We just ran into a subtle bug in remote execution ...
# development
f
We just ran into a subtle bug in remote execution for the special case of requesting the entire working directory as the
output_directories
. We've opened https://github.com/pantsbuild/pants/issues/23372 for it that has a proposed quick fix in the 1st comment (h/t Claude). I think that fix is probably good enough for pushing into 2.32.0 (hopefully!), but for 2.33 we might want to think about a more extensive improvement to the remote execution & cache handling. Rumor is that BuildBarn is going to drop support for
output_files
and
output_directories
and force everything to
output_paths
.
There's also a side question of why pants writes the entry into the cache after getting the answer back from the remote engine. At least BB has an "auto-write" mechanism. I don't know about other RE engines.
(and, in case anyone was wondering, of course it was discovered over the holiday weekend, blocking CI for an end user)
😢 1
Can someone please poke CI for https://github.com/pantsbuild/pants/pull/23381 ?
✅ 1
w
Did this only get merged into 2.32.x or also into 2.33.0?
f
Only 2.32.x. I saw @fast-nail-55400 (sorry for the early morning tag) has a PR open to make adjustments for the new rev of the RE protocol that'll require redoing this fix differently.
f
I got stuck on that PR thinking about how to deal with remote cache entries.
Tentative fix: • Introduce BuildBarn end-to-end integration test: github.com/pantsbuild/pants/pull/23484 • Use `output_paths`: github.com/pantsbuild/pants/pull/23294