. So, for example, this scheduler could return the following series: 0, 0.119, 0.622, 1.29, 1.343, 1.984, 2.233. Note that this implementation is usually not sufficient as it does not take into account the number of clients. Therefore, you will typically want to implement a full-blown scheduler which can also take parameters. Below is an example for our random scheduler:

import random

class RandomScheduler:
    def __init__(self, params):
        # assume one client by default
        clients = self.params.get("clients", 1)
        # scale accordingly with the number of clients!
        self.lower_bound = clients * self.params.get("lower-bound-millis", 10)
        self.upper_bound = clients * self.params.get("upper-bound-millis", 900)

    def next(self, current):
        return current + random.randint(self.lower_bound, self.upper_bound) / 1000.0


def register(registry):
    registry.register_scheduler("my_random", RandomScheduler)

This implementation will now achieve the same rate independent of the number of clients. Additionally, we can pass the lower and upper bound for the random function from our track:

{
    "operation": "term",
    "schedule": "my_random",
    "clients": 4,
    "lower-bound-millis": 50,
    "upper-bound-millis": 250
}

Developing Rally¶

Prerequisites¶

Please ensure that the following packages are installed before installing Rally in development mode:

Python 3.4 or better available as python3 on the path (verify with: python3 --version)
pip3 available on the path (verify with pip3 --version)
JDK 8 or 9
git 1.9 or better
Gradle 3.3 or better

Please check the installation guide for detailed installation instructions for these packages.

Rally does not support Windows and is only actively tested on MacOS and Linux.

Installation Instructions for Development¶

git clone https://github.com/elastic/rally.git
cd rally
./rally

If you get errors during installation, it is probably due to the installation of psutil which we use to gather system metrics like CPU utilization. Please check the installation instructions of psutil in this case. Keep in mind that Rally is based on Python 3 and you need to install the Python 3 header files instead of the Python 2 header files on Linux.

Automatic Updates¶

Rally has a built-in auto-update feature when you install it from sources. By default, it will update from the remote named origin. If you want to auto-update from a different remote, provide --update-from-remote=YOUR_REMOTE_NAME as first parameter.

To work conveniently with Rally, we suggest that you add the Rally project directory to your PATH. In case you use a different remote, you should also define aliases in your shell’s config file, e.g.:

alias rally='rally --update-from-remote=elastic '
alias rallyd='rallyd --update-from-remote=elastic '

Then you can invoke Rally or the Rally daemon as usual and have auto-update still work.

Also note that automatic updates are disabled in the following cases:

There are local (uncommitted) changes in the Rally project directory
A different branch than master is checked out
You have specified --skip-update as the first command line parameter
You have specified --offline as a command line parameter for Rally

Configuring Rally¶

Before we can run our first benchmark, we have to configure Rally. Just invoke ./rally configure and Rally will automatically detect that its configuration file is missing and prompt you for some values and write them to ~/.rally/rally.ini. After you’ve configured Rally, it will exit.

For more information see configuration help page.

Key Components of Rally¶

To get a rough understanding of Rally, it makes sense to get to know its key components:

Race Control: is responsible for proper execution of the race. It sets up all components and acts as a high-level controller.
Mechanic: can build and prepare a benchmark candidate for the race. It checks out the source, builds Elasticsearch, provisions and starts the cluster.
Track: is a concrete benchmarking scenario, e.g. the http_logs benchmark. It defines the data set to use.
Challenge: is the specification on what benchmarks should be run and its configuration (e.g. index, then run a search benchmark with 1000 iterations)
Car: is a concrete system configuration for a benchmark, e.g. an Elasticsearch single-node cluster with default settings.
Driver: drives the race, i.e. it is executing the benchmark according to the track specification.
Reporter: A reporter tells us how the race went (currently only after the fact).

There is a dedicated tutorial on how to add new tracks to Rally.

How to contribute code¶

First of all, please read the contributors guide.

We strive to be PEP-8 compliant but don’t follow it to the letter.

Command Line Reference¶

You can control Rally with subcommands and command line flags:

Subcommands determine which task Rally performs.
Command line flags are used to change Rally’s behavior but not all command line flags can be used for each subcommand. To find out which command line flags are supported by a specific subcommand, just run esrally <<subcommand>> --help.

Subcommands¶

`race`¶

The race subcommand is used to actually run a benchmark. It is the default one and chosen implicitly if none is given.

`list`¶

The list subcommand is used to list different configuration options:

telemetry: Will show all telemetry devices that are supported by Rally.
tracks: Will show all tracks that are supported by Rally. As this may depend on the Elasticsearch version that you want to benchmark, you can specify --distribution-version and also --distribution-repository as additional options.
pipelines: Will show all pipelines that are supported by Rally.
races: Will show a list of the most recent races. This is needed for the tournament mode.
cars: Will show all cars that are supported by Rally (i.e. Elasticsearch configurations).
elasticsearch-plugins: Will show all Elasticsearch plugins and their configurations that are supported by Rally.

To list a specific configuration option, place it after the list subcommand. For example, esrally list pipelines will list all pipelines known to Rally.

`compare`¶

This subcommand is needed for tournament mode and its usage is described there.

`configure`¶

This subcommand is needed to configure Rally. It is implicitly chosen if you start Rally for the first time but you can rerun this command at any time.

Command Line Flags¶

`track-path`¶

Can be either a directory that contains a track.json file or a .json file with an arbitrary name that contains a track specification. --track-path and --track-repository as well as --track are mutually exclusive. See the track reference to decide whether you should use --track-path or --track-repository / --track.

Examples:

# provide a directory - Rally searches for a track.json file in this directory
# Track name is "app-logs"
esrally --track-path=~/Projects/tracks/app-logs
# provide a file name - Rally uses this file directly
# Track name is "syslog"
esrally --track-path=~/Projects/tracks/syslog.json

`track-repository`¶

Selects the track repository that Rally should use to resolve tracks. By default the default track repository is used, which is available in the Github project rally-tracks. See the track reference on how to add your own track repositories. --track-path and --track-repository as well as --track are mutually exclusive.

`track`¶

Selects the track that Rally should run. By default the geonames track is run. For more details on how tracks work, see adding tracks or the track reference. --track-path and --track-repository as well as --track are mutually exclusive.

`track-params`¶

With this parameter you can inject variables into tracks. The supported variables depend on the track and you should check the track JSON file to see which variables can be provided.

It accepts a list of comma-separated key-value pairs. The key-value pairs have to be delimited by a colon.

Examples:

Consider the following track snippet showing a single challenge:

{
  "name": "index-only",
  "index-settings": {
    "index.number_of_replicas": {{ replica_count|default(0) }},
    "index.number_of_shards": {{ shard_count|default(5) }},
  },
  "schedule": [
    {
      "operation": "bulk-index",
      "warmup-time-period": 120,
      "clients": 8
    }
  ]
}

Rally tracks can use the Jinja templating language and the construct {{ some_variable|default(0) }} that you can see above is a feature of Jinja to define default values for variables.

We can see that it defines two variables:

replica_count with a default value of 0
shard_count with a default value of 5

When we run this track, we can override these defaults:

--track-params="replica_count:1,shard_count:3" will set the number of replicas to 1 and the number of shards to 3.
--track-params="replica_count:1" will just set the number of replicas to 1 and just keep the default value of 5 shards.

All track parameters are recorded for each metrics record in the metrics store. Also, when you run esrally list races, it will show all track parameters:

Race Timestamp    Track    Track Parameters               Challenge            Car       User Tag
----------------  -------  ------------------------------ -------------------  --------  ---------
20160518T122341Z  pmc      replica_count=1                append-no-conflicts  defaults
20160518T112341Z  pmc      replica_count=1,shard_count=3  append-no-conflicts  defaults

Note that the default values are not recorded or shown (Rally does not know about them).

`challenge`¶

A track consists of one or more challenges. With this flag you can specify which challenge should be run. If you don’t specify a challenge, Rally derives the default challenge itself. To see the default challenge of a track, run esrally list tracks.

`include-tasks`¶

Each challenge consists of one or more tasks but sometimes you are only interested to run a subset of all tasks. For example, you might have prepared an index already and want only to repeatedly run search benchmarks. Or you want to run only the indexing task but nothing else.

You can use --include-tasks to specify a comma-separated list of tasks that you want to run. Each item in the list defines either the name of a task or the operation type of a task. Only the tasks that match will be executed. Currently there is also no command that lists the tasks of a challenge so you need to look at the track source.

Examples:

Execute only the tasks with the name index and term: --include-tasks="index,term"
Execute only tasks of type search: --include-tasks="type:search"
You can also mix and match: --include-tasks="index,type:search"

`team-repository`¶

Selects the team repository that Rally should use to resolve cars. By default the default team repository is used, which is available in the Github project rally-teams. See the documentation about cars on how to add your own team repositories.

`car`¶

A car defines the Elasticsearch configuration that will be used for the benchmark. To see a list of possible cars, issue esrally list cars. You can specify one or multiple comma-separated values, for details please refer to the documentation about cars.

Example

esrally --car="4gheap,ea"

Rally will configure Elasticsearch with 4GB of heap (4gheap) and enable Java assertions (ea).

`car-params`¶

Allows to override config variables of Elasticsearch.

Example

esrally --car="4gheap" --car-params="data_paths:'/opt/elasticsearch'"

The variables that are exposed depend on the car’s configuration. In addition, Rally implements special handling for the variable data_paths (by default the values for variable is determined by Rally).

`elasticsearch-plugins`¶

A comma-separated list of Elasticsearch plugins to install for the benchmark. If a plugin supports multiple configurations you need to specify the configuration after the plugin name. To see a list of possible plugins and configurations, issue esrally list elasticsearch-plugins.

Example:

esrally --elasticsearch-plugins="analysis-icu,xpack:security"

In this example, Rally will install the analysis-icu plugin and the x-pack plugin with the security configuration. See the reference documentation about Elasticsearch plugins for more details.

`plugin-params`¶

Allows to override variables of Elasticsearch plugins.

Example:

esrally --distribution-version=6.1.1. --elasticsearch-plugins="x-pack:monitoring-http" --plugin-params="monitoring_type:'https',monitoring_host:'some_remote_host',monitoring_port:10200,monitoring_user:'rally',monitoring_password:'m0n1t0r1ng'"

This enables the HTTP exporter of X-Pack Monitoring and exports the data to the configured monitoring host.

`pipeline`¶

Selects the pipeline that Rally should run.

Rally can autodetect the pipeline in most cases. If you specify --distribution-version it will auto-select the pipeline from-distribution otherwise it will use from-sources-complete.

`laps`¶

Allows to run the benchmark for multiple laps (defaults to 1 lap). Each lap corresponds to one full execution of a track but note that the benchmark candidate is not restarted in between.

`enable-driver-profiling`¶

This option enables a profiler on all tasks that the load test driver performs. It is intended to help track authors spot accidental bottlenecks, especially if they implement their own runners or parameter sources. When this mode is enabled, Rally will enable a profiler in the load driver module. After each task and for each client, Rally will add the profile information to a dedicated profile log file. For example:

2017-02-09 08:23:24,35 rally.profile INFO
=== Profile START for client [0] and task [index-append-1000] ===
   16052402 function calls (15794402 primitive calls) in 180.221 seconds

   Ordered by: cumulative time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
      130    0.001    0.000  168.089    1.293 /Users/dm/Projects/rally/esrally/driver/driver.py:908(time_period_based)
      129    0.260    0.002  168.088    1.303 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/bulk_source.py:79(params)
   129000    0.750    0.000  167.791    0.001 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/randomevent.py:142(generate_event)
   516000    0.387    0.000  160.485    0.000 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/weightedarray.py:20(get_random)
   516000    6.199    0.000  160.098    0.000 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/weightedarray.py:23(__random_index)
   516000    1.292    0.000  152.289    0.000 /usr/local/Cellar/python3/3.6.0/Frameworks/Python.framework/Versions/3.6/lib/python3.6/random.py:96(seed)
   516000  150.783    0.000  150.783    0.000 {function Random.seed at 0x10b7fa2f0}
   129000    0.363    0.000   45.686    0.000 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/randomevent.py:48(add_fields)
   129000    0.181    0.000   41.742    0.000 /Users/dm/.rally/benchmarks/tracks/develop/bottleneck/parameter_sources/randomevent.py:79(add_fields)
   ....

=== Profile END for client [0] and task [index-append-1000] ===

In this example we can spot quickly that Random.seed is called excessively, causing an accidental bottleneck in the load test driver.

`test-mode`¶

Allows you to test a track without running it for the whole duration. This mode is only intended for quick sanity checks when creating a track. Please don’t rely on these numbers at all (they are meaningless).

If you write your own track, please keep in mind that you need prepare your track to support this mode.

`telemetry`¶

Activates the provided telemetry devices for this race.

Example

esrally --telemetry=jfr,jit

This activates Java flight recorder and the JIT compiler telemetry devices.

`revision`¶

If you actively develop Elasticsearch and want to benchmark a source build of Elasticsearch (which will Rally create for you), you can specify the git revision of Elasticsearch that you want to benchmark. But note that Rally does only support Gradle as build tool which effectively means that it will only support this for Elasticsearch 5.0 or better. The default value is current.

You can specify the revision in different formats:

--revision=latest: Use the HEAD revision from origin/master.
--revision=current: Use the current revision (i.e. don’t alter the local source tree).
--revision=abc123: Where abc123 is some git revision hash.
--revision=@2013-07-27T10:37:00Z: Determines the revision that is closest to the provided date. Rally logs to which git revision hash the date has been resolved and if you use Elasticsearch as metrics store (instead of the default in-memory one), each metric record will contain the git revision hash also in the meta-data section.

Supported date format: If you specify a date, it has to be ISO-8601 conformant and must start with an @ sign to make it easier for Rally to determine that you actually mean a date.

