Tiago Quelhas

> Note that the files that symlinks point to are not in the inputs of the action that creates the symlink, and i don't think it matters much. I think...

(To further clarify: if I understand the scenario correctly, your build is already non-hermetic; without the input dependency, nothing guarantees that the output pointed to by the symlink is produced...

> It's just odd given the action succeeds with other spawn strategies I agree that it's unfortunate that the behavior depends on the contents of the output tree. We could...

This is trivial to repro. The issue is this [logic](https://cs.opensource.google/bazel/bazel/+/master:src/main/java/com/google/devtools/build/lib/exec/StandaloneTestStrategy.java;l=819;drc=874a050aa7ab3f2034bbd4afb3a5fbf4ed5109d1) in StandaloneTestStrategy that appends the test stderr file into the stdout file and then deletes the former. With `--experimental_remote_cache_async`, this...

In addition, the `test.err` files are guaranteed to always be empty, because `test-setup.sh` [redirects](https://cs.opensource.google/bazel/bazel/+/master:tools/test/test-setup.sh;l=18;drc=8e639dfa8c7aa0723123ee90dc933d528a4f2b68) stderr to stdout. I believe the merge logic is just a last resort in case the...

...except on Windows, where the test wrapper doesn't redirect. Sigh.

I'm having trouble to repro (assuming `@//:third_party/folly/build_file.bazel` the file containing the `cc_library`, and the build command is `bazel build @folly//:header_only`). Maybe something else in your workspace is required to trigger...

Please reopen with a repro.

> @coeuvre @tjgq If I understand the logic around `RemoteCacheClient` correctly, Bazel will not deduplicate inflight requests to the remote cache/executor - which makes sense without path mapping, as an...

Thinking some more about it: is this really worth doing if we assume that the remote executor is able to deduplicate identical concurrent requests on its own? The spec doesn't...