Add watch parameter to {module,repo}_ctx.path
#23386
Conversation
github-actions
bot
added
team-ExternalDeps
|
FYI @aignas |
src/test/java/com/google/devtools/build/lib/bazel/repository/starlark/BUILD
Show resolved
Hide resolved
|
Will this be backported to |
|
@bazel-io fork 7.4.0 |
| // This can never happen because we check it in the Starlark interpreter. | ||
| default -> throw new IllegalArgumentException("expected string or label for path"); | ||
| }; | ||
| if (shouldWatch == ShouldWatch.YES) { |
There was a problem hiding this comment.
why not just maybeWatch(result, shouldWatch) here?
There was a problem hiding this comment.
I added a comment, this would be backwards incompatible (caught by one of the new tests).
There was a problem hiding this comment.
(sorry for the delay)
You're talking about testPath_watchAuto_starlarkPathAndAbsolutePathNotWatched, right? I agree that's a change in behavior, but I don't know if it counts as backwards-incompatible. Didn't we do the same behavior change for repo rules?
Plus, we already have the condition where watch = 'auto' means that paths outside the workspace aren't watched (https://cs.opensource.google/bazel/bazel/+/master:src/main/java/com/google/devtools/build/lib/bazel/repository/starlark/StarlarkBaseExternalContext.java;drc=6052ad65e4185f19944e5138e9b70f6a4eaf9e76;l=1457). Isn't that enough?
There was a problem hiding this comment.
The PR title was misleading, it adds this parameter to both repo and module context, so we don't have prior art for this function to refer to.
I agree that the change is minor, but I think that it would end up breaking the code in https://github.com/bazelbuild/rules_python/pull/2135/files#diff-ca74c876e894b0bd25547ca666e562281e5c8fb2abc4a51446f410371224aea2 by reintroducing a dependency into the lockfile. Depending on a sibling label, manipulating the path and then obtaining a StarlarkPath for it via this method is the kind of existing "pattern" that changing this would break.
I'm not in love with that edge case and we could decide to point users to bazel_features instead, but it would still be breaking.
There was a problem hiding this comment.
hmm, I see. Could you add some more details in the PR description about the motivation of this change? It got lost in the commotion.
I could see an API where auto (or whatever the default value for [mr]ctx.path(watch=) ends up being) means "don't watch". Or just make it default to no. Or actually, maybe just make [mr]ctx.watch return the path?
There was a problem hiding this comment.
oh, I remember now -- so rctx.path(label) used to watch, and rctx.path(path) and rctx.path(string) used to not watch.
This PR retains this old rctx.path(X) watch behavior (which is based on the type of X, not the actual location of X or whether it's rctx or mctx). It just allows the caller to be explicit about watching or not by specifying watch='yes' or watch='no'. The awkward part is that watch='auto' means a different thing for rctx.path() than rctx.read(), rctx.extract(), etc.
I wonder if we'd want rctx.path(X) to never watch (just use rctx.watch() instead). That's obviously backwards-incompatible because some people rely on rctx.path(label) watching, but maybe it's worth introducing an incompatible flag flip for? Undecided.
Also, the fact that rctx.read(label, watch='no') actually still watches is a bit of a subtle hiccup. I'll need to think about that too...
There was a problem hiding this comment.
Now that I've thought more about it, I think that I would prefer the following changes:
ctx.pathnever watches and doesn't have awatchparameter. It only ever watched because that was the only way to watch anything at all. It doesn't make much sense that resolving a label to a path is linked to caring about the content at that path.ctx.watchcan replace it.ctx.read(label, watch = "no")shouldn't watch.
The latter clearly violates the intent, I would consider changing that behavior a bug fix and don't think it would break anyone.
The former is more impactful to change as it has been the only way to watch for a long time. rules_go would definitely need to change. Luckily ctx.watch can be trivially feature detected.
There was a problem hiding this comment.
I fully agree. Just a bit nervous about not having an incompatible flag for it. (so maybe not so "fully" 😅) Adding an incompatible flag for this one isn't super bad, right? It's just a single if in StarlarkBaseExternalContext so shouldn't be a lot of maintenance burden. And we can make it no-op in Bazel 9.
There was a problem hiding this comment.
Yeah, I agree the change for path should come with a flag. But since it will mostly require fixing in rulesets, not user projects, we would need to look into that regardless.
fmeum
changed the title
watch parameter to module_ctx.pathwatch parameter to {module,repo}_ctx.path
|
Closing in favor of #23750 |
github-actions
bot
removed
the
awaiting-review
No description provided.