TensorFlow testing best practices

These are the recommended practices for testing code in the TensorFlow repository.

Before you get started

Before you contribute source code to a TensorFlow project, please review the CONTRIBUTING.md file in the GitHub repo of the project. (For example, see the CONTRIBUTING.md file for the core TensorFlow repo.) All code contributors are required to sign a Contributor License Agreement (CLA).

General principles

Only depend on what you use in your BUILD rules

TensorFlow is a large library, and depending on the full package when writing a unit test for its submodules has been a common practice. However, this disables the bazel dependency-based analysis. This means that continuous integration systems cannot intelligently eliminate unrelated tests for presubmit/postsubmit runs. If you only depend on the submodules that you are testing in your BUILD file, you will save time for all TensorFlow developers, and a lot of valuable computation power.

However, modifying your build dependency to omit the full TF targets brings some limitations for what you can import in your Python code. You will not be able to use the import tensorflow as tf statement in your unit tests anymore. But this is a worthwhile tradeoff since as it saves all developers from running thousands of unnecessary tests.

All code should have unit tests

For any code you write, you should also write its unit tests. If you write a new file foo.py, you should place its unit tests in foo_test.py and submit it within the same change. Aim for >90% incremental test coverage for all your code.

Avoid using native bazel test rules in TF

TF has a lot of subtleties when running tests. We have worked to hide all of those complexities in our bazel macros. To avoid having to deal with those, use the following instead of the native test rules. Note that all of these are defined in tensorflow/tensorflow.bzl For CC tests, use tf_cc_test, tf_gpu_cc_test, tf_gpu_only_cc_test. For python tests, use tf_py_test or gpu_py_test. If you need something really close to the native py_test rule, please use the one defined in tensorflow.bzl instead. You just need to add the following line at the top of the BUILD file: load(“tensorflow/tensorflow.bzl”, “py_test”)

Be aware where the test executes

When you write a test, our test infra can take care of running your tests on CPU, GPU and accelerators if you write them accordingly. We have automated tests that run on Linux, macos, windows, that have systems with or without GPUs. You simply need to pick one of the macros listed above, and then use tags to limit where they are executed.

  • manual tag will exclude your test from running anywhere. This includes manual test executions that use patterns such as bazel test tensorflow/…

  • no_oss will exclude your test from running in the official TF OSS test infrastructure.

  • no_mac or no_windows tags can be used to exclude your test from relevant operating system test suites.

  • no_gpu tag can be used to exclude your test from running in GPU test suites.

Verify tests run in expected test suites

TF has quite a few test suites. Sometimes, they may be confusing to set up. There might be different problems that cause your tests to be omitted from continuous builds. Thus, you should verify your tests are executing as expected. To do this:

  • Wait for your presubmits on your Pull Request(PR) to run to completion.
  • Scroll to the bottom of your PR to see the status checks.
  • Click the “Details” link at the right side of any Kokoro check.
  • Check the “Targets” list to find your newly added targets.

Each class/unit should have its own unit test file

Separate test classes help us better isolate failures and resources. They lead to much shorter and easier to read test files. Therefore, all your Python files should have at least one corresponding test file (For each foo.py, it should have foo_test.py). For more elaborate tests, such as integration tests that require different setups, it is fine to add more test files.

Speed and running times

Sharding should be used as little as possible

Instead of sharding please consider:

  • Making your tests smaller
  • If the above is not possible, split the tests up

Sharding helps reduce the overall latency of a test, but the same can be achieved by breaking up tests to smaller targets. Splitting tests gives us a finer level of control on each test, minimizing unnecessary presubmit runs and reducing the coverage loss from a buildcop disabling an entire target due to a misbehaving testcase. Moreover, sharding incurs hidden costs that are not so obvious, such as running all test initialization code for all shards. This issue has been escalated to us by infra teams as a source that creates extra load.

Smaller tests are better

The quicker your tests run, the more likely people will be to run your tests. One extra second for your test can accumulate to hours of extra time spent running your test by developers and our infrastructure. Try to make your tests run under 30 seconds (in non-opt mode!), and make them small. Only mark your tests as medium as a last resort. The infra does not run any large tests as presubmits or postsubmits! Therefore, only write a large test if you are going to arrange where it is going to run. Some tips to make tests run faster:

  • Run less iterations of training in your test
  • Consider using dependency injection to replace heavy dependencies of system under test with simple fakes.
  • Consider using smaller input data in unit tests
  • If nothing else works, try splitting up your test file.

Test times should aim for half of test size timeout to avoid flakes

With bazel test targets, small tests have 1 minute timeouts. Medium test timeouts are 5 minutes. Large tests are just not executed by the TensorFlow test infra. However, many tests are not deterministic in the amount of time they take. For various reasons your tests might take more time every now and then. And, if you mark a test that runs for 50 seconds on the average as small, your test will flake if it schedules on a machine with an old CPU. Therefore, aim for 30 second average running time for small tests. Aim for 2 minutes 30 seconds of average running time for medium tests.

Reduce the number of samples and increase tolerances for training

Slow running tests deter contributors. Running training in tests can be very slow. Prefer higher tolerances to be able to use less samples in your tests to keep your tests sufficiently fast (2.5 minutes max).

Eliminate non-determinism and flakes

Write deterministic tests

Unit tests should always be deterministic. All tests running on TAP and guitar should run the same way every single time, if there is no code change affecting them. To ensure this, below are some points to consider.

Always seed any source of stochasticity

Any random number generator, or any other sources of stochasticity can cause flakiness. Therefore, each of these must be seeded. In addition to making tests less flaky, this makes all tests reproducible. Different ways to set some seeds you may need to set in TF tests are:

# Python RNG
import random

# Numpy RNG
import numpy as np

from tensorflow.python.framework import random_seed

Avoid using sleep in multithreaded tests

Using sleep function in tests can be a major cause of flakiness. Especially when using multiple threads, using sleep to wait for another thread will never be determistic. This is due to system not being able to guarantee any ordering of execution of different threads or processes. Therefore, prefer deterministic synchronization constructs such as mutexes.

Check if the test is flaky

Flakes cause buildcops and developers to lose many hours. They are difficult to detect, and they are difficult to debug. Even though there are automated systems to detect flakiness, they need to accumulate hundreds of test runs before they can accurately denylist tests. Even when they detect, they denylist your tests and test coverage is lost. Therefore, test authors should check if their tests are flaky when writing tests. This can be easily done by running your test with the flag: --runs_per_test=1000

Use TensorFlowTestCase

TensorFlowTestCase takes necessary precautions such as seeding all random number generators used to reduce flakiness as much as possible. As we discover and fix more flakiness sources, these all will be added to TensorFlowTestCase. Therefore, you should use TensorFlowTestCase when writing tests for tensorflow. TensorFlowTestCase is defined here: tensorflow/python/framework/test_util.py

Write hermetic tests

Hermetic tests do not need any outside resources. They are packed with everything they need, and they just start any fake services they might need. Any services other than your tests are sources for non determinism. Even with 99% availability of other services, network can flake, rpc response can be delayed, and you might end up with an inexplicable error message. Outside services may be, but not limited to, GCS, S3 or any website.