If you want to create source builds of Elasticsearch plugins, you need to specify the revision for Elasticsearch and all relevant plugins separately. Revisions for Elasticsearch and each plugin need to be comma-separated (,). Each revision is prefixed either by elasticsearch or by the plugin name and separated by a colon (:). As core plugins are contained in the Elasticsearch repo, there is no need to specify a revision for them (the revision would even be ignored in fact).

Examples:

Build latest Elasticsearch and plugin “my-plugin”: --revision="elasticsearch:latest,my-plugin:latest"
Build Elasticsearch tag v5.6.1 and revision abc123 of plugin “my-plugin”: --revision="elasticsearch:v5.6.1,my-plugin:abc123"

Note that it is still required to provide the parameter --elasticsearch-plugins. Specifying a plugin with --revision just tells Rally which revision to use for building the artifact. See the documentation on Elasticsearch plugins for more details.

`distribution-version`¶

If you want to benchmark a binary distribution, you can specify the version here.

Example

esrally --distribution-version=2.3.3

Rally will then benchmark the official Elasticsearch 2.3.3 distribution.

Rally works with all releases of Elasticsearch that are supported by Elastic.

The following versions are already end-of-life:

0.x: Rally is not tested, and not expected to work for this version; we will make no effort to make Rally work.
1.x: Rally works on a best-effort basis with this version but support may be removed at any time.

Additionally, Rally will always work with the current development version of Elasticsearch (by using either a snapshot repository or by building Elasticsearch from sources).

`distribution-repository`¶

Rally does not only support benchmarking official distributions but can also benchmark snapshot builds. This is option is really just intended for our benchmarks that are run in continuous integration but if you want to, you can use it too. The only supported value out of the box is release (default) but you can define arbitrary repositories in ~/.rally/rally.ini.

Example

Say, you have an in-house repository where Elasticsearch snapshot builds get published. Then you can add the following in the distributions section of your Rally config file:

in_house_snapshot.url = https://www.example.org/snapshots/elasticsearch/elasticsearch-{{VERSION}}.tar.gz
in_house_snapshot.cache = false

The url property defines the URL pattern for this repository. The cache property defines whether Rally should always download a new archive (cache=false) or just reuse a previously downloaded version (cache=true). Rally will replace the {{VERSION}} placeholder of in the url property with the value of distribution-version provided by the user on the command line.

You can use this distribution repository with the name “in_house_snapshot” as follows:

esrally --distribution-repository=in_house_snapshot --distribution-version=7.0.0-SNAPSHOT

This will benchmark the latest 7.0.0 snapshot build of Elasticsearch.

`report-format`¶

The command line reporter in Rally displays a table with key metrics after a race. With this option you can specify whether this table should be in markdown format (default) or csv.

`show-in-report`¶

By default, the command line reporter will only show values that are available (available). With all you can force it to show a line for every value, even undefined ones, and with all-percentiles it will show only available values but force output of all possible percentile values.

This command line parameter is not available for comparing races.

`report-file`¶

By default, the command line reporter will print the results only on standard output, but can also write it to a file.

Example

esrally --report-format=csv --report-file=~/benchmarks/result.csv

`client-options`¶

With this option you can customize Rally’s internal Elasticsearch client.

It accepts a list of comma-separated key-value pairs. The key-value pairs have to be delimited by a colon. These options are passed directly to the Elasticsearch Python client API. See their documentation on a list of supported options.

We support the following data types:

Strings: Have to be enclosed in single quotes. Example: ca_certs:'/path/to/CA_certs'
Numbers: There is nothing special about numbers. Example: sniffer_timeout:60
Booleans: Specify either true or false. Example: use_ssl:true

In addition to the options, supported by the Elasticsearch client, it is also possible to enable HTTP compression by specifying compressed:true

Default value: timeout:60

Warning

If you provide your own client options, the default value will not be magically merged. You have to specify all client options explicitly. The only exceptions to this rule is ca_cert (see below).

Examples

Here are a few common examples:

Enable HTTP compression: --client-options="compressed:true"
Enable SSL (e.g. if you have X-Pack Security installed): --client-options="use_ssl:true,verify_certs:true". Note that you don’t need to set ca_cert (which defines the path to the root certificates). Rally does this automatically for you.
Enable SSL with a client key and certificate: --client-options="use_ssl:true,verify_certs:true,ca_certs:'/path/to/cacert.pem',client_cert:'/path/to/client_cert.pem',client_key='/path/to/client_key.pem" (see also the Elasticsearch Python client docs)
Enable basic authentication: --client-options="basic_auth_user:'user',basic_auth_password:'password'". Please avoid the characters ', , and : in user name and password as Rally’s parsing of these options is currently really simple and there is no possibility to escape characters.

`on-error`¶

This option controls whether Rally will continue or abort when a request error occurs. By default, Rally will just record errors and report the error rate at the end of a race. With --on-error=abort, Rally will immediately abort the race on the first error and print a detailed error message.

`load-driver-hosts`¶

By default, Rally will run its load driver on the same machine where you start the benchmark. However, if you benchmark larger clusters, one machine may not be enough to generate sufficient load. Hence, you can specify a comma-separated list of hosts which should be used to generate load with --load-driver-hosts.

Example

esrally --load-driver-hosts=10.17.20.5,10.17.20.6

In the example, above Rally will generate load from the hosts 10.17.20.5 and 10.17.20.6. For this to work, you need to start a Rally daemon on these machines, see distributing the load test driver for a complete example.

`target-hosts`¶

If you run the benchmark-only pipeline or you want Rally to benchmark a remote cluster, then you can specify a comma-delimited list of hosts:port pairs to which Rally should connect. The default value is 127.0.0.1:9200.

Example

esrally --pipeline=benchmark-only --target-hosts=10.17.0.5:9200,10.17.0.6:9200

This will run the benchmark against the hosts 10.17.0.5 and 10.17.0.6 on port 9200. See client-options if you use X-Pack Security and need to authenticate or Rally should use https.

`quiet`¶

Suppresses some output on the command line.

`offline`¶

Tells Rally that it should assume it has no connection to the Internet when checking for track data. The default value is false. Note that Rally will only assume this for tracks but not for anything else, e.g. it will still try to download Elasticsearch distributions that are not locally cached or fetch the Elasticsearch source tree.

`preserve-install`¶

Rally usually installs and launches an Elasticsearch cluster internally and wipes the entire directory after the benchmark is done. Sometimes you want to keep this cluster including all data after the benchmark has finished and that’s what you can do with this flag. Note that depending on the track that has been run, the cluster can eat up a very significant amount of disk space (at least dozens of GB). The default value is configurable in the advanced configuration but usually false.

Note

This option does only affect clusters that are provisioned by Rally. More specifically, if you use the pipeline benchmark-only, this option is ineffective as Rally does not provision a cluster in this case.

`cluster-health`¶

Warning

This option is deprecated and will be removed in a future version of Rally. For details, please see the respective Github ticket #364.

Rally checks whether the cluster health is “green” before it runs a benchmark against it. The main reason is that we don’t want to benchmark a cluster which is shuffling shards around or might start doing so. If you really need to run a benchmark against a cluster that is “yellow” or “red”, then you can explicitly override Rally’s default behavior. It is even possible to skip this check entirely by providing --cluster-health=skip. But please think twice before doing so and rather eliminate the root cause.

Example

esrally --cluster-health=yellow

`advanced-config`¶

This flag determines whether Rally should present additional (advanced) configuration options. The default value is false.

Example

esrally configure --advanced-config

`assume-defaults`¶

This flag determines whether Rally should automatically accept all values for configuration options that provide a default. This is mainly intended to configure Rally automatically in CI runs. The default value is false.

Example

esrally configure --assume-defaults=true

`user-tag`¶

This is only relevant when you want to run tournaments. You can use this flag to attach an arbitrary text to the meta-data of each metric record and also the corresponding race. This will help you to recognize a race when you run esrally list races as you don’t need to remember the concrete timestamp on which a race has been run but can instead use your own descriptive names.

The required format is key “:” value. You can choose key and value freely.

Example

esrally --user-tag="intention:github-issue-1234-baseline,gc:cms"

You can also specify multiple tags. They need to be separated by a comma.

Example

esrally --user-tag="disk:SSD,data_node_count:4"

When you run esrally list races, this will show up again:

Race Timestamp    Track    Track Parameters   Challenge            Car       User Tag
----------------  -------  ------------------ -------------------  --------  ------------------------------------
20160518T122341Z  pmc                         append-no-conflicts  defaults  intention:github-issue-1234-baseline
20160518T112341Z  pmc                         append-no-conflicts  defaults  disk:SSD,data_node_count:4

This will help you recognize a specific race when running esrally compare.

Offline Usage¶

In some corporate environments servers do not have Internet access. You can still use Rally in such environments and this page summarizes all information that you need to get started.

Installation and Configuration¶

We provide a special offline installation package. Please refer to the offline installation guide for detailed instructions. After the installation you can just follow the normal configuration procedure.

Command Line Usage¶

Rally will automatically detect upon startup that no Internet connection is available and print the following warning:

[WARNING] No Internet connection detected. Automatic download of track data sets etc. is disabled.

It detects this by trying to connect to github.com. If you want to disable this probing you can explicitly specify --offline.

Using tracks¶

A Rally track describes a benchmarking scenario. You can either write your own tracks or use the tracks that Rally provides out of the box. In the former case, Rally will work just fine in an offline environment. In the latter case, Rally would normally download the track and its associated data from the Internet. If you want to use one of Rally’s standard tracks in offline mode, you need to download all relevant files first on a machine that has Internet access and copy it to the target machine(s).

Use the download script to download all data for a track on a machine that has access to the Internet. Example:

# downloads the script from Github
curl -O https://raw.githubusercontent.com/elastic/rally-tracks/master/download.sh
chmod u+x download.sh
# download all data for the geonames track
./download.sh geonames

This will download all data for the geonames track and create a tar file rally-track-data-geonames.tar in the current directory. Copy this file to the home directory of the user which will execute Rally on the target machine (e.g. /home/rally-user).

On the target machine, run:

cd ~
tar -xf rally-track-data-geonames.tar

The download script does not require a Rally installation on the machine with Internet access but assumes that git and curl are available.

After you’ve copied the data, you can list the available tracks with esrally list tracks. If a track shows up in this list, it just means that the track description is available locally but not necessarily all data files.

Using cars¶

Note

You can skip this section if you use Rally only as a load generator.

If you have Rally configure and start Elasticsearch then you also need the out-of-the-box configurations available. Run the following command on a machine with Internet access:

git clone https://github.com/elastic/rally-teams.git ~/.rally/benchmarks/teams/default
tar -C ~ -czf rally-teams.tar.gz .rally/benchmarks/teams/default

Copy that file to the target machine(s) and run on the target machine:

cd ~
tar -xzf rally-teams.tar.gz

After you’ve copied the data, you can list the available tracks with esrally list cars.

Track Reference¶

Definition¶

A track is a specification of one or more benchmarking scenarios with a specific document corpus. It defines for example the involved indices, data files and the operations that are invoked. Its most important attributes are:

One or more indices, each with one or more types
The queries to issue
Source URL of the benchmark data
A list of steps to run, which we’ll call “challenge”, for example indexing data with a specific number of documents per bulk request or running searches for a defined number of iterations.

Track File Format and Storage¶

A track is specified in a JSON file.

Ad-hoc use¶

For ad-hoc use you can store a track definition anywhere on the file system and reference it with --track-path, e.g:

# provide a directory - Rally searches for a track.json file in this directory
# Track name is "app-logs"
esrally --track-path=~/Projects/tracks/app-logs
# provide a file name - Rally uses this file directly
# Track name is "syslog"
esrally --track-path=~/Projects/tracks/syslog.json

Rally will also search for additional files like mappings or data files in the provided directory. If you use advanced features like custom runners or parameter sources we recommend that you create a separate directory per track.

Custom Track Repositories¶

Alternatively, you can store Rally tracks also in a dedicated git repository which we call a “track repository”. Rally provides a default track repository that is hosted on Github. You can also add your own track repositories although this requires a bit of additional work. First of all, track repositories need to be managed by git. The reason is that Rally can benchmark multiple versions of Elasticsearch and we use git branches in the track repository to determine the best match for each track (based on the command line parameter --distribution-version). The versioning scheme is as follows:

The master branch needs to work with the latest master branch of Elasticsearch.
All other branches need to match the version scheme of Elasticsearch, i.e. MAJOR.MINOR.PATCH-SUFFIX where all parts except MAJOR are optional.

Rally implements a fallback logic so you don’t need to define a branch for each patch release of Elasticsearch. For example:

The branch 6.0.0-alpha1 will be chosen for the version 6.0.0-alpha1 of Elasticsearch.
The branch 5 will be chosen for all versions for Elasticsearch with the major version 5, e.g. 5.0.0, 5.1.3 (provided there is no specific branch).

Rally tries to use the branch with the best match to the benchmarked version of Elasticsearch.

Rally will also search for related files like mappings or custom runners or parameter sources in the track repository. However, Rally will use a separate directory to look for data files (~/.rally/benchmarks/data/$TRACK_NAME/). The reason is simply that we do not want to check multi-GB data files into git.

Creating a new track repository¶

All track repositories are located in ~/.rally/benchmarks/tracks. If you want to add a dedicated track repository, called private follow these steps:

cd ~/.rally/benchmarks/tracks
mkdir private
cd private
git init
# add your track now
git add .
git commit -m "Initial commit"

If you want to share your tracks with others you need to add a remote and push it:

git remote add origin git@git-repos.acme.com:acme/rally-tracks.git
git push -u origin master

If you have added a remote you should also add it in ~/.rally/rally.ini, otherwise you can skip this step. Open the file in your editor of choice and add the following line in the section tracks:

private.url = <<URL_TO_YOUR_ORIGIN>>

Rally will then automatically update the local tracking branches before the benchmark starts.

