What's the purpose of setting shard_count in java_test?

[GinFungYJF]

I want to make "bazel test" run the methods in a single java class concurrently in order to speed up the test. And then, I set the shard_count=n(n is the total number of methods in the java class) in java_test rule. But I found that setting shard_count takes more time than no setting shard_count. So, is the option shard_count aimed at speeding up the test? If not, how could I make "bazel test" run the methods in a single java class concurrently? What's more, my test framework is spring and the Test Runner is JUnit4.

Environment info

• Operating System:
  ubuntu 14.04

• Bazel version (output of bazel info release):
  0.4.0

[ulfjack]

Yes, shard_count is aimed at speeding up the test. The idea is indeed that Bazel runs multiple instances with each only running a subset of the tests. That said, if the tests are all very quick-running, then the overhead of starting up multiple Jvms can be larger than the benefits of sharding. Also, if you're running it on a single machine, you're limited by the parallelism you can get on that machine, especially if the tests are internally using multiple threads. Even for single-threaded tests, setting shard_count higher than the number of cores on your machine is (very) unlikely to be a win. At Google, we distribute tests to multiple machines, which allows us to run with larger number of shards.

All that said, if you see no reduction in per-shard times when comparing one vs. multiple shards, then that's an indication that there's still a bug in Bazel or the Java test runner wrt. the sharding protocol. Can you check the logs from the shards to see if each shard is actually only running a single test method?