Bazel: Provide tempdir in genrule()

[kylecordes]

Description of the problem / feature request:

A genrule is sometimes used to wrap legacy build tooling. Such tooling sometimes demands a temp directory in which it can write scratch files.

A workaround is to get this is suggested on StackOverflow (link below). “(1) wrap the entire cmd body in a shell function, and call it (2) create a temp directory as the first thing in the function, store its path in some variable (3) add a trap to clean up the temp directory.”

That's not a great thing to reinvent at each point of use, and it may sometimes put the tempdir in a non-optimal place (like an expensive sandboxfs rather than fast cheap tmpfs).

Thus this feature request: provide a tempdir mechanism in genrule to:

• obtain a temp dir, either $(temp) in the cmd, or in an environment variable
• automatically cleanup tempdir contents (on success or failure)
• enable other machinery (remote build executors, for example) to control where these temps land (i.e. the right filesystem etc for the situation).

Feature requests: what underlying problem are you trying to solve with this feature?

See above. Some tools demand a temp directory, and sometimes we need to wrap old things in a genrule() rather than create new Bazel rules for every part of a sprawling project.

Have you found anything relevant by searching the web?

https://stackoverflow.com/questions/55001748/do-bazel-genrules-offer-a-temp-directory

[kylecordes]

An extra wrinkle: some tools may demand an absolute path of a tempdir, which is unfortunate as it violates the general Bazel guidance to use only relative paths.

[aiuto]

Assigning to RemoteExec team, but this kind of thing is really a collaboration with LocalExec as well.
It is a pretty big design space.
I think the concept might apply to all rules, not just genrule. There are often helpers that will be use space at the path specified in $TMPDIR. If you know the behavior of the helper, you could tune the action to use temp space in memory or disk.

[kylecordes]

A couple more idea variants:

• Provide two tempdirs to each rule: one that is small and fast, the other large and slow.
• Provide one tempdir, but have a rule setting to state the desired size of the temp storage - which the strategy code would use to choose what kind of tempdir fs to provide.

[davido]

Gerrit Code Review project extracted its genrule2 Starlark macro in bazlets repository that would defined ROOT and TMP for you and that you can use like this:

load("//tools/bzl:genrule2.bzl", "genrule2")

genrule2(
    name = "%s-static" % name,
    cmd = " && ".join([
        "mkdir -p $$TMP/static",
        "unzip -qd $$TMP/static $(location %s__gwt_application)" % name,
        "cd $$TMP",
        "zip -qr $$ROOT/$@ .",
    ]),
    tools = [":%s__gwt_application" % name],
    outs = ["%s-static.jar" % name],
)

The definition is here: [1].

def genrule2(cmd, **kwargs):
    cmd = " && ".join([
        "ROOT=$$PWD",
        "TMP=$$(mktemp -d || mktemp -d -t bazel-tmp)",
        "(" + cmd + ")",
    ])
    native.genrule(
        cmd = cmd,
        **kwargs
    )

[1] https://github.com/GerritCodeReview/bazlets/blob/master/tools/genrule2.bzl#L19-L28

[kylecordes]

@davido Excellent. What do you think of the advice from the StackOverflow page commenter, "add a trap to clean up the temp directory"? Without that, and unless Bazel itself hooks mktemp (which as far as I know, it does not), it seems genrule2() as implemented above will leave tempfiles to accumulate until eventual OS-level cleanup (at restart, etc.)?

[davido]

As the matter of fact the quoted Starlark macro could be improved:

diff --git a/tools/bzl/genrule2.bzl b/tools/bzl/genrule2.bzl
index 3113022104..d0b0969438 100644
--- a/tools/bzl/genrule2.bzl
+++ b/tools/bzl/genrule2.bzl
@@ -21,6 +21,7 @@ def genrule2(cmd, **kwargs):
         "ROOT=$$PWD",
         "TMP=$$(mktemp -d || mktemp -d -t bazel-tmp)",
         "(" + cmd + ")",
+        "rm -rf $$TMP",
     ])
     native.genrule(
         cmd = cmd,

[kylecordes]

In case anyone comes along and needs this, it's also pretty easy to use trap to ensure the tempdir is always cleaned up. Here is a particularly helpful page about it.

https://www.putorius.net/using-trap-to-exit-bash-scripts-cleanly.html