Deterministic builds

Chromium‘s build is deterministic. This means that building Chromium at the same revision will produce exactly the same binary in two builds, even if these builds are on different machines, in build directories with different names, or if one build is a clobber build and the other build is an incremental build with the full build done at a different revision. This is a project goal, and we have bots that verify that it’s true.

Furthermore, even if a binary is built at two different revisions but none of the revisions in between logically affect a binary, then builds at those two revisions should produce exactly the same binary too (imagine a revision that modifies code chrome/ while we‘re looking at base_unittests). This isn’t enforced by bots, and it‘s currently not always true in Chromium’s build -- but it‘s true for some binaries at least, and it’s supposed to become more true over time.

Having deterministic builds is important, among other things, so that swarming can cache test results based on the hash of test inputs.

This document currently describes how to handle failures on the deterministic bots.

There's also https://www.chromium.org/developers/testing/isolated-testing/deterministic-builds; over time all documentation over there will move to here.

Handling failures on the deterministic bots

This section describes what to do when compare_build_artifacts is failing on a bot.

The deterministic bots make sure that building the same revision of chromium always produces the same output.

To analyze the failing step, it's useful to understand what the step is doing.

There are two types of checks.

  1. The full determinism check makes sure that build artifacts are independent of the name of the build directory, and that full and incremental builds produce the same output. This is done by having bots that have two build directories: out/Release does incremental builds, and out/Release.2 does full clobber builds. After doing the two builds, the bot checks that all built files needed to run tests on swarming are identical in the two build directories. The full determinism check is currently used on Linux and Windows bots. (Deterministic Linux (dbg) has one more check: it doesn‘t use reclient for the incremental build, to check that using reclient doesn’t affect built files either.)

  2. The simple determinism check does a clobber build in out/Release, moves this to a different location (out/Release.1), then does another clobber build in out/Release, moves that to another location (out/Release.2), and then does the same comparison as done in the full build. Since both builds are done at the same path, and since both are clobber builds, this doesn‘t check that the build is independent of the name of the build directory, and it doesn’t check that incremental and full builds produce the same results. This check is used on Android and macOS, but over time all platforms should move to the full determinism check.

Understanding compare_build_artifacts error output

compare_build_artifacts prints a list of all files it compares, followed by ": None" for files that have no difference. Files that are different between the two build directories are followed by ": DIFFERENT(expected)" or ": DIFFERENT(unexpected)", followed by e.g. "different size: 195312640 != 195311616" if the two files have different size, or by e.g. "70 out of 5091840 bytes are different (0.00%)" if they're the same size.

You can ignore lines that say ": None" or ": DIFFERENT(expected)", these don't turn the step red. ": DIFFERENT(expected)" is for files that are known to not yet be deterministic; these are listed in src/tools/determinism/deterministic_build_ignorelist.pyl. If the deterministic bots turn red, you usually do not want to add an entry to this list, but figure out what introduced the nondeterminism and revert that.

If only a few bytes are different, the script prints a diff of the hexdump of the two files. Most of the time, you can ignore this.

After this list of filenames, the script prints a summary that looks like

Equals:           5454
Expected diffs:   3
Unexpected diffs: 60
Unexpected files with diffs:

followed by a list of all files that contained ": DIFFERENT(unexpected)". This is the most interesting part of the output.

After that, the script tries to compute all build inputs of each file with a difference, and compares the inputs. For example, if a .exe is different, this will try to find all .obj files the .exe consists of, and try to compare these too. Nowadays, the compile step is usually deterministic, so this can usually be ignored too. Here's an example output:

fixed_build_dir C:\b\s\w\ir\cache\builder\src\out\Release exists. will try to use orig dir.
Checking verifier_test_dll_2.dll.pdb difference: (1 deps)

Diagnosing bot redness

Things to do, in order of involvedness and effectiveness:

  • Look at the list of files following "Unexpected files with diffs:" and check if they have something in common. If the blame list on the first red build has a change to that common thing, try reverting it and see if it helps. If many, seemingly unrelated files have differences, look for changes to the build config (Ctrl-F “.gn”) or for toolchain changes (Ctrl-F “clang”).

  • The deterministic bots try to upload a tar archive to Google Storage. Use gsutil.py ls gs://chrome-determinism to see available archives, and use e.g. gsutil.py cp gs://chrome-determinism/Windows\ deterministic/9998/deterministic_build_diffs.tgz . to copy one archive to your workstation. You can then look at the diffs in more detail. See https://bugs.chromium.org/p/chromium/issues/detail?id=985285#c6 for an example.

  • Try to reproduce the problem locally. First, set up two build directories with identical args.gn. Then do a full build at the last known green revision in the first build directory:

    $ gn clean out/gn
    $ autoninja -C out/gn base_unittests
    

    Then, sync to the first bad revision (make sure to also run gclient sync to update dependencies), do an incremental build in the first build directory and a full build in the second build directory, and run compare_build_artifacts.py to compare the outputs:

    $ autoninja -C out/gn base_unittests
    $ gn clean out/gn2
    $ autoninja -C out/gn2 base_unittests
    $ tools/determinism/compare_build_artifacts.py \
         --first-build-dir out/gn \
         --second-build-dir out/gn2 \
         --target-platform linux
    

    This will hopefully reproduce the error, and then you can binary search between good and bad revisions to identify the bad commit.

Things not to do:

  • Don‘t clobber the deterministic bots. Clobbering a deterministic bot will turn it green if build nondeterminism is caused by incremental and full clobber builds producing different outputs. However, this is one of the things we want these bots to catch, and clobbering them only removes the symptom on this one bot -- all CQ bots will still have nondeterministic incremental builds, which is (among other things) bad for caching. So while clobbering a deterministic bot might make it green, it’s papering over issues that the deterministic bots are supposed to catch.

  • Don't add entries to src/tools/determinism/deterministic_build_ignorelist.pyl. Instead, try to revert commits introducing nondeterminism.