You can now verify that everything works by listing all tracks in this track repository:

esrally list tracks --track-repository=private

This shows all tracks that are available on the master branch of this repository. Suppose you only created tracks on the branch 2 because you’re interested in the performance of Elasticsearch 2.x, then you can specify also the distribution version:

esrally list tracks --track-repository=private --distribution-version=2.0.0

Rally will follow the same branch fallback logic as described above.

Adding an already existing track repository¶

If you want to add a track repository that already exists, just open ~/.rally/rally.ini in your editor of choice and add the following line in the section tracks:

your_repo_name.url = <<URL_TO_YOUR_ORIGIN>>

After you have added this line, have Rally list the tracks in this repository:

esrally list tracks --track-repository=your_repo_name

When to use what?¶

We recommend the following path:

Start with a simple json file. The file name can be arbitrary.
If you need custom runners or parameter sources, create one directory per track. Then you can keep everything that is related to one track in one place. Remember that the track JSON file needs to be named track.json.
If you want to version your tracks so they can work with multiple versions of Elasticsearch (e.g. you are running benchmarks before an upgrade), use a track repository.

Anatomy of a track¶

A track JSON file consists of the following sections:

indices
templates
corpora
operations
challenges

In the indices and templates sections you define the relevant indices and index templates. These sections are optional but recommended if you want to create indices and index templates with the help of Rally.

In the corpora section you define all document corpora (i.e. data files) that Rally should use for this track.

In the operations section you describe which operations are available for this track and how they are parametrized. This section is optional and you can also define any operations directly per challenge. You can use it, if you want to share operation definitions between challenges.

In the challenge or challenges section you describe one or more execution schedules respectively. Each schedule either uses the operations defined in the operations block or defines the operations to execute inline. Think of a challenge as a scenario that you want to test for your data. An example challenge is to index with two clients at maximum throughput while searching with another two clients with ten operations per second.

Track elements¶

The track elements that are described here are defined in Rally’s JSON schema for tracks. Rally uses this track schema to validate your tracks when it is loading them.

Each track defines the following info attributes:

version (optional): An integer describing the track specification version in use. Rally uses it to detect incompatible future track specification versions and raise an error. See the table below for a reference of valid versions.
description (optional): A human-readable description of the track. Although it is optional, we recommend providing it.

Track Specification Version	Rally version
1	>=0.7.3, <0.10.0
2	>=0.9.0

The version property has been introduced with Rally 0.7.3. Rally versions before 0.7.3 do not recognize this property and thus cannot detect incompatible track specification versions.

Example:

{
    "version": 1,
    "description": "POIs from Geonames"
}

meta¶

For each track, an optional structure, called meta can be defined. You are free which properties this element should contain.

This element can also be defined on the following elements:

challenge
operation
task

If the meta structure contains the same key on different elements, more specific ones will override the same key of more generic elements. The order from generic to most specific is:

track
challenge
operation
task

E.g. a key defined on a task, will override the same key defined on a challenge. All properties defined within the merged meta structure, will get copied into each metrics record.

indices¶

The indices section contains a list of all indices that are used by this track. By default Rally will assume that it can destroy and create these indices at will.

Each index in this list consists of the following properties:

name (mandatory): The name of the index.
body (optional): File name of the corresponding index definition that will be used as body in the create index API call.
auto-managed (optional, defaults to true): Controls whether Rally or the user takes care of creating / destroying the index. If this setting is false, Rally will neither create nor delete this index but just assume its presence.
types (optional): A list of type names in this index.

Example:

"indices": [
    {
      "name": "geonames",
      "body": "geonames-index.json",
      "types": ["docs"]
    }
]

templates¶

The templates section contains a list of all index templates that Rally should create.

name (mandatory): Index template name
index-pattern (mandatory): Index pattern that matches the index template. This must match the definition in the index template file.
delete-matching-indices (optional, defaults to true): Delete all indices that match the provided index pattern before start of the benchmark.
template (mandatory): Index template file name

Example:

"templates": [
    {
        "name": "my-default-index-template",
        "index-pattern": "my-index-*",
        "delete-matching-indices": true,
        "template": "default-template.json"
    }
]

corpora¶

The corpora section contains all document corpora that are used by this track. Note that you can reuse document corpora across tracks; just copy & paste the respective corpora definitions. It consists of the following properties:

name (mandatory): Name of this document corpus. As this name is also used by Rally in directory names, it is recommended to only use lower-case names without whitespaces for maximum compatibility across file systems.
documents (mandatory): A list of documents files.

Each entry in the documents list consists of the following properties:

base-url (optional): A http or https URL that points to the root path where Rally can obtain the corresponding source file.
source-format (optional, default: bulk): Defines in which format Rally should interpret the data file specified by source-file. Currently, only bulk is supported.
source-file (mandatory): File name of the corresponding documents. For local use, this file can be a .json file. If you provide a base-url we recommend that you provide a compressed file here. The following extensions are supported: .zip, .bz2, .gz, .tar, .tar.gz, .tgz or .tar.bz2. It must contain exactly one JSON file with the same name. The preferred file extension for our official tracks is .bz2.
includes-action-and-meta-data (optional, defaults to false): Defines whether the documents file contains already an action and meta-data line (true) or only documents (false).
document-count (mandatory): Number of documents in the source file. This number is used by Rally to determine which client indexes which part of the document corpus (each of the N clients gets one N-th of the document corpus). If you are using parent-child, specify the number of parent documents.
compressed-bytes (optional but recommended): The size in bytes of the compressed source file. This number is used to show users how much data will be downloaded by Rally and also to check whether the download is complete.
uncompressed-bytes (optional but recommended): The size in bytes of the source file after decompression. This number is used by Rally to show users how much disk space the decompressed file will need and to check that the whole file could be decompressed successfully.
target-index: Defines the name of the index which should be targeted for bulk operations. Rally will automatically derive this value if you have defined exactly one index in the indices section. Ignored if includes-action-and-meta-data is true.
target-type: Defines the name of the document type which should be targeted for bulk operations. Rally will automatically derive this value if you have defined exactly one index in the indices section and this index has exactly one type. Ignored if includes-action-and-meta-data is true.

To avoid repetition, you can specify default values on document corpus level for the following properties:

base-url
source-format
includes-action-and-meta-data
target-index
target-type

Examples

Here we define a single document corpus with one set of documents:

"corpora": [
  {
    "name": "geonames",
    "documents": [
      {
        "base-url": "http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/geonames",
        "source-file": "documents.json.bz2",
        "document-count": 11396505,
        "compressed-bytes": 264698741,
        "uncompressed-bytes": 3547614383,
        "target-index": "geonames",
        "target-type": "docs"
      }
    ]
  }
]

We can also define default values on document corpus level but override some of them (base-url for the last entry):

"corpora": [
  {
    "name": "http_logs",
    "base-url": "http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/http_logs",
    "target-type": "docs",
    "documents": [
      {
        "source-file": "documents-181998.json.bz2",
        "document-count": 2708746,
        "target-index": "logs-181998"
      },
      {
        "source-file": "documents-191998.json.bz2",
        "document-count": 9697882,
        "target-index": "logs-191998"
      },
      {
        "base-url": "http://example.org/corpora/http_logs",
        "source-file": "documents-201998.json.bz2",
        "document-count": 13053463,
        "target-index": "logs-201998"
      }
    ]
  }
]

operations¶

The operations section contains a list of all operations that are available later when specifying challenges. Operations define the static properties of a request against Elasticsearch whereas the schedule element defines the dynamic properties (such as the target throughput).

Each operation consists of the following properties:

name (mandatory): The name of this operation. You can choose this name freely. It is only needed to reference the operation when defining schedules.
operation-type (mandatory): Type of this operation. See below for the operation types that are supported out of the box in Rally. You can also add arbitrary operations by defining custom runners.
include-in-reporting (optional, defaults to true for normal operations and to false for administrative operations): Whether or not this operation should be included in the command line report. For example you might want Rally to create an index for you but you are not interested in detailed metrics about it. Note that Rally will still record all metrics in the metrics store.

Depending on the operation type a couple of further parameters can be specified.

bulk¶

With the operation type bulk you can execute bulk requests. It supports the following properties:

bulk-size (mandatory): Defines the bulk size in number of documents.
corpora (optional): A list of document corpus names that should be targeted by this bulk-index operation. Only needed if the corpora section contains more than one document corpus and you don’t want to index all of them with this operation.
indices (optional): A list of index names that defines which indices should be used by this bulk-index operation. Rally will then only select the documents files that have a matching target-index specified.
batch-size (optional): Defines how many documents Rally will read at once. This is an expert setting and only meant to avoid accidental bottlenecks for very small bulk sizes (e.g. if you want to benchmark with a bulk-size of 1, you should set batch-size higher).
pipeline (optional): Defines the name of an (existing) ingest pipeline that should be used (only supported from Elasticsearch 5.0).
conflicts (optional): Type of index conflicts to simulate. If not specified, no conflicts will be simulated. Valid values are: ‘sequential’ (A document id is replaced with a document id with a sequentially increasing id), ‘random’ (A document id is replaced with a document id with a random other id).

Example:

{
  "name": "index-append",
  "operation-type": "bulk",
  "bulk-size": 5000
}

Throughput will be reported as number of indexed documents per second.

force-merge¶

With the operation type force-merge you can call the force merge API. On older versions of Elasticsearch (prior to 2.1), Rally will use the optimize API instead. It supports the following parameter:

max_num_segments (optional) The number of segments the index should be merged into. Defaults to simply checking if a merge needs to execute, and if so, executes it.

This is an administrative operation. Metrics are not reported by default. If reporting is forced by setting include-in-reporting to true, then throughput is reported as the number of completed force-merge operations per second.

index-stats¶

With the operation type index-stats you can call the indices stats API. It does not support any parameters.

Throughput will be reported as number of completed index-stats operations per second.

node-stats¶

With the operation type nodes-stats you can execute nodes stats API. It does not support any parameters.

Throughput will be reported as number of completed node-stats operations per second.

search¶

With the operation type search you can execute request body searches. It supports the following properties:

index (optional): An index pattern that defines which indices should be targeted by this query. Only needed if the index section contains more than one index. Otherwise, Rally will automatically derive the index to use. If you have defined multiple indices and want to query all of them, just specify "index": "_all".
type (optional): Defines the type within the specified index for this query.
cache (optional): Whether to use the query request cache. By default, Rally will define no value thus the default depends on the benchmark candidate settings and Elasticsearch version.
request-params (optional): A structure containing arbitrary request parameters. The supported parameters names are documented in the Python ES client API docs. Parameters that are implicitly set by Rally (e.g. body or request_cache) are not supported (i.e. you should not try to set them and if so expect unspecified behavior).
body (mandatory): The query body.
pages (optional): Number of pages to retrieve. If this parameter is present, a scroll query will be executed. If you want to retrieve all result pages, use the value “all”.
results-per-page (optional): Number of documents to retrieve per page for scroll queries.

Example:

{
  "name": "default",
  "operation-type": "search",
  "body": {
    "query": {
      "match_all": {}
    }
  },
  "request-params": {
    "_source_include": "some_field",
    "analyze_wildcard": false
  }
}

For scroll queries, throughput will be reported as number of retrieved scroll pages per second. The unit is ops/s, where one op(eration) is one page that has been retrieved. The rationale is that each HTTP request corresponds to one operation and we need to issue one HTTP request per result page. Note that if you use a dedicated Elasticsearch metrics store, you can also use other request-level meta-data such as the number of hits for your own analyses.

For other queries, throughput will be reported as number of search requests per second, also measured as ops/s.

put-pipeline¶

With the operation-type put-pipeline you can execute the put pipeline API. Note that this API is only available from Elasticsearch 5.0 onwards. It supports the following properties:

id (mandatory): Pipeline id
body (mandatory): Pipeline definition

Example:

{
  "name": "define-ip-geocoder",
  "operation-type": "put-pipeline",
  "id": "ip-geocoder",
  "body": {
    "description": "Extracts location information from the client IP.",
    "processors": [
      {
        "geoip": {
          "field": "clientip",
          "properties": [
            "city_name",
            "country_iso_code",
            "country_name",
            "location"
          ]
        }
      }
    ]
  }
}

This example requires that the ingest-geoip Elasticsearch plugin is installed.

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

cluster-health¶

With the operation cluster-health you can execute the cluster health API. It supports the following properties:

request-params (optional): A structure containing any request parameters that are allowed by the cluster health API.
index (optional): The name of the index that should be used to check.

The cluster-health operation will check whether the expected cluster health and will report a failure if this is not the case. Use --on-error on the command line to control Rally’s behavior in case of such failures.

Example:

{
  "name": "check-cluster-green",
  "operation-type": "cluster-health",
  "index": "logs-*",
  "request-params": {
    "wait_for_status": "green",
    "wait_for_no_relocating_shards": "true"
  }
}

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

refresh¶

With the operation refresh you can execute the refresh API. It supports the following properties:

index (optional, defaults to _all): The name of the index that should be refreshed.

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

create-index¶

With the operation create-index you can execute the create index API. It supports two modes: it creates either all indices that are specified in the track’s indices section or it creates one specific index defined by this operation.

If you want it to create all indices that have been declared in the indices section you can specify the following properties:

settings (optional): Allows to specify additional index settings that will be merged with the index settings specified in the body of the index in the indices section.
request-params (optional): A structure containing any request parameters that are allowed by the create index API.

If you want it to create one specific index instead, you can specify the following properties:

index (mandatory): One or more names of the indices that should be created. If only one index should be created, you can use a string otherwise this needs to be a list of strings.
body (optional): The body for the create index API call.
request-params (optional): A structure containing any request parameters that are allowed by the create index API.

