testsharder is a tool that takes a single list of tests and splits that list into one or more new lists (“shards”) so that each shard only contains tests that run on a single device type.
Specifically, testsharder takes as input (via the
-build-dir flag) a fuchsia build out directory containing a tests.json produced by the fuchsia build. tests.json contains a list of JSON objects conforming to the schema of the
TestSpec struct from
testsharder's output is another JSON file, whose location is specified by the
-output-file flag. This file contains a list of JSON objects conforming to the schema of the
Shard struct from
//tools/integration/testsharder/shard.go. Each shard has a name that will end in a number if there are multiple shards with the same device type, e.g. “QEMU-(1)”.
testsharder‘s primary consumer is the infra recipes, specifically the
fuchsia/build recipe. That recipe uses testsharder’s output to schedule a set of Swarming tasks, each of which runs the tests from one shard.
testsharder has two flags to control the size of shards:
-max-shard-size specifies a target number of tests for each shard. If set, testsharder will divide the input tests.json into shards of approximately this size.
-target-duration-secs specifies a target duration for each shard. If set, testsharder will divide the input tests.json into shards whose tests are expected to complete in approximately this amount of time. See the “Sharding by time” section below for details on how testsharder uses this flag.
It's invalid to specify both of these flags at the same time.
testsharder also has a
-max-shards-per-environment flag that sets a hard maximum on the number of shards for each device type. If it's impossible to divide the input
tests.json into fewer than the maximum number of shards while satisfying
-max-shard-size, some or all of the shards will exceed
tests.json, testsharder also reads a
test_durations.json file from the build output directory. This file is copied into the output directory from the
//integration/infra/test_durations directory, which contains a file for every infra builder that runs tests using the
fuchsia/fuchsia recipe. These files are updated periodically with recent duration data for each test.
Each builder takes the path to its corresponding test duration file as an argument and passes that path into the build via the
test_durations_file GN argument. The file is then renamed to
test_durations.json and copied into the output directory by the
test_durations target from the root
testsharder then reads
test_durations.json from the build output directory and uses the duration data to divide tests into shards of approximately equal expected duration, using a greedy approximation algorithm for static multiprocessor scheduling.
Within a shard, tests are ordered by expected duration descending (longest tests first). However, this should be considered an infra implementation detail and is subject to change.
If any test does not appear in
test_durations.json (for example, because the duration files have not been updated since it was added) then testsharder will use the duration file entry with name
* for that test. That entry‘s data is an average of all existing tests’ data. So any newly added tests will be scheduled close to the middle of one of the shards.
Given an input
-multipliers file, testsharder's output is completely deterministic.
However, adding, deleting, or renaming a test can completely change the output. For example, adding a new test that takes an average amount of time will add that test to the middle of one of the shards. This will generally push some faster tests into new shards, leading to a complete reshuffling of all faster tests between shards.
Updates to the duration files may also affect testsharder‘s output, so duration file updates go through CQ before landing to ensure that they don’t cause any ordering-related test breakages.
Testsharder has an optional
-modifiers flag that allows customization of test retry strategies and sharding, such as the
MULTIPLY feature and separate sharding of affected tests in CQ.
-modifiers should point to a JSON file containing a list of objects conforming to the
TestModifier schema (see
TestModifier applies to a single test or a subset of tests (in order to support fuzzy matching for
MULTIPLY) and determines how many times the test should run, and whether the test must pass on every run to be considered successful, or whether it need only pass once.