Force output checking for incremental run commands without the bytes.

[tjgq]

When building without the bytes, outputs are only materialized when either (1) an action prefetches them, or (2) they've been explicitly requested. When both --remote_download_minimal and --noexperimental_check_output_files are set, this presents a problem for a build command followed by a run command for the same target: the build command won't download the outputs and the run command won't check for their presence, proceeding without them. A non-incremental run command works, because the outputs are considered top-level even in minimal mode.

This is only a problem for a run command because it exists outside of action execution, and doesn't get a chance to prefetch inputs; it would have been better to implement the run command by injecting a specially crafted action into the build graph, but that will have to wait for another day. The present fix might theoretically slow things down if output checking is truly unnecessary, but I doubt that would cause a significant regression in practice.

Fixes #20843.

[tjgq]

I don't think I can make the check narrower, because the remote options can change between builds without invalidating the Skyframe state. (But maybe you can see a way.)

[tjgq]

@bazel-io fork 7.0.1

[iancha1992]

@bazel-io fork 7.1.0

[tomdegoede]

@tjgq This change slows down our bazel run invocations by 5 seconds compared to bazel build. I think your theoretical case of switching remote options between builds will be very uncommon in practice, where these settings are pinned in the .bazelrc. Could you reconsider narrowing down the check to also return NOTHING_MODIFIED when remote download is set to all? This would allow us to speed up common dev tasks by 5 seconds.

[tjgq]

I think your theoretical case of switching remote options between builds will be very uncommon in practice, where these settings are pinned in the .bazelrc

Switching options is not the problem; the problem is a bazel build followed by a bazel run, even if the output mode hasn't changed. The test case added by this PR demonstrates this.

Could you reconsider narrowing down the check to also return NOTHING_MODIFIED when remote download is set to all? This would allow us to speed up common dev tasks by 5 seconds.

To be absolutely clear, the condition for NOTHING_MODIFIED isn't whether --remote_download_all is set in the current invocation; the condition for a complete fix is whether the previous invocation either set --remote_download_all, or it set --remote_download_toplevel and the bazel run target was toplevel. It can definitely be done, but it requires keeping this additional state somewhere, so it's not a localized fix.

(Switching between output modes might be infrequent, but we don't have the luxury of saying "we'll be correct unless you happen to do something unusual".)

As mentioned in the PR description, there's also the alternative of reimplementing bazel run as a specifically crafted build action, but I think that's more difficult (though more principled).

[tjgq]

To be absolutely clear, the condition for NOTHING_MODIFIED isn't whether --remote_download_all is set in the current invocation; the condition for a complete fix is whether the previous invocation either set --remote_download_all, or it set --remote_download_toplevel and the bazel run target was toplevel. It can definitely be done, but it requires keeping this additional state somewhere, so it's not a localized fix.

It turns out that I'm wrong, and the fix is much simpler than anticipated: the changes in 4c42edb also made this workaround unnecessary, so we can literally revert this PR and fix the performance issue while preserving correctness (if the integration test is to be believed). I'll send it in soon.

[tomdegoede]

Amazing! Thanks @tjgq :)