Examples

The following snippet will create all indices that have been defined in the indices section. It will reuse all settings defined but override the number of shards:

{
  "name": "create-all-indices",
  "operation-type": "create-index",
  "settings": {
    "index.number_of_shards": 1
  },
  "request-params": {
    "wait_for_active_shards": true
  }
}

With the following snippet we will create a new index that is not defined in the indices section. Note that we specify the index settings directly in the body:

{
  "name": "create-an-index",
  "operation-type": "create-index",
  "index": "people",
  "body": {
    "settings": {
      "index.number_of_shards": 0
    },
    "mappings": {
      "docs": {
        "properties": {
          "name": {
            "type": "text"
          }
        }
      }
    }
  }
}

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

delete-index¶

With the operation delete-index you can execute the delete index API. It supports two modes: it deletes either all indices that are specified in the track’s indices section or it deletes one specific index (pattern) defined by this operation.

If you want it to delete all indices that have been declared in the indices section, you can specify the following properties:

only-if-exists (optional, defaults to true): Defines whether an index should only be deleted if it exists.
request-params (optional): A structure containing any request parameters that are allowed by the delete index API.

If you want it to delete one specific index (pattern) instead, you can specify the following properties:

index (mandatory): One or more names of the indices that should be deleted. If only one index should be deleted, you can use a string otherwise this needs to be a list of strings.
only-if-exists (optional, defaults to true): Defines whether an index should only be deleted if it exists.
request-params (optional): A structure containing any request parameters that are allowed by the delete index API.

Examples

With the following snippet we will delete all indices that are declared in the indices section but only if they existed previously (implicit default):

{
  "name": "delete-all-indices",
  "operation-type": "delete-index"
}

With the following snippet we will delete all logs-* indices:

{
  "name": "delete-logs",
  "operation-type": "delete-index",
  "index": "logs-*",
  "only-if-exists": false,
  "request-params": {
    "expand_wildcards": "all",
    "allow_no_indices": true,
    "ignore_unavailable": true
  }
}

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

create-index-template¶

With the operation create-index-template you can execute the create index template API. It supports two modes: it creates either all index templates that are specified in the track’s templates section or it creates one specific index template defined by this operation.

If you want it to create index templates that have been declared in the templates section you can specify the following properties:

template (optional): If you specify a template name, only the template with this name will be created.
settings (optional): Allows to specify additional settings that will be merged with the settings specified in the body of the index template in the templates section.
request-params (optional): A structure containing any request parameters that are allowed by the create index template API.

If you want it to create one specific index instead, you can specify the following properties:

template (mandatory): The name of the index template that should be created.
body (mandatory): The body for the create index template API call.
request-params (optional): A structure containing any request parameters that are allowed by the create index template API.

Examples

The following snippet will create all index templates that have been defined in the templates section:

{
  "name": "create-all-templates",
  "operation-type": "create-index-template",
  "request-params": {
    "create": true
  }
}

With the following snippet we will create a new index template that is not defined in the templates section. Note that we specify the index template settings directly in the body:

{
  "name": "create-a-template",
  "operation-type": "create-index-template",
  "template": "defaults",
  "body": {
    "index_patterns": ["*"],
    "settings": {
      "number_of_shards": 3
    },
    "mappings": {
      "docs": {
        "_source": {
          "enabled": false
        }
      }
    }
  }
}

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

delete-index-template¶

With the operation delete-index-template you can execute the delete index template API. It supports two modes: it deletes either all index templates that are specified in the track’s templates section or it deletes one specific index template defined by this operation.

If you want it to delete all index templates that have been declared in the templates section, you can specify the following properties:

only-if-exists (optional, defaults to true): Defines whether an index template should only be deleted if it exists.
request-params (optional): A structure containing any request parameters that are allowed by the delete index template API.

If you want it to delete one specific index template instead, you can specify the following properties:

template (mandatory): The name of the index that should be deleted.
only-if-exists (optional, defaults to true): Defines whether the index template should only be deleted if it exists.
delete-matching-indices (optional, defaults to false): Whether to delete indices that match the index template’s index pattern.
index-pattern (mandatory iff delete-matching-indices is true): Specifies the index pattern to delete.
request-params (optional): A structure containing any request parameters that are allowed by the delete index template API.

Examples

With the following snippet we will delete all index templates that are declared in the templates section but only if they existed previously (implicit default):

{
  "name": "delete-all-index-templates",
  "operation-type": "delete-index-template"
}

With the following snippet we will delete the default` index template:

{
  "name": "delete-default-template",
  "operation-type": "delete-index-template",
  "template": "default",
  "only-if-exists": false,
  "delete-matching-indices": true,
  "index-pattern": "*"
}

Note

If delete-matching-indices is set to true, indices with the provided index-pattern are deleted regardless whether the index template has previously existed.

This is an administrative operation. Metrics are not reported by default. Reporting can be forced by setting include-in-reporting to true.

raw-request¶

With the operation raw-request you can execute arbitrary HTTP requests against Elasticsearch. This is a low-level operation that should only be used if no high-level operation is available. Note that it is always possible to write a custom runner. The raw-request operation supports the following parameters:

method (optional, defaults to GET): The HTTP request method to use
path (mandatory): Path for the API call (excluding host and port)
header (optional): A structure containing any request headers as key-value pairs.
body (optional): The document body.
request-params (optional): A structure containing HTTP request parameters.
ignore (optional): An array of HTTP response status codes to ignore (i.e. consider as successful).

challenge¶

If you track has only one challenge, you can use the challenge element. If you have multiple challenges, you can define an array of challenges.

This section contains one or more challenges which describe the benchmark scenarios for this data set. A challenge can reference all operations that are defined in the operations section.

Each challenge consists of the following properties:

name (mandatory): A descriptive name of the challenge. Should not contain spaces in order to simplify handling on the command line for users.
description (optional): A human readable description of the challenge.
default (optional): If true, Rally selects this challenge by default if the user did not specify a challenge on the command line. If your track only defines one challenge, it is implicitly selected as default, otherwise you need define "default": true on exactly one challenge.
index-settings (optional): Defines the index settings of the benchmark candidate when an index is created. Note that these settings are only applied if the index is auto-managed.
schedule (mandatory): Defines the concrete execution order of operations. It is described in more detail below.

Note

You should strive to minimize the number of challenges. If you just want to run a subset of the tasks in a challenge, use task filtering.

schedule¶

The schedule element contains a list of tasks that are executed by Rally. Each task consists of the following properties:

name (optional): This property defines an explicit name for the given task. By default the operation’s name is implicitly used as the task name but if the same operation is run multiple times, a unique task name must be specified using this property.
operation (mandatory): This property refers either to the name of an operation that has been defined in the operations section or directly defines an operation inline.
clients (optional, defaults to 1): The number of clients that should execute a task concurrently.
warmup-iterations (optional, defaults to 0): Number of iterations that each client should execute to warmup the benchmark candidate. Warmup iterations will not show up in the measurement results.
iterations (optional, defaults to 1): Number of measurement iterations that each client executes. The command line report will automatically adjust the percentile numbers based on this number (i.e. if you just run 5 iterations you will not get a 99.9th percentile because we need at least 1000 iterations to determine this value precisely).
warmup-time-period (optional, defaults to 0): A time period in seconds that Rally considers for warmup of the benchmark candidate. All response data captured during warmup will not show up in the measurement results.
time-period (optional): A time period in seconds that Rally considers for measurement. Note that for bulk indexing you should usually not define this time period. Rally will just bulk index all documents and consider every sample after the warmup time period as measurement sample.
schedule (optional, defaults to deterministic): Defines the schedule for this task, i.e. it defines at which point in time during the benchmark an operation should be executed. For example, if you specify a deterministic schedule and a target-interval of 5 (seconds), Rally will attempt to execute the corresponding operation at second 0, 5, 10, 15 … . Out of the box, Rally supports deterministic and poisson but you can define your own custom schedules.
target-throughput (optional): Defines the benchmark mode. If it is not defined, Rally assumes this is a throughput benchmark and will run the task as fast as it can. This is mostly needed for batch-style operations where it is more important to achieve the best throughput instead of an acceptable latency. If it is defined, it specifies the number of requests per second over all clients. E.g. if you specify target-throughput: 1000 with 8 clients, it means that each client will issue 125 (= 1000 / 8) requests per second. In total, all clients will issue 1000 requests each second. If Rally reports less than the specified throughput then Elasticsearch simply cannot reach it.
target-interval (optional): This is just 1 / target-throughput (in seconds) and may be more convenient for cases where the throughput is less than one operation per second. Define either target-throughput or target-interval but not both (otherwise Rally will raise an error).

Defining operations¶

In the following snippet we define two operations force-merge and a match-all query separately in an operations block:

{
  "operations": [
    {
      "name": "force-merge",
      "operation-type": "force-merge"
    },
    {
      "name": "match-all-query",
      "operation-type": "search",
      "body": {
        "query": {
          "match_all": {}
        }
      }
    }
  ],
  "challenge": {
    "name": "just-query",
    "schedule": [
      {
        "operation": "force-merge",
        "clients": 1
      },
      {
        "operation": "match-all-query",
        "clients": 4,
        "warmup-iterations": 1000,
        "iterations": 1000,
        "target-throughput": 100
      }
    ]
  }
}

If we do not want to reuse these operations, we can also define them inline. Note that the operations section is gone:

{
  "challenge": {
    "name": "just-query",
    "schedule": [
      {
        "operation": {
          "name": "force-merge",
          "operation-type": "force-merge"
        },
        "clients": 1
      },
      {
        "operation": {
          "name": "match-all-query",
          "operation-type": "search",
          "body": {
            "query": {
              "match_all": {}
            }
          }
        },
        "clients": 4,
        "warmup-iterations": 1000,
        "iterations": 1000,
        "target-throughput": 100
      }
    ]
  }
}

Contrary to the query, the force-merge operation does not take any parameters, so Rally allows us to just specify the operation-type for this operation. It’s name will be the same as the operation’s type:

{
  "challenge": {
    "name": "just-query",
    "schedule": [
      {
        "operation": "force-merge",
        "clients": 1
      },
      {
        "operation": {
          "name": "match-all-query",
          "operation-type": "search",
          "body": {
            "query": {
              "match_all": {}
            }
          }
        },
        "clients": 4,
        "warmup-iterations": 1000,
        "iterations": 1000,
        "target-throughput": 100
      }
    ]
  }
}

Choosing a schedule¶

Rally allows you to choose between the following schedules to simulate traffic:

The diagram below shows how different schedules in Rally behave during the first ten seconds of a benchmark. Each schedule is configured for a (mean) target throughput of one operation per second.

Comparison of Scheduling Strategies in Rally

If you want as much reproducibility as possible you can choose the deterministic schedule. A Poisson distribution models random independent arrivals of clients which on average match the expected arrival rate which makes it suitable for modelling the behaviour of multiple clients that decide independently when to issue a request. For this reason, Poisson processes play an important role in queueing theory.

If you have more complex needs on how to model traffic, you can also implement a custom schedule.

Time-based vs. iteration-based¶

You should usually use time periods for batch style operations and iterations for the rest. However, you can also choose to run a query for a certain time period.

All tasks in the schedule list are executed sequentially in the order in which they have been defined. However, it is also possible to execute multiple tasks concurrently, by wrapping them in a parallel element. The parallel element defines of the following properties:

clients (optional): The number of clients that should execute the provided tasks. If you specify this property, Rally will only use as many clients as you have defined on the parallel element (see examples)!
warmup-time-period (optional, defaults to 0): Allows to define a default value for all tasks of the parallel element.
time-period (optional, no default value if not specified): Allows to define a default value for all tasks of the parallel element.
warmup-iterations (optional, defaults to 0): Allows to define a default value for all tasks of the parallel element.
iterations (optional, defaults to 1): Allows to define a default value for all tasks of the parallel element.
completed-by (optional): Allows to define the name of one task in the tasks list. As soon as this task has completed, the whole parallel task structure is considered completed. If this property is not explicitly defined, the parallel task structure is considered completed as soon as all its subtasks have completed. A task is completed if and only if all associated clients have completed execution.
tasks (mandatory): Defines a list of tasks that should be executed concurrently. Each task in the list can define the following properties that have been defined above: clients, warmup-time-period, time-period, warmup-iterations and iterations.

Note

parallel elements cannot be nested.

Warning

Specify the number of clients on each task separately. If you specify this number on the parallel element instead, Rally will only use that many clients in total and you will only want to use this behavior in very rare cases (see examples)!

Examples¶

A track with a single task¶

To get started with custom tracks, you can benchmark a single task, e.g. a match_all query:

{
  "challenge": {
    "name": "just-search",
    "schedule": [
      {
        "operation": {
          "operation-type": "search",
          "index": "_all",
          "body": {
            "query": {
              "match_all": {}
            }
          }
        },
        "warmup-iterations": 100,
        "iterations": 100,
        "target-throughput": 10
      }
    ]
  }
}

This track assumes that you have an existing cluster with pre-populated data. It will run the provided match_all query at 10 operations per second with one client and use 100 iterations as warmup and the next 100 iterations to measure.

For the examples below, note that we do not show the operation definition but you should be able to infer from the operation name what it is doing.

Running unthrottled¶

In this example Rally will run a bulk index operation unthrottled for one hour:

"schedule": [
  {
    "operation": "bulk",
    "warmup-time-period": 120,
    "time-period": 3600,
    "clients": 8
  }
]

Running tasks in parallel¶

Note

You cannot nest parallel tasks.

