Tezt: Long Tests

Tezt can also be used for long tests. The differences with regular Tezt tests are:

  • long tests are not run in the CI but on dedicated machines with stable, predictable performance, and with no global timeout like in the CI;

  • long tests are declared in tezt/long_tests/main.ml instead of tezt/tests/main.ml;

  • long tests are registered with Long_test.register instead of Test.register;

  • long tests have easy access to an InfluxDB database to send data points such as how long it takes to do something;

  • long tests have easy access to a Grafana instance that they can update to create graphs displaying the data points sent to InfluxDB;

  • long tests have easy access to a Slack webhook that can be used to send alerts when measurements differ significantly from previous tests.

Adding a Long Test

To add a long test, you need to register it with Long_test.register (respectively Long_test.register_with_protocol) instead of Test.register (respectively Protocol.register_test). The test must be registered in tezt/long_tests/main.ml instead of tezt/tests/main.ml. Long_test.register is very similar to Test.register. The main difference is that you have to declare a timeout (that should be significantly overestimated, see the documentation of the timeout type in tezt/long_tests/long_test.mli) and that Long_test.register will handle InfluxDB, Grafana and Slack alerts. Just like regular Tezt tests, your test should be implemented in the same file with other thematically-related tests, with only a single line in tezt/long_tests/main.ml.

If you declare a test in tezt/long_tests/main.ml and merge it into master, it will be automatically run on dedicated machines that regularly pull the latest version of master and run long tests. If your test takes a significant time to run (days), you should however ask that a new dedicated machine is created to run your test. Please ask on the #tests Slack channel of tezos-dev before merging.

Time Series, Alerts and Graphs

Long tests can use functions from the Long_test module (tezt/long_tests/long_test.mli) to send data points to InfluxDB, which is a time-series database. A time-series database stores values annotated with a timestamp. In the particular case of InfluxDB, data points are composed of:

  • a timestamp (usually the time the measure was taken);

  • a measurement name (such as client load time);

  • optionally, some tags that can be used to refine the measurement name (such as a tag mode equal to client, mockup or proxy);

  • fields, which are composed of field names and field values (such as a field duration equal to how long it took for the client to load).

The Long_test module allows you to send any data point using the add_data_point function. They will be sent at the end of the test execution. You can query statistics about data points from previous tests using function get_previous_stats to compare them with your new data points and send an alert on Slack with function alert if you detect that something is wrong. Note that function check_regression, which is essentially a combination of get_previous_stats and alert, can be used in most cases. And if all you want to do is measure the time it takes for a chunk of code to run though, you can use the time function instead, which does everything for you: measure time, send data points, compare them with previous runs, and send an alert if the difference is too large.

The Long_test module also provides function update_grafana_dashboard which is called in tezt/long_tests/main.ml with a specification of graphs to display in Grafana at https://graf.nomadic-labs.com/?orgId=1 (in the “Long Tezts” dashboard by default but you can create your own dashboards). To add graphs for your tests, define them next to your test in the same file as your test, and declare them in tezt/long_tests/main.ml in the call to update_grafana_dashboard.

As always in Tezt, the above functions try to provide flexibility. The time function in particular is parameterized by settings like the number of times the test should be repeated, how many previous data points should be fetched from InfluxDB when comparing with new measurements, how much of a difference to tolerate before alerting, etc. time itself being a combination of other lower-level functions that are also provided and which you can combine to fit your needs. And of course you can contribute to improve them.

Example

See tezt/long_tests/prt_client.ml, which is a very simple test that measures how long it takes for tezos-client to load. It uses Long_test.time_lwt to measure how long it takes for Client.version to run and to emit alerts if this time is significantly higher than usual. It also defines a graph of this time. This test and its graph are registered in tezt/long_tests/main.ml.

One-Shot Tests

You may be interested in running some long tests using this framework on your own branch instead of master. This is work-in-progress, but the goal is to make it easy to do, possibly with a manual job in your branch’s CI pipeline that you could trigger to spawn the tests for your branch. Stay tuned.

Providing Large Data

Your test may require data that is too large to commit in tezos/tezos. For example, a benchmark in which measurement is dependent on some block’s context would need to load the same data directory on each execution. This is also a work-in-progress, but a dedicacted AWS storage will be provided as well as a way to dowload files and archives from it to fullfill these needs.