Do not print errors when repository rules are interrupted

[fmeum]

When the user interrupts a build with Ctrl+C during the download or extraction of a file, no Starlark error should be printed. This is achieved by propagating InterruptedExceptions instead of failing the repository rule with an IOException.

[Wyverald]

I'm going to trust that you did some testing, as I'm not really sure how the interaction works here :)

[fmeum]

I'm going to trust that you did some testing, as I'm not really sure how the interaction works here :)

I tested interrupting downloads and extractions and they both interrupted cleanly.

Based on the contract of SkyFunction#compute, swallowing InterruptedExceptions caused by you and not your dependencies does seem wrong, which gives me some confidence that this is also conceptually correct.

I would still suggest not to cherry pick this into 7.0.0.

[katre]

Another case of #18629.

[fmeum]

@lberki While taking a deeper look at how Skyframe handles interrupts, I noticed that it seems to drop the Skyframe compute state only if a SkyFunction errors out (https://cs.opensource.google/bazel/bazel/+/master:src/main/java/com/google/devtools/build/skyframe/AbstractParallelEvaluator.java;l=479;drc=69359279d4ff40b25b6a69219f0576bae19f0297;bpv=1;bpt=1?q=Skyfunc&ss=bazel%2Fbazel), but not when it is interrupted. Do you know what the precise invariant on state maintained through interrupts is?

[lberki]

I don't, not off the bat. Maybe @justinhorvitz does?

[justinhorvitz]

I think we haven't thought to clear the compute state on interrupt because an interrupted skyframe thread typically means that the whole evaluation is being interrupted, and the state cache is per-evaluation. Do you have a case where there's an interrupt but the overall evaluation proceeds?

Tagging @haxorz because he was more involved in the state cache implementation.

[fmeum]

If that's the case then all is well: I was worried about state being preserved across evaluations, and with it the possibly invalid contents of the external repository. Just for the sake of my own education, could you point me to where this per-evaluation state clearing happens?

[justinhorvitz]

AbstractParallelEvaluator#stateCache is initialized per evaluator instance. A new instance is created for each evaluation.

[fmeum]

@bazel-io flag

[meteorcloudy]

@fmeum You can now do @bazel-io flag <version> to indicate which release you want this to be cherry picked. I guess this one will be 7.1

[meteorcloudy]

@bazel-io fork 7.1.0

[iancha1992]

The changes in this PR have been included in Bazel 7.1.0 RC1. Please test out the release candidate and report any issues as soon as possible. If you're using Bazelisk, you can point to the latest RC by setting USE_BAZEL_VERSION=last_rc.
Thanks!