If we want to run tasks in parallel, we can use the parallel element. In the simplest case, you let Rally decide the number of clients needed to run the parallel tasks (note how we can define default values on the parallel element):

    {
      "parallel": {
        "warmup-iterations": 50,
        "iterations": 100,
        "tasks": [
          {
            "operation": "default",
            "target-throughput": 50
          },
          {
            "operation": "term",
            "target-throughput": 200
          },
          {
            "operation": "phrase",
            "target-throughput": 200
          }
        ]
      }
    }
  ]
}

Rally will determine that three clients are needed to run each task in a dedicated client. You can also see that each task can have different settings.

However, you can also explicitly define the number of clients:

"schedule": [
  {
    "parallel": {
      "warmup-iterations": 50,
      "iterations": 100,
      "tasks": [
        {
          "operation": "match-all",
          "clients": 4,
          "target-throughput": 50
        },
        {
          "operation": "term",
          "clients": 2,
          "target-throughput": 200
        },
        {
          "operation": "phrase",
          "clients": 2,
          "target-throughput": 200
        }
      ]
    }
  }
]

This schedule will run a match all query, a term query and a phrase query concurrently. It will run with eight clients in total (four for the match all query and two each for the term and phrase query).

In this scenario, we run indexing and a few queries in parallel with a total of 14 clients:

"schedule": [
  {
    "parallel": {
      "tasks": [
        {
          "operation": "bulk",
          "warmup-time-period": 120,
          "time-period": 3600,
          "clients": 8,
          "target-throughput": 50
        },
        {
          "operation": "default",
          "clients": 2,
          "warmup-iterations": 50,
          "iterations": 100,
          "target-throughput": 50
        },
        {
          "operation": "term",
          "clients": 2,
          "warmup-iterations": 50,
          "iterations": 100,
          "target-throughput": 200
        },
        {
          "operation": "phrase",
          "clients": 2,
          "warmup-iterations": 50,
          "iterations": 100,
          "target-throughput": 200
        }
      ]
    }
  }
]

We can use completed-by to stop querying as soon as bulk-indexing has completed:

"schedule": [
  {
    "parallel": {
      "completed-by": "bulk",
      "tasks": [
        {
          "operation": "bulk",
          "warmup-time-period": 120,
          "time-period": 3600,
          "clients": 8,
          "target-throughput": 50
        },
        {
          "operation": "default",
          "clients": 2,
          "warmup-time-period": 480,
          "time-period": 7200,
          "target-throughput": 50
        }
      ]
    }
  }
]

We can also mix sequential tasks with the parallel element. In this scenario we are indexing with 8 clients and continue querying with 6 clients after indexing has finished:

"schedule": [
  {
    "operation": "bulk",
    "warmup-time-period": 120,
    "time-period": 3600,
    "clients": 8,
    "target-throughput": 50
  },
  {
    "parallel": {
      "warmup-iterations": 50,
      "iterations": 100,
      "tasks": [
        {
          "operation": "default",
          "clients": 2,
          "target-throughput": 50
        },
        {
          "operation": "term",
          "clients": 2,
          "target-throughput": 200
        },
        {
          "operation": "phrase",
          "clients": 2,
          "target-throughput": 200
        }
      ]
    }
  }
]

Be aware of the following case where we explicitly define that we want to run only with two clients in total:

"schedule": [
  {
    "parallel": {
      "warmup-iterations": 50,
      "iterations": 100,
      "clients": 2,
      "tasks": [
        {
          "operation": "match-all",
          "target-throughput": 50
        },
        {
          "operation": "term",
          "target-throughput": 200
        },
        {
          "operation": "phrase",
          "target-throughput": 200
        }
      ]
    }
  }
]

Rally will not run all three tasks in parallel because you specified that you want only two clients in total. Hence, Rally will first run “match-all” and “term” concurrently (with one client each). After they have finished, Rally will run “phrase” with one client. You could also specify more clients than there are tasks but these will then just idle.

You can also specify a number of clients on sub tasks explicitly (by default, one client is assumed per subtask). This allows to define a weight for each client operation. Note that you need to define the number of clients also on the parallel parent element, otherwise Rally would determine the number of totally needed clients again on its own:

{
  "parallel": {
    "clients": 3,
    "warmup-iterations": 50,
    "iterations": 100,
    "tasks": [
      {
        "operation": "default",
        "target-throughput": 50
      },
      {
        "operation": "term",
        "target-throughput": 200
      },
      {
        "operation": "phrase",
        "target-throughput": 200,
        "clients": 2
      }
    ]
  }
}

This will ensure that the phrase query will be executed by two clients. All other ones are executed by one client.

Configure Elasticsearch: Cars¶

Note

You can skip this section if you use Rally only as a load generator.

Definition¶

A Rally “car” is a specific configuration of Elasticsearch. You can list the available cars with esrally list cars:

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/

Name        Type    Description
----------  ------  ----------------------------------
16gheap     car     Sets the Java heap to 16GB
1gheap      car     Sets the Java heap to 1GB
2gheap      car     Sets the Java heap to 2GB
4gheap      car     Sets the Java heap to 4GB
8gheap      car     Sets the Java heap to 8GB
defaults    car     Sets the Java heap to 1GB
verbose_iw  car     Log more detailed merge time stats
ea          mixin   Enables Java assertions

You can specify the car that Rally should use with e.g. --car="4gheap". It is also possible to specify one or more “mixins” to further customize the configuration. For example, you can specify --car="4gheap,ea" to run with a 4GB heap and enable Java assertions (they are disabled by default).

Similar to custom tracks, you can also define your own cars.

The Anatomy of a car¶

The default car definitions of Rally are stored in ~/.rally/benchmarks/teams/default/cars. There we find the following structure:

├── 16gheap.ini
├── 1gheap.ini
├── 2gheap.ini
├── 4gheap.ini
├── defaults.ini
├── ea
│   └── config
│       └── jvm.options
├── ea.ini
├── vanilla
│   └── config
│       ├── elasticsearch.yml
│       ├── jvm.options
│       └── log4j2.properties
├── verbose_iw
│   └── config
│       ├── elasticsearch.yml
│       ├── jvm.options
│       └── log4j2.properties
└── verbose_iw.ini

Each .ini file in the top level directory defines a car. And each directory (ea, vanilla or verbose_iw) contains templates for the config files.

Let’s have a look at the 1gheap car by inspecting 1gheap.ini:

[meta]
description=Sets the Java heap to 1GB
type=car

[config]
base=vanilla

[variables]
heap_size=1g

The name of the car is derived from the ini file name. In the meta section we can provide a description and the type. Use car if a configuration can be used standalone and mixin if it needs to be combined with other configurations. In the config section we define that this definition is based on the vanilla configuration. We also define a variable heap_size and set it to 1g.

Let’s open vanilla/config/jvm.options to see how this variable is used (we’ll only show the relevant part here):

# Xms represents the initial size of total heap space
# Xmx represents the maximum size of total heap space

-Xms{{heap_size}}
-Xmx{{heap_size}}

So Rally reads all variables and the template files and replaces the variables in the final configuration. Note that Rally does not know anything about jvm.options or elasticsearch.yml. For Rally, these are just plain text templates that need to be copied to the Elasticsearch directory before running a benchmark. Under the hood, Rally uses Jinja2 as template language. This allows you to use Jinja2 expressions in your car configuration files.

If you open vanilla/config/elasticsearch.yml you will see a few variables that are not defined in the .ini file:

network_host
http_port
node_count_per_host

These values are derived by Rally internally based on command line flags and you cannot override them in your car definition. You also cannot use these names as names for variables because Rally would simply override them.

If you specify multiple configurations, e.g. --car="4gheap,ea", Rally will apply them in order. It will first read all variables in 4gheap.ini, then in ea.ini. Afterwards, it will copy all configuration files from the corresponding config base of 4gheap and append all configuration files from ea. This also shows when to define a separate “car” and when to define a “mixin”: If you need to amend configuration files, use a mixin, if you need to have a specific configuration, define a car.

Custom Team Repositories¶

Rally provides a default team repository that is hosted on Github. You can also add your own team repositories although this requires a bit of additional work. First of all, team repositories need to be managed by git. The reason is that Rally can benchmark multiple versions of Elasticsearch and we use git branches in the track repository to determine the best match. The versioning scheme is as follows:

The master branch needs to work with the latest master branch of Elasticsearch.
All other branches need to match the version scheme of Elasticsearch, i.e. MAJOR.MINOR.PATCH-SUFFIX where all parts except MAJOR are optional.

Rally implements a fallback logic so you don’t need to define a branch for each patch release of Elasticsearch. For example:

The branch 6.0.0-alpha1 will be chosen for the version 6.0.0-alpha1 of Elasticsearch.
The branch 5 will be chosen for all versions for Elasticsearch with the major version 5, e.g. 5.0.0, 5.1.3 (provided there is no specific branch).

Rally tries to use the branch with the best match to the benchmarked version of Elasticsearch.

Creating a new team repository¶

All team repositories are located in ~/.rally/benchmarks/teams. If you want to add a dedicated team repository, called private follow these steps:

cd ~/.rally/benchmarks/teams
mkdir private
cd private
git init
# add your team now (don't forget to add the subdirectory "cars").
git add .
git commit -m "Initial commit"

If you want to share your teams with others (or you want to run remote benchmarks) you need to add a remote and push it:

git remote add origin git@git-repos.acme.com:acme/rally-teams.git
git push -u origin master

If you have added a remote you should also add it in ~/.rally/rally.ini, otherwise you can skip this step. Open the file in your editor of choice and add the following line in the section teams:

private.url = <<URL_TO_YOUR_ORIGIN>>

Rally will then automatically update the local tracking branches before the benchmark starts.

Warning

If you run benchmarks against a remote machine that is under the control of Rally then you need to add the custom team configuration on every node!

You can now verify that everything works by listing all teams in this team repository:

esrally list cars --team-repository=private

This shows all teams that are available on the master branch of this repository. Suppose you only created tracks on the branch 2 because you’re interested in the performance of Elasticsearch 2.x, then you can specify also the distribution version:

esrally list teams --team-repository=private --distribution-version=2.0.0

Rally will follow the same branch fallback logic as described above.

Adding an already existing team repository¶

If you want to add a team repository that already exists, just open ~/.rally/rally.ini in your editor of choice and add the following line in the section teams:

your_repo_name.url = <<URL_TO_YOUR_ORIGIN>>

After you have added this line, have Rally list the tracks in this repository:

esrally list cars --team-repository=your_repo_name

Using Elasticsearch Plugins¶

You can have Rally setup an Elasticsearch cluster with plugins for you. However, there are a couple of restrictions:

This feature is only supported from Elasticsearch 5.0.0 onwards
Whereas Rally caches downloaded Elasticsearch distributions, plugins will always be installed via the Internet and thus each machine where an Elasticsearch node will be installed, requires an active Internet connection.

Listing plugins¶

To see which plugins are available, run esrally list elasticsearch-plugins:

Available Elasticsearch plugins:

Name                     Configuration
-----------------------  ----------------
analysis-icu
analysis-kuromoji
analysis-phonetic
analysis-smartcn
analysis-stempel
analysis-ukrainian
discovery-azure-classic
discovery-ec2
discovery-file
discovery-gce
ingest-attachment
ingest-geoip
ingest-user-agent
lang-javascript
lang-python
mapper-attachments
mapper-murmur3
mapper-size
repository-azure
repository-gcs
repository-hdfs
repository-s3
store-smb
x-pack                   monitoring-local
x-pack                   monitoring-http
x-pack                   security

Rally supports plugins only for Elasticsearch 5.0 or better. As the availability of plugins may change from release to release we recommend that you include the --distribution-version parameter when listing plugins. By default Rally assumes that you want to benchmark the latest master version of Elasticsearch.

Let’s see what happens if we run esrally list elasticsearch-plugins --distribution-version=2.4.0:

No Elasticsearch plugins are available.

As mentioned before, this is expected as only Elasticsearch 5.0 or better is supported.

Running a benchmark with plugins¶

In order to tell Rally to install a plugin, use the --elasticsearch-plugins parameter when starting a race. You can provide multiple plugins (comma-separated) and they will be installed in the order to that you define on the command line.

Example:

esrally --distribution-version=5.5.0 --elasticsearch-plugins="analysis-icu,analysis-phonetic"

This will install the plugins analysis-icu and analysis-phonetic (in that order). In order to use the features that these plugins provide, you need to write a custom track.

Rally will use several techniques to install and configure plugins:

First, Rally checks whether directory plugins/PLUGIN_NAME in the currently configured team repository exists. If this is the case, then plugin installation and configuration details will be read from this directory.
Next, Rally will use the provided plugin name when running the Elasticsearch plugin installer. With this approach we can avoid to create a plugin configuration directory in the team repository for very simple plugins that do not need any configuration.

As mentioned above, Rally also allows you to specify a plugin configuration and you can even combine them. Here are some examples:

Run a benchmark with the x-pack plugin in the security configuration: --elasticsearch-plugins=x-pack:security
Run a benchmark with the x-pack plugin in the security and the graph configuration: --elasticsearch-plugins=x-pack:security+graph

Note

To benchmark the security configuration of x-pack you need to add the following command line options: --client-options="use_ssl:true,verify_certs:false,basic_auth_user:'rally',basic_auth_password:'rally-password'" --cluster-health=yellow

You can also override plugin variables with --plugin-params which is needed for example if you want to use the monitoring-http configuration in order to export monitoring data. You can export monitoring data e.g. with the following configuration:

--elasticsearch-plugins="x-pack:monitoring-http" --plugin-params="monitoring_type:'https',monitoring_host:'some_remote_host',monitoring_port:10200,monitoring_user:'rally',monitoring_password:'m0n1t0r1ng'"

The monitoring_user and monitoring_password parameters are optional, the other parameters are mandatory. For more details on the configuration options, please see the Monitoring plugin documentation.

If you are behind a proxy, please set the environment variable ES_JAVA_OPTS accordingly on each target machine as described in the Elasticsearch plugin documentation.

