pantsbuild / pants / 33038133062

93%

push

github

Restart pantsd when `[GLOBAL].plugins` changes (#23650)

Fixes #23649.

`[GLOBAL].plugