commit 6a97e137a4f61e53591f8b7737b04b841ca04ece
author Benjamin Peterson <bp@benjamin.pe> Mon Jan 28 03:36:16 2019 -0800
committer Copybara-Service <copybara-piper@google.com> Mon Jan 28 03:37:42 2019 -0800
tree dfbaf86677ad2e27d89ae00c35ed56f17842a36a
parent 8ccbb0f76fe5a67e2072b35e5702b52423e07d99

Fix a race in ByteStreamUploader deduplication.

The callback to add a completed digest to the uploadedBlobs set was added to the result future of uploadBlobAsync. The non-deterministic ordering of future callbacks meant a client of ByteStreamUploader could observe a completed upload before the digest was added to uploadedBlobs. The solution is to chain futures so that the digest is placed in uploadedBlobs before the final future fires.

I noticed this race caused a test failure in CI: https://storage.googleapis.com/bazel-buildkite-artifacts/f1b54300-cf87-4516-8767-410b93ab1ee9/src/test/java/com/google/devtools/build/lib/remote-tests/test.log

It also was readily reproducible for me with this command:
```
$ bazel test //src/test/java/com/google/devtools/build/lib:remote-tests --runs_per_test 50 --test_filter deduplicationOfUploadsShouldWork
There was 1 failure:
1) deduplicationOfUploadsShouldWork(com.google.devtools.build.lib.remote.ByteStreamUploaderTest)
expected: 1
but was : 2
        at com.google.devtools.build.lib.remote.ByteStreamUploaderTest.deduplicationOfUploadsShouldWork(ByteStreamUploaderTest.java:731)
        at java.base/jdk.internal.reflect.NativeMethodAccessorImpl.invoke0(Native Method)
        at java.base/jdk.internal.reflect.NativeMethodAccessorImpl.invoke(NativeMethodAccessorImpl.java:62)
        at java.base/jdk.internal.reflect.DelegatingMethodAccessorImpl.invoke(DelegatingMethodAccessorImpl.java:43)
```

We can also simplify things by removing the second parameter of startAsyncUpload. This in turn permits us to delete a Retrier.executeAsync overload.

Closes #7256.

PiperOrigin-RevId: 231190201