Building plugins from sources¶

Plugin authors may want to benchmark source builds of their plugins. Your plugin is either:

built alongside Elasticsearch
built against a released version of Elasticsearch

Plugins built alongside Elasticsearch¶

To make this work, you need to manually edit Rally’s configuration file in ~/.rally/rally.ini. Suppose, we want to benchmark the plugin “my-plugin”. Then you need to add the following entries in the source section:

plugin.my-plugin.remote.repo.url = git@github.com:example-org/my-plugin.git
plugin.my-plugin.src.subdir = elasticsearch-extra/my-plugin
plugin.my-plugin.build.task = :my-plugin:plugin:assemble
plugin.my-plugin.build.artifact.subdir = plugin/build/distributions

Let’s discuss these properties one by one:

plugin.my-plugin.remote.repo.url (optional): This is needed to let Rally checkout the source code of the plugin. If this is a private repo, credentials need to be setup properly. If the source code is already locally available you may not need to define this property. The remote’s name is assumed to be “origin” and this is not configurable. Also, only git is supported as revision control system.
plugin.my-plugin.src.subdir (mandatory): This is the directory to which the plugin will be checked out relative to src.root.dir. In order to allow to build the plugin alongside Elasticsearch, the plugin needs to reside in a subdirectory of elasticsearch-extra (see also the Elasticsearch testing documentation.
plugin.my-plugin.build.task (mandatory): The Gradle task to run in order to build the plugin artifact. Note that this command is run from the Elasticsearch source directory as Rally assumes that you want to build your plugin alongside Elasticsearch (otherwise, see the next section).
plugin.my-plugin.build.artifact.subdir (mandatory): This is the subdirectory relative to plugin.my-plugin.src.subdir in which the final plugin artifact is located.

In order to run a benchmark with my-plugin, you’d invoke Rally as follows: esrally --revision="elasticsearch:some-elasticsearch-revision,my-plugin:some-plugin-revision" --elasticsearch-plugins="my-plugin" where you need to replace some-elasticsearch-revision and some-plugin-revision with the appropriate git revisions. Adjust other command line parameters (like track or car) accordingly. In order for this to work, you need to ensure that:

All prerequisites for source builds are installed.
The Elasticsearch source revision is compatible with the chosen plugin revision. Note that you do not need to know the revision hash to build against an already released version and can use git tags instead. E.g. if you want to benchmark against Elasticsearch 5.6.1, you can specify --revision="elasticsearch:v5.6.1,my-plugin:some-plugin-revision" (see e.g. the Elasticsearch tags on Github or use git tag in the Elasticsearch source directory on the console).
If your plugin needs to be configured, please ensure to create a proper plugin specification (see below).

Note

Rally can build all Elasticsearch core plugins out of the box without any further configuration.

Plugins based on a released Elasticsearch version¶

To make this work, you need to manually edit Rally’s configuration file in ~/.rally/rally.ini. Suppose, we want to benchmark the plugin “my-plugin”. Then you need to add the following entries in the source section:

plugin.my-plugin.remote.repo.url = git@github.com:example-org/my-plugin.git
plugin.my-plugin.src.dir = /path/to/your/plugin/sources
plugin.my-plugin.build.task = :my-plugin:plugin:assemble
plugin.my-plugin.build.artifact.subdir = build/distributions

Let’s discuss these properties one by one:

plugin.my-plugin.remote.repo.url (optional): This is needed to let Rally checkout the source code of the plugin. If this is a private repo, credentials need to be setup properly. If the source code is already locally available you may not need to define this property. The remote’s name is assumed to be “origin” and this is not configurable. Also, only git is supported as revision control system.
plugin.my-plugin.src.dir (mandatory): This is the absolute directory to which the source code will be checked out.
plugin.my-plugin.build.task (mandatory): The Gradle task to run in order to build the plugin artifact. This command is run from the plugin project’s root directory.
plugin.my-plugin.build.artifact.subdir (mandatory): This is the subdirectory relative to plugin.my-plugin.src.dir in which the final plugin artifact is located.

In order to run a benchmark with my-plugin, you’d invoke Rally as follows: esrally --distribution-version="elasticsearch-version" --revision="my-plugin:some-plugin-revision" --elasticsearch-plugins="my-plugin" where you need to replace elasticsearch-version with the correct release (e.g. 6.0.0) and some-plugin-revision with the appropriate git revisions. Adjust other command line parameters (like track or car) accordingly. In order for this to work, you need to ensure that:

All prerequisites for source builds are installed.
The Elasticsearch release is compatible with the chosen plugin revision.
If your plugin needs to be configured, please ensure to create a proper plugin specification (see below).

Anatomy of a plugin specification¶

Simple plugins¶

You can use Rally to benchmark community-contributed or even your own plugins. In the simplest case, the plugin does not need any custom configuration. Then you just need to add the download URL to your Rally configuration file. Consider we want to benchmark the plugin “my-plugin”:

[distributions]
plugin.my-plugin.release.url=https://example.org/my-plugin/releases/{{VERSION}}/my-plugin-{{VERSION}}.zip

Then you can use --elasticsearch-plugins=my-plugin to run a benchmark with your plugin. Rally will also replace {{VERSION}} with the distribution version that you have specified on the command line.

Plugins which require configuration¶

If the plugin needs a custom configuration we recommend to fork the official Rally teams repository and add your plugin configuration there. Suppose, you want to benchmark “my-plugin” which has the following settings that can be configured in elasticsearch.yml:

myplugin.active: a boolean which activates the plugin
myplugin.mode: Either simple or advanced

We want to support two configurations for this plugin: simple which will set myplugin.mode to simple and advanced which will set myplugin.mode to advanced.

First, we need a template configuration. We will call this a “config base” in Rally. We will just need one config base for this example and will call it “default”.

In $TEAM_REPO_ROOT create the directory structure for the plugin and its config base with mkdir -p myplugin/default/config and add the following elasticsearch.yml in the new directory:

myplugin.active: true
myplugin.mode={{my_plugin_mode}}

That’s it. Later, Rally will just copy all files in myplugin/default to the home directory of the Elasticsearch node that it configures. First, Rally will always apply the car’s configuration and then plugins can add their configuration on top. This also explains why we have created a config/elasticsearch.yml. Rally will just copy this file and replace template variables on the way.

Note

If you create a new customization for a plugin, ensure that the plugin name in the team repository matches the core plugin name. Note that hyphens need to be replaced by underscores (e.g. “x-pack” becomes “x_pack”). The reason is that Rally allows to write custom install hooks and the plugin name will become the root package name of the install hook. However, hyphens are not supported in Python which is why we use underscores instead.

The next step is now to create our two plugin configurations where we will set the variables for our config base “default”. Create a file simple.ini in the myplugin directory:

[config]
# reference our one and only config base here
base=default

[variables]
my_plugin_mode=simple

Similarly, create advanced.ini in the myplugin directory:

[config]
# reference our one and only config base here
base=default

[variables]
my_plugin_mode=advanced

Rally will now know about myplugin and its two configurations. Let’s check that with esrally list elasticsearch-plugins:

Available Elasticsearch plugins:

Name                     Configuration
-----------------------  ----------------
analysis-icu
analysis-kuromoji
analysis-phonetic
analysis-smartcn
analysis-stempel
analysis-ukrainian
discovery-azure-classic
discovery-ec2
discovery-file
discovery-gce
ingest-attachment
ingest-geoip
ingest-user-agent
lang-javascript
lang-python
mapper-attachments
mapper-murmur3
mapper-size
myplugin                 simple
myplugin                 advanced
repository-azure
repository-gcs
repository-hdfs
repository-s3
store-smb
x-pack                   monitoring-local
x-pack                   monitoring-http
x-pack                   security

As myplugin is not a core plugin, the Elasticsearch plugin manager does not know from where to install it, so we need to add the download URL to ~/.rally/rally.ini as before:

[distributions]
plugin.myplugin.release.url=https://example.org/myplugin/releases/{{VERSION}}/myplugin-{{VERSION}}.zip

Now you can run benchmarks with the custom Elasticsearch plugin, e.g. with esrally --distribution-version=5.5.0 --elasticsearch-plugins="myplugin:simple".

For this to work you need ensure two things:

The plugin needs to be available for the version that you want to benchmark (5.5.0 in the example above).
Rally will choose the most appropriate branch in the team repository before starting the benchmark. In practice, this will most likely be branch “5” for this example. Therefore you need to ensure that your plugin configuration is also available on that branch. See the README in the team repository to learn how the versioning scheme works.

Telemetry Devices¶

You probably want to gain additional insights from a race. Therefore, we have added telemetry devices to Rally. If you invoke esrally list telemetry, it will show which telemetry devices are available:

dm@io:Projects/rally ‹master*›$ esrally list telemetry

    ____        ____
   / __ \____ _/ / /_  __
  / /_/ / __ `/ / / / / /
 / _, _/ /_/ / / / /_/ /
/_/ |_|\__,_/_/_/\__, /
                /____/


Available telemetry devices:

Command    Name                   Description
---------  ---------------------  -----------------------------------------------------
jit        JIT Compiler Profiler  Enables JIT compiler logs.
gc         GC log                 Enables GC logs.
jfr        Flight Recorder        Enables Java Flight Recorder (requires an Oracle JDK)
perf       perf stat              Reads CPU PMU counters (requires Linux and perf)

Keep in mind that each telemetry device may incur a runtime overhead which can skew results.

You can attach one or more of these telemetry devices to the benchmarked cluster. However, this only works if Rally provisions the cluster (i.e. it does not work with --pipeline=benchmark-only).

jfr¶

The jfr telemetry device enables the Java Flight Recorder on the benchmark candidate. Java Flight Recorder ships only with Oracle JDK, so Rally assumes that Oracle JDK is used for benchmarking.

To enable jfr, invoke Rally with esrally --telemetry jfr. jfr will then write a flight recording file which can be opened in Java Mission Control. Rally prints the location of the flight recording file on the command line.

Note

The licensing terms of Java flight recorder do not allow you to run it in production environments without a valid license (for details, please refer to the Oracle Java SE Advanced & Suite Products page). However, running in a QA environment is fine.

jit¶

The jit telemetry device enables JIT compiler logs for the benchmark candidate. If the HotSpot disassembler library is available, the logs will also contain the disassembled JIT compiler output which can be used for low-level analysis. We recommend to use JITWatch for analysis.

hsdis can be built for JDK 8 on Linux with (based on a description by Alex Blewitt):

curl -O -O -O -O https://raw.githubusercontent.com/dmlloyd/openjdk/jdk8u/jdk8u/hotspot/src/share/tools/hsdis/{hsdis.c,hsdis.h,Makefile,README}
mkdir -p build/binutils
curl http://ftp.gnu.org/gnu/binutils/binutils-2.27.tar.gz | tar --strip-components=1 -C build/binutils -z -x -f -
make BINUTILS=build/binutils ARCH=amd64

After it has been built, the binary needs to be copied to the JDK directory (see README of hsdis for details).

gc¶

The gc telemetry device enables GC logs for the benchmark candidate. You can use tools like GCViewer to analyze the GC logs.

perf¶

The perf telemetry device runs perf stat on each benchmarked node and writes the output to a log file. It can be used to capture low-level CPU statistics. Note that the perf tool, which is only available on Linux, must be installed before using this telemetry device.

Rally Daemon¶

At its heart, Rally is a distributed system, just like Elasticsearch. However, in its simplest form you will not notice, because all components of Rally can run on a single node too. If you want Rally to configure and start Elasticsearch nodes remotely or distribute the load test driver to apply load from multiple machines, you need to use Rally daemon.

Rally daemon needs to run on every machine that should be under Rally’s control. We can consider three different roles:

Benchmark coordinator: This is the machine where you invoke esrally. It is responsible for user interaction, coordinates the whole benchmark and shows the results. Only one node can be the benchmark coordinator.
Load driver: Nodes of this type will interpret and run tracks.
Provisioner: Nodes of this type will configure an Elasticsearch cluster according to the provided car and Elasticsearch plugin configurations.

The two latter roles are not statically preassigned but rather determined by Rally based on the command line parameter --load-driver-hosts (for the load driver) and --target-hosts (for the provisioner).

Preparation¶

First, install and configure Rally on all machines that are involved in the benchmark. If you want to automate this, there is no need to use the interactive configuration routine of Rally. You can copy ~/.rally/rally.ini to the target machines adapting the paths in the file as necessary. We also recommend that you copy ~/.rally/benchmarks/data to all load driver machines before-hand. Otherwise, each load driver machine will need to download a complete copy of the benchmark data.

Note

Rally Daemon will listen on port 1900 and the actor system that Rally uses internally require access to arbitrary (unprivileged) ports. Please be sure to open up these ports between the Rally nodes.

Starting¶

For all this to work, Rally needs to form a cluster. This is achieved with the binary esrallyd (note the “d” - for daemon - at the end). You need to start the Rally daemon on all nodes: First on the coordinator node, then on all others. The order does matter, because nodes attempt to connect to the coordinator on startup.

On the benchmark coordinator, issue:

esrallyd start --node-ip=IP_OF_COORDINATOR_NODE --coordinator-ip=IP_OF_COORDINATOR_NODE

On all other nodes, issue:

esrallyd start --node-ip=IP_OF_THIS_NODE --coordinator-ip=IP_OF_COORDINATOR_NODE

After that, all Rally nodes, know about each other and you can use Rally as usual. Please see the tips and tricks for more examples.

Stopping¶

You can leave the Rally daemon processes running in case you want to run multiple benchmarks. When you are done, you can stop the Rally daemon on each node with:

esrallyd stop

Contrary to startup, order does not matter here.

