Isolated Testing for SWEs
This page explains how to convert a googletest executable into a fully isolated test. This is done by describing runtime dependencies for a test.
For information about the infrastructure itself and the roadmap, see Isolated Testing Infrastructure.
Note: If you're a Googler who needs to perform a manual test on a platform for which you cannot readily build (e.g. Windows Developer with no development environment on your Mac book), follow these instructions to get the build from a try run.
The isolate project is all the file formats, tools and server code to make this work fast and efficiently. Every test executable list the data files it depends on, which allows the testers bots to only download the files they need instead of the whole Chromium source tree.
They are described in BUILD.gn files.
Goal: distribute tasks fast and efficiently in an heterogeneous fleet of bots.
The swarming project is all the tools, server and bot code to run a step (like a unit test) on a remote bot and get results back. It has native google-test sharding capability.
By reducing the amount of data that we need to transfer to the tester machines, it becomes much easier to increase the number of tester bots that we have using Swarming (part of LUCI).
- Make sure the test exist in a BUILD.gn file.
- Add missing items to the "data" section until the test runs locally isolated via mb.py run.
- Add the test to the corresponding json file in //testing/buildbot/.
- All the required data files, executables started by the test
executable* required by the test to succeed are listed as a
- For example,
chrome.exe, which needs
- For example,
Expectations of the tests:
- Must not do directory walking or otherwise try to guess what should be tested.
- Must not edit any input file.
- Must not write at all in the current directory. It must only use the temporary directory for temporary files.
- Visit the try job build, e.g. from https://ci.chromium.org/.
- Search for [trigger] my_test and click Swarming task "shard #0" link
- Scroll down a bit, follow the instructions in the Reproducing this Task Locally section
echo gn > out/Release/mb_type # Must be done once to avoid mb.py from performing a clobber tools/mb/mb.py run //out/Release base_unittests # Builds, generates .isolate (via gn desc runtime_deps), and runs (via "isolate.py run")
Right now, only users Chromium team members can use the infrastructure directly. Note that the whole Swarming infrastructure is open source so if any other company would help to recreate the same infrastructure internally, send us a note at firstname.lastname@example.org
By login first, you have access tokens so that the following commands do not have to constantly prompt for your identity. You only need to do this once:
If you are running through a text only session on a remote machine, append
python3 tools/mb/mb.py run --swarmed out/Release \ base_unittests \ -- --gtest_filter=<test filter>
This will perform the steps below, picking defaults based on your local setup. If it fails, you can re-run the last command run by mb.py to see the cause of the failure. You may need to login again.
If you want to pick different defaults, e.g., run tests on Windows-11 from a Windows 10 machine, you need to use --no-default-dimensions and explicitly specify some dimensions, e.g.,
python3 tools/mb/mb.py run \ --swarmed out/Release \ -d os Windows-11 \ -d pool chromium.tests \ --no-default-dimensions base_unittests \ -- --gtest_filter=<test filter>
Or, you can run the individual steps below manually.
The isolate file describes what are the files that needs to be mapped on the Swarming bot. It is generated via GN "data" and "data_deps" statements, and is processed by mb.py:
ninja -C out/Release base_unittests.exe echo gn > out/Release/mb_type # Must be done once to avoid mb.py from performing a clobber python3 tools\mb\mb.py isolate //out/Release base_unittests # Creates out/Release/base_unittests.isolate
The isolated file contains the SHA-1 of each input files. It is archives along all the inputs to the Isolate server. Since the isolate server is a content-addressed cache, only the ones missing from the cache are uploaded. Depending on your connection speed and the size of the executable, it may take up to a minute:
tools\luci-go\isolate archive \ -i out\Release\base_unittests.isolate \ -cas-instance chromium-swarm
This will output a digest string that you use in step 5.
That's where a task is requested, specifying the isolated (tree of SHA-1 of files to map in):
tools\luci-go\swarming trigger \ -digest <digest from step 4> \ -server chromium-swarm.appspot.com \ -d os=Windows-10-19042 \ -d pool=chromium.tests \ -d cpu=x86-64 -- out\Release\base_unittests.exe
Wait for results. OS currently available:
- Windows-7-SP1 (64 bits)
- Windows-2008ServerR2-SP1 (64 bits)
- Windows-10-15063 (64 bits)
- Ubuntu-14.04 and Ubuntu-16.04 (64 bits)
- Mac-10.13 to Mac-13
For other available --dimension values, look at Swarming bots e.g.: https://chromium-swarm.appspot.com/botlist.
That's it. Feel free to contact the team at email@example.com for any chromium open source specific questions.
-d pool Chromeis needed!
- -d cpu x86-64 is useful to prevent your 64 bit build from being run on 32-bit Windows trybots.
- Running an executable on a swarming bot is documented at in the user guide.
- Monitor the running tasks by visiting chromium-swarm.appspot.com.
It is possible that all the bots are currently fully utilized. In this case, the task is in PENDING state.
In theory yes, in practice please don't and keep the list to the strict minimum. The goal is not to run the tests more slowly and having the bots download 20 gb of data. Reasons includes:
- Isolate Server is optimized for < 50000 files scenario. There's a 2ms/file cost per cache hit. So for example, layout tests are currently out of the use case since there's > 80000 files.
- It's always possible to go coarser but it's much harder to get back stricter.
The .isolated files are generated when you build the isolation specific version of a target, e.g. out/Debug or out/Release. The isolation target name is just the normal target name with a _run added to the end.
If you have a test that passes locally and fails on some trybots, then it could be related.
This error can be seen when a browser test js file is not found:
TypeError: Cannot read property 'testCaseBodies' of undefined