Write a Sharded IRemoteTest Test Runner

When writing a test runner, it's important to think about scalability. Ask yourself, "if my test runner had to run 200K test cases" how long would it take?

Sharding is one of the answers available in Trade Federation. It requires splitting all the tests the runner needs into several chunks that can be parallelized.

This page describes how to make your runner shardable for Tradefed.

Interface to implement

The single most important interface to implement to be considered shardable by TF is IShardableTest, which contains two methods: split(int numShard) and split().

If your sharding is going to depend on the number of shards requested, you should implement split(int numShard). Otherwise, implement split().

When a TF test command is executed with sharding parameters --shard-count and --shard-index, TF iterates through all IRemoteTest to look for ones implementing IShardableTest. If found, it will call split to get a new IRemoteTest object to run a subset of test cases for a specific shard.

What should I know about the split implementation?

  • You runner may shard upon some conditions only; in that case return null when you did not shard.
  • Try to split as much as it makes sense: split your runner into unit of execution that makes sense for it. It really depends of your runner. For example: HostTest is sharded at the Class level, each test class is put in a separate shard.
  • If it makes sense, add some options to control the sharding a little bit. For example: AndroidJUnitTest has a ajur-max-shard to specify the maximum number of shards it could split in, regardless of the number requested.

Detailed example implementation

Here is an example code snippet implementing IShardableTest you can reference. The full code is avaiable at (https://android.googlesource.com/platform/tools/tradefederation/+/refs/heads/master/src/com/android/tradefed/testtype/InstalledInstrumentationsTest.java)

/**
 * Runs all instrumentation found on current device.
 */
@OptionClass(alias = "installed-instrumentation")
public class InstalledInstrumentationsTest
        implements IDeviceTest, IResumableTest, IShardableTest {
    ...

    /** {@inheritDoc} */
    @Override
    public Collection<IRemoteTest> split(int shardCountHint) {
        if (shardCountHint > 1) {
            Collection<IRemoteTest> shards = new ArrayList<>(shardCountHint);
            for (int index = 0; index < shardCountHint; index++) {
                shards.add(getTestShard(shardCountHint, index));
            }
            return shards;
        }
        // Nothing to shard
        return null;
    }

    private IRemoteTest getTestShard(int shardCount, int shardIndex) {
        InstalledInstrumentationsTest shard = new InstalledInstrumentationsTest();
        try {
            OptionCopier.copyOptions(this, shard);
        } catch (ConfigurationException e) {
            CLog.e("failed to copy instrumentation options: %s", e.getMessage());
        }
        shard.mShardIndex = shardIndex;
        shard.mTotalShards = shardCount;
        return shard;
    }
    ...
}

This exampl simply creates a new instance of itself and sets shard parameters to it. However, the splitting logic can be totally different from test to test; and as long as it is deterministic and yields collectively exhaustive subsets, it is okay.

Independence

Shards need to be independent! Two shards created by your implementation of split in your runner should not have dependencies on each other or share resources.

Shards splitting needs to be deterministic! This is also mandatory, given the same conditions, your split method should always return the exact same list of shards in the same order.

NOTE: Since each shard can run on different TF instances, it is critical to ensure the split logic yields subsets that are mutually exclusive and collectively exhaustive in a deterministic manner.

How to shard a test locally

To shard a test on a local TF, you can simply add the --shard-count option to the command line.

tf >run host --class com.android.tradefed.UnitTests --shard-count 3

Then TF will automatically spawn commands for each shard and run them.

tf >l i
Command Id  Exec Time  Device          State
3           0m:03      [null-device-2]  running stub on build 0 (shard 1 of 3)
3           0m:03      [null-device-1]  running stub on build 0 (shard 0 of 3)
3           0m:03      [null-device-3]  running stub on build 0 (shard 2 of 3)

Test result aggregation

Since TF does not do any test result aggregation for sharded invocations, you need to make sure your reporting service supports it.