Status¶

You can query the status of the local Rally daemon with:

esrallyd status

Pipelines¶

A pipeline is a series of steps that are performed to get benchmark results. This is not intended to customize the actual benchmark but rather what happens before and after a benchmark.

An example will clarify the concept: If you want to benchmark a binary distribution of Elasticsearch, Rally has to download a distribution archive, decompress it, start Elasticsearch and then run the benchmark. However, if you want to benchmark a source build of Elasticsearch, it first has to build a distribution with Gradle. So, in both cases, different steps are involved and that’s what pipelines are for.

You can get a list of all pipelines with esrally list pipelines:

Available pipelines:

Name                     Description
-----------------------  ---------------------------------------------------------------------------------------------
from-distribution        Downloads an Elasticsearch distribution, provisions it, runs a benchmark and reports results.
from-sources-complete    Builds and provisions Elasticsearch, runs a benchmark and reports results.
benchmark-only           Assumes an already running Elasticsearch instance, runs a benchmark and reports results
from-sources-skip-build  Provisions Elasticsearch (skips the build), runs a benchmark and reports results.

benchmark-only¶

This is intended if you want to provision a cluster by yourself. Do not use this pipeline unless you are absolutely sure you need to. As Rally has not provisioned the cluster, results are not easily reproducable and it also cannot gather a lot of metrics (like CPU usage).

To benchmark a cluster, you also have to specify the hosts to connect to. An example invocation:

esrally --pipeline=benchmark-only --target-hosts=search-node-a.intranet.acme.com:9200,search-node-b.intranet.acme.com:9200

from-distribution¶

This pipeline allows to benchmark an official Elasticsearch distribution which will be automatically downloaded by Rally. The earliest supported version is Elasticsearch 1.7.0. An example invocation:

esrally --pipeline=from-distribution --distribution-version=1.7.5

The version numbers have to match the name in the download URL path.

You can also benchmark Elasticsearch snapshot versions by specifying the snapshot repository:

esrally --pipeline=from-distribution --distribution-version=5.0.0-SNAPSHOT --distribution-repository=snapshot

However, this feature is mainly intended for continuous integration environments and by default you should just benchmark official distributions.

Note

This pipeline is just mentioned for completeness but Rally will autoselect it for you. All you need to do is to define the --distribution-version flag.

from-sources-complete¶

You should use this pipeline when you want to build and benchmark Elasticsearch from sources. This pipeline will only work from Elasticsearch 5.0 onwards because Elasticsearch switched from Maven to Gradle and Rally only supports one build tool in the interest of maintainability.

Remember that you also need to install git and Gradle before and Rally needs to be configured for building for sources. If that’s not the case you’ll get an error and have to run esrally configure first. An example invocation:

esrally --pipeline=from-sources-complete --revision=latest

You have to specify a revision.

Note

This pipeline is just mentioned for completeness but Rally will automatically select it for you. All you need to do is to define the --revision flag.

from-sources-skip-build¶

This pipeline is similar to from-sources-complete except that it assumes you have built the binary once. It saves time if you want to run a benchmark twice for the exact same version of Elasticsearch. Obviously it doesn’t make sense to provide a revision: It is always the previously built revision. An example invocation:

esrally --pipeline=from-sources-skip-build

Metrics¶

Metrics Records¶

At the end of a race, Rally stores all metrics records in its metrics store, which is a dedicated Elasticsearch cluster. Rally stores the metrics in the indices rally-metrics-*. It will create a new index for each month.

Here is a typical metrics record:

{
      "environment": "nightly",
      "track": "geonames",
      "track-params": {
        "shard-count": 3
      },
      "challenge": "append-no-conflicts",
      "car": "defaults",
      "sample-type": "normal",
      "trial-timestamp": "20160421T042749Z",
      "@timestamp": 1461213093093,
      "relative-time": 10507328,
      "name": "throughput",
      "value": 27385,
      "unit": "docs/s",
      "task": "index-append-no-conflicts",
      "operation": "index-append-no-conflicts",
      "operation-type": "Index",
      "lap": 1,
      "meta": {
        "cpu_physical_cores": 36,
        "cpu_logical_cores": 72,
        "cpu_model": "Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz",
        "os_name": "Linux",
        "os_version": "3.19.0-21-generic",
        "host_name": "beast2",
        "node_name": "rally-node0",
        "source_revision": "a6c0a81",
        "distribution_version": "5.0.0-SNAPSHOT",
        "tag_reference": "Github ticket 1234",
      }
    }

As you can see, we do not only store the metrics name and its value but lots of meta-information. This allows you to create different visualizations and reports in Kibana.

Below we describe each field in more detail.

environment¶

The environment describes the origin of a metric record. You define this value in the initial configuration of Rally. The intention is to clearly separate different benchmarking environments but still allow to store them in the same index.

track, track-params, challenge, car¶

This is the track, challenge and car for which the metrics record has been produced. If the user has provided track parameters with the command line parameter, --track-params, each of them is listed here too.

If you specify a car with mixins, it will be stored as one string separated with “+”, e.g. --car="4gheap,ea" will be stored as 4gheap+ea in the metrics store in order to simplify querying in Kibana. For more details, please see the cars documentation.

sample-type¶

Rally runs warmup trials but records all samples. Normally, we are just interested in “normal” samples but for a full picture we might want to look also at “warmup” samples.

trial-timestamp¶

A constant timestamp (always in UTC) that is determined when Rally is invoked. It is intended to group all samples of a benchmark trial.

@timestamp¶

The timestamp in milliseconds since epoch determined when the sample was taken.

relative-time¶

The relative time in microseconds since the start of the benchmark. This is useful for comparing time-series graphs over multiple trials, e.g. you might want to compare the indexing throughput over time across multiple benchmark trials. Obviously, they should always start at the same (relative) point in time and absolute timestamps are useless for that.

name, value, unit¶

This is the actual metric name and value with an optional unit (counter metrics don’t have a unit). Depending on the nature of a metric, it is either sampled periodically by Rally, e.g. the CPU utilization or query latency or just measured once like the final size of the index.

task, operation, operation-type¶

task is the name of the task (as specified in the track file) that ran when this metric has been gathered. Most of the time, this value will be identical to the operation’s name but if the same operation is ran multiple times, the task name will be unique whereas the operation may occur multiple times. It will only be set for metrics with name latency and throughput.

operation is the name of the operation (as specified in the track file) that ran when this metric has been gathered. It will only be set for metrics with name latency and throughput.

operation-type is the more abstract type of an operation. During a race, multiple queries may be issued which are different operation``s but they all have the same ``operation-type (Search). For some metrics, only the operation type matters, e.g. it does not make any sense to attribute the CPU usage to an individual query but instead attribute it just to the operation type.

lap¶

The lap number in which this metric was gathered. Laps start at 1. See the command line reference for more info on laps.

meta¶

Rally captures also some meta information for each metric record:

CPU info: number of physical and logical cores and also the model name
OS info: OS name and version
Host name
Node name: If Rally provisions the cluster, it will choose a unique name for each node.
Source revision: We always record the git hash of the version of Elasticsearch that is benchmarked. This is even done if you benchmark an official binary release.
Distribution version: We always record the distribution version of Elasticsearch that is benchmarked. This is even done if you benchmark a source release.
Custom tag: You can define one custom tag with the command line flag --user-tag. The tag is prefixed by tag_ in order to avoid accidental clashes with Rally internal tags.
Operation-specific: The optional substructure operation contains additional information depending on the type of operation. For bulk requests, this may be the number of documents or for searches the number of hits.

Note that depending on the “level” of a metric record, certain meta information might be missing. It makes no sense to record host level meta info for a cluster wide metric record, like a query latency (as it cannot be attributed to a single node).

Metric Keys¶

Rally stores the following metrics:

latency: Time period between submission of a request and receiving the complete response. It also includes wait time, i.e. the time the request spends waiting until it is ready to be serviced by Elasticsearch.
service_time Time period between start of request processing and receiving the complete response. This metric can easily be mixed up with latency but does not include waiting time. This is what most load testing tools refer to as “latency” (although it is incorrect).
throughput: Number of operations that Elasticsearch can perform within a certain time period, usually per second. See the track reference for a definition of what is meant by one “operation” for each operation type.
merge_parts_total_time_*: Different merge times as reported by Lucene. Only available if Lucene index writer trace logging is enabled.
merge_parts_total_docs_*: See merge_parts_total_time_*
disk_io_write_bytes: number of bytes that have been written to disk during the benchmark. On Linux this metric reports only the bytes that have been written by Elasticsearch, on Mac OS X it reports the number of bytes written by all processes.
disk_io_read_bytes: number of bytes that have been read from disk during the benchmark. The same caveats apply on Mac OS X as for disk_io_write_bytes.
cpu_utilization_1s: CPU usage in percent of the Elasticsearch process based on a one second sample period. The maximum value is N * 100% where N is the number of CPU cores available.
node_startup_time: The time in seconds it took from process start until the node is up.
node_total_old_gen_gc_time: The total runtime of the old generation garbage collector across the whole cluster as reported by the node stats API.
node_total_young_gen_gc_time: The total runtime of the young generation garbage collector across the whole cluster as reported by the node stats API.
segments_count: Total number of segments as reported by the indices stats API.
segments_memory_in_bytes: Number of bytes used for segments as reported by the indices stats API.
segments_doc_values_memory_in_bytes: Number of bytes used for doc values as reported by the indices stats API.
segments_stored_fields_memory_in_bytes: Number of bytes used for stored fields as reported by the indices stats API.
segments_terms_memory_in_bytes: Number of bytes used for terms as reported by the indices stats API.
segments_norms_memory_in_bytes: Number of bytes used for norms as reported by the indices stats API.
segments_points_memory_in_bytes: Number of bytes used for points as reported by the indices stats API.
merges_total_time: Total runtime of merges as reported by the indices stats API. Note that this is not Wall clock time (i.e. if M merge threads ran for N minutes, we will report M * N minutes, not N minutes).
merges_total_throttled_time: Total time within merges have been throttled as reported by the indices stats API. Note that this is not Wall clock time.
indexing_total_time: Total time used for indexing as reported by the indices stats API. Note that this is not Wall clock time.
indexing_throttle_time: Total time that indexing has been throttled as reported by the indices stats API. Note that this is not Wall clock time.
refresh_total_time: Total time used for index refresh as reported by the indices stats API. Note that this is not Wall clock time.
flush_total_time: Total time used for index flush as reported by the indices stats API. Note that this is not Wall clock time.
final_index_size_bytes: Final resulting index size after the benchmark.

Summary Report¶

At the end of each race, Rally shows a summary report. Below we’ll explain the meaning of each line including a reference to its corresponding metrics key which can be helpful if you want to build your own reports in Kibana. Note that not every summary report will show all lines.

Indexing time¶

Definition: Total time used for indexing as reported by the indices stats API. Note that this is not Wall clock time (i.e. if M indexing threads ran for N minutes, we will report M * N minutes, not N minutes).
Corresponding metrics key: indexing_total_time

Indexing throttle time¶

Definition: Total time that indexing has been throttled as reported by the indices stats API. Note that this is not Wall clock time (i.e. if M indexing threads ran for N minutes, we will report M * N minutes, not N minutes).
Corresponding metrics key: indexing_throttle_time

Merge time¶

Definition: Total runtime of merges as reported by the indices stats API. Note that this is not Wall clock time.
Corresponding metrics key: merges_total_time

Refresh time¶

Definition: Total time used for index refresh as reported by the indices stats API. Note that this is not Wall clock time.
Corresponding metrics key: refresh_total_time

Flush time¶

Definition: Total time used for index flush as reported by the indices stats API. Note that this is not Wall clock time.
Corresponding metrics key: flush_total_time

Merge throttle time¶

Definition: Total time within merges have been throttled as reported by the indices stats API. Note that this is not Wall clock time.
Corresponding metrics key: merges_total_throttled_time

Merge time (`X`)¶

Where X is one of:

postings
stored fields
doc values
norms
vectors
points

Definition: Different merge times as reported by Lucene. Only available if Lucene index writer trace logging is enabled (use the car verbose_iw for that).
Corresponding metrics keys: merge_parts_total_time_*

Median CPU usage¶

Definition: Median CPU usage in percent of the Elasticsearch process during the whole race based on a one second sample period. The maximum value is N * 100% where N is the number of CPU cores available
Corresponding metrics key: cpu_utilization_1s

Total Young Gen GC¶

Definition: The total runtime of the young generation garbage collector across the whole cluster as reported by the node stats API.
Corresponding metrics key: node_total_young_gen_gc_time

Total Old Gen GC¶

Definition: The total runtime of the old generation garbage collector across the whole cluster as reported by the node stats API.
Corresponding metrics key: node_total_old_gen_gc_time

Index size¶

Definition: Final resulting index size after the benchmark.
Corresponding metrics key: final_index_size_bytes

Totally written¶

Definition: number of bytes that have been written to disk during the benchmark. On Linux this metric reports only the bytes that have been written by Elasticsearch, on Mac OS X it reports the number of bytes written by all processes.
Corresponding metrics key: disk_io_write_bytes

Heap used for `X`¶

Where X is one of:

doc values
terms
norms
points
stored fields

Definition: Number of bytes used for the corresponding item as reported by the indices stats API.
Corresponding metrics keys: segments_*_in_bytes

Segment count¶

Definition: Total number of segments as reported by the indices stats API.
Corresponding metrics key: segments_count

Throughput¶

Rally reports the minimum, median and maximum throughput for each task.

Definition: Number of operations that Elasticsearch can perform within a certain time period, usually per second.
Corresponding metrics key: throughput

Latency¶

Rally reports several percentile numbers for each task. Which percentiles are shown depends on how many requests Rally could capture (i.e. Rally will not show a 99.99th percentile if it could only capture five samples because that would be a vanity metric).

Definition: Time period between submission of a request and receiving the complete response. It also includes wait time, i.e. the time the request spends waiting until it is ready to be serviced by Elasticsearch.
Corresponding metrics key: latency

Service time¶

Rally reports several percentile numbers for each task. Which percentiles are shown depends on how many requests Rally could capture (i.e. Rally will not show a 99.99th percentile if it could only capture five samples because that would be a vanity metric).

Definition: Time period between start of request processing and receiving the complete response. This metric can easily be mixed up with latency but does not include waiting time. This is what most load testing tools refer to as “latency” (although it is incorrect).
Corresponding metrics key: service_time

Error rate¶

Definition: The ratio of erroneous responses relative to the total number of responses. Any exception thrown by the Python Elasticsearch client is considered erroneous (e.g. HTTP response codes 4xx, 5xx or network errors (network unreachable)). For specific details, please check the reference documentation of the Elasticsearch client. Usually any error rate greater than zero is alerting. You should investigate the root cause by inspecting Rally and Elasticsearch logs and rerun the benchmark.
Corresponding metrics key: service_time. Each service_time record has a meta.success flag. Rally simply counts how often this flag is true and false respectively.

Migration Guide¶

Migrating to Rally 0.9.0¶

Track Syntax¶

With Rally 0.9.0, we have changed the track file format. While the previous format is still supported with deprecation warnings, we recommend that you adapt your tracks as we will remove the deprecated syntax with the next minor release.

Below is an example of a track with the previous syntax:

{
  "description": "Tutorial benchmark for Rally",
  "data-url": "http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/geonames",
  "indices": [
    {
      "name": "geonames",
      "types": [
        {
          "name": "type",
          "mapping": "mappings.json",
          "documents": "documents.json",
          "document-count": 8647880,
          "uncompressed-bytes": 2790927196
        }
      ]
    }
  ],
  "challenge": {
    "name": "index-only",
    "index-settings": {
      "index.number_of_replicas": 0
    },
    "schedule": [
      {
        "operation": {
          "operation-type": "bulk",
          "bulk-size": 5000
        },
        "warmup-time-period": 120,
        "clients": 8
      }
    ]
  }
}

Before Rally 0.9.0, indices have been created implicitly. We will remove this ability and thus you need to tell Rally explicitly that you want to create indices. With Rally 0.9.0 your track should look as follows:

{
  "description": "Tutorial benchmark for Rally",
  "indices": [
    {
      "name": "geonames",
      "body": "index.json",
      "auto-managed": false,
      "types": [ "type" ]
    }
  ],
  "corpora": [
    {
      "name": "geonames",
      "documents": [
        {
          "base-url": "http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/geonames",
          "source-file": "documents.json",
          "document-count": 8647880,
          "uncompressed-bytes": 2790927196
        }
      ]
    }
  ],
  "challenge": {
    "name": "index-only",
    "schedule": [
      {
        "operation": "delete-index"
      },
      {
        "operation": {
          "operation-type": "create-index",
          "settings": {
            "index.number_of_replicas": 0
          }
        }
      },
      {
        "operation": {
          "operation-type": "cluster-health",
          "request-params": {
            "wait_for_status": "green"
          }
        }
      },
      {
        "operation": {
          "operation-type": "bulk",
          "bulk-size": 5000
        },
        "warmup-time-period": 120,
        "clients": 8
      }
    ]
  }
}

Let’s go through the necessary changes one by one.

Define the document corpus separately¶

Previously you had to define the document corpus together with the document type. In order to allow you to reuse existing document corpora across tracks, you now need to specify any document corpora separately:

"corpora": [
  {
    "name": "geonames",
    "documents": [
      {
        "base-url": "http://benchmarks.elasticsearch.org.s3.amazonaws.com/corpora/geonames",
        "source-file": "documents.json",
        "document-count": 8647880,
        "uncompressed-bytes": 2790927196
      }
    ]
  }
]

Note that this is just a simple example that should cover the most basic case. Be sure to check the track reference for all details.

Change the index definition¶

The new index definition now looks as follows:

{
  "name": "geonames",
  "body": "index.json",
  "auto-managed": false,
  "types": [ "type" ]
}

We have added a body property to the index and removed the mapping property from the type. In fact, the only information that we need about the document type is its name, hence it is now a simple list of strings. Just put all type mappings now into the mappings property of the index definition. For more details, please refer to the create index API documentation.

Secondly, we have disabled index auto-management by setting auto-managed to false. This allows us to define explicit tasks below to manage our index. Note that index auto-management is still working in Rally 0.9.0 but it will be removed with the next minor release Rally 0.10.0.

Explicitly delete and recreate the index¶

We have also added three tasks at the beginning of the schedule:

{
  "operation": "delete-index"
},
{
  "operation": {
    "operation-type": "create-index",
    "settings": {
      "index.number_of_replicas": 0
    }
  }
},
{
  "operation": {
    "operation-type": "cluster-health",
    "request-params": {
      "wait_for_status": "green"
    }
  }
}

These tasks represent what Rally previously did implicitly.

The first task will delete all indices that have been declared in the indices section if they existed previously. This ensures that we don’t have any leftovers from previous benchmarks.

After that we will create all indices that have been declared in the indices section. Note that we have also removed the special property index-settings and moved it to the settings parameter of create-index. Rally will merge any settings from the index body definition with these settings. This means you should define settings that are always the same in the index body and settings that change from challenge to challenge in the settings property.

Finally, Rally will check that the cluster health is green. If you want to be able to override the cluster health check parameters from the command line, you can leverage Rally’s track parameter feature:

{
  "operation": {
    "operation-type": "cluster-health",
    "request-params": {
      "wait_for_status": "{{ cluster_health|default('green') }}"
    }
  }
}

If you don’t specify anything on the command line, Rally will use the default value but you can e.g. specify --track-params="cluster_health:'yellow'" so Rally will check for (at least) a yellow cluster health status.

Note that you can customize these operations. Please see the track reference for all details.

Custom Parameter Sources¶

With Rally 0.9.0, the API for custom parameter sources has changed. Previously, the following syntax was valid:

# for parameter sources implemented as functions
def custom_param_source(indices, params):

# for parameter sources implemented as classes
class CustomParamSource:
    def __init__(self, indices, params):

With Rally 0.9.0, the signatures need to be changed to:

# for parameter sources implemented as functions
def custom_param_source(track, params, **kwargs):

# for parameter sources implemented as classes
class CustomParamSource:
    def __init__(self, track, params, **kwargs):

Rally will issue a warning along the lines of Parameter source 'custom_param_source' is using deprecated method signature if your track is affected. If you need access to the indices list, you can call track.indices to retrieve it from the track.

Frequently Asked Questions (FAQ)¶

A benchmark aborts with `Couldn't find a tar.gz distribution`. What’s the problem?¶

This error occurs when Rally cannot build an Elasticsearch distribution from source code. The most likely cause is that there is some problem in the build setup.

To see what’s the problem, try building Elasticsearch yourself. First, find out where the source code is located (run grep src ~/.rally/rally.ini). Then change to the directory (src.root.dir + elasticsearch.src.subdir which is usually ~/.rally/benchmarks/src/elasticsearch) and run the following commands:

gradle clean
gradle :distribution:tar:assemble

By that you are mimicking what Rally does. Fix any errors that show up here and then retry.

Where does Rally get the benchmark data from?¶

Rally comes with a set of tracks out of the box which we maintain in the rally-tracks repository on Github. This repository contains the track descriptions. The actual data are stored as compressed files in an S3 bucket.

Will Rally destroy my existing indices?¶

First of all: Please (please, please) do NOT run Rally against your production cluster if you are just getting started with it. You have been warned.

Depending on the track, Rally will delete and create one or more indices. For example, the geonames track specifies that Rally should create an index named “geonames” and Rally will assume it can do to this index whatever it wants. Specifically, Rally will check at the beginning of a race if the index “geonames” exists and delete it. After that it creates a new empty “geonames” index and runs the benchmark. So if you benchmark against your own cluster (by specifying the benchmark-only pipeline) and this cluster contains an index that is called “geonames” you will lose (all) data if you run Rally against it. Rally will neither read nor write (or delete) any other index. So if you apply the usual care nothing bad can happen.

Where and how long does Rally keep its data?¶

Rally stores a lot of data (this is just the nature of a benchmark) so you should keep an eye on disk usage. All data are kept in ~/.rally and Rally does not implicitly delete them. These are the most important directories:

~/.rally/logs: Contains all log files. Logs are rotated daily. If you don’t need the logs anymore, you can safely wipe this directory.
~/.rally/benchmarks/races: telemetry data, Elasticsearch logs and even complete Elasticsearch installations including the data directory if a benchmark failed. If you don’t need the data anymore, you can safely wipe this directory.
~/.rally/benchmarks/src: the Elasticsearch Github repository (only if you had Rally build Elasticsearch from sources at least once).
~/.rally/benchmarks/data: the benchmark data sets. This directory can get very huge (way more than 100 GB if you want to try all default tracks). You can delete the files in this directory but keep in mind that Rally may needs to download them again.
~/.rally/benchmarks/distributions: Contains all downloaded Elasticsearch distributions.

There are a few more directories but the ones above are the most disk-hogging ones.

Does Rally spy on me?¶

No. Rally does not collect or send any usage data and also the complete source code is open. We do value your feedback a lot though and if you got any ideas for improvements, found a bug or have any other feedback, please head over to Rally’s Discuss forum or raise an issue on Github.

Do I need an Internet connection?¶

You do NOT need Internet access on any node of your Elasticsearch cluster but the machine where you start Rally needs an Internet connection to download track data sets and Elasticsearch distributions. After it has downloaded all data, an Internet connection is not required anymore and you can specify --offline. If Rally detects no active Internet connection, it will automatically enable offline mode and warn you.

We have a dedicated documentation page for running Rally offline which should cover all necessary details.

Glossary¶

track: A track is the description of one or more benchmarking scenarios with a specific document corpus. It defines for example the involved indices, data files and which operations are invoked. List the available tracks with esrally list tracks. Although Rally ships with some tracks out of the box, you should usually create your own track based on your own data.
challenge: A challenge describes one benchmarking scenario, for example indexing documents at maximum throughput with 4 clients while issuing term and phrase queries from another two clients rate-limited at 10 queries per second each. It is always specified in the context of a track. See the available challenges by listing the corresponding tracks with esrally list tracks.
car: A car is a specific configuration of an Elasticsearch cluster that is benchmarked, for example the out-of-the-box configuration, a configuration with a specific heap size or a custom logging configuration. List the available cars with esrally list cars.
telemetry: Telemetry is used in Rally to gather metrics about the car, for example CPU usage or index size.
race: A race is one invocation of the Rally binary. Another name for that is one “benchmarking trial”. During a race, Rally runs one challenge on a track with the given car.
tournament: A tournament is a comparison of two races. You can use Rally’s tournament mode for that.

Community Resources¶

Below are a few community resources about Rally. If you find an interesting article, talk or custom tracks, please raise an issue or open a pull request.

Talks¶

Articles¶

Using Rally to benchmark Elasticsearch queries by Darren Smith

License¶

This software is licensed under the Apache License, version 2 (“ALv2”), quoted below.

Licensed under the Apache License, Version 2.0 (the “License”); you may not use this file except in compliance with the License. You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an “AS IS” BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Rally 0.9.3¶

Getting Help or Contributing to Rally¶

Source Code¶

Quickstart¶

Install¶

Configure¶

Run your first race¶

Next steps¶

Installation¶

Hardware Requirements¶

Prerequisites¶

Python¶

git¶

JDK¶

Installing Rally¶

Non-sudo Install¶

VirtualEnv Install¶

Kubernetes Job¶

Docker¶

Offline Install¶

Next Steps¶

Configuration¶

Simple Configuration¶

Advanced Configuration¶

Prerequisites¶

Preparation¶

Configuration Options¶

Proxy Configuration¶

Run a Benchmark: Races¶

Definition¶

List Tracks¶

Starting a Race¶

Compare Results: Tournaments¶

Tips and Tricks¶

Benchmarking an existing cluster¶

Benchmarking a remote cluster¶

Distributing the load test driver¶

Changing the default track repository¶

Define Custom Workloads: Tracks¶

Definition¶

Example track¶

Adding support for test mode¶

Structuring your track¶

Sharing your track with others¶

Advanced topics¶

Template Language¶

Extension Points¶

Custom parameter sources¶

Custom runners¶

Custom schedulers¶

Developing Rally¶

Prerequisites¶

Installation Instructions for Development¶

Automatic Updates¶

Configuring Rally¶

Key Components of Rally¶

How to contribute code¶

Command Line Reference¶

Subcommands¶

race¶

list¶

compare¶

configure¶

Command Line Flags¶

track-path¶

track-repository¶

track¶

track-params¶

challenge¶

include-tasks¶

team-repository¶

car¶

car-params¶

elasticsearch-plugins¶

plugin-params¶

pipeline¶

laps¶

enable-driver-profiling¶

test-mode¶

telemetry¶

`race`¶

`list`¶

`compare`¶

`configure`¶

`track-path`¶

`track-repository`¶

`track`¶

`track-params`¶

`challenge`¶

`include-tasks`¶

`team-repository`¶

`car`¶

`car-params`¶

`elasticsearch-plugins`¶

`plugin-params`¶

`pipeline`¶

`laps`¶

`enable-driver-profiling`¶

`test-mode`¶

`telemetry`¶

`revision`¶

`distribution-version`¶

`distribution-repository`¶

`report-format`¶

`show-in-report`¶

`report-file`¶

`client-options`¶

`on-error`¶

`load-driver-hosts`¶

`target-hosts`¶

`quiet`¶

`offline`¶

`preserve-install`¶

`cluster-health`¶

`advanced-config`¶

`assume-defaults`¶

`user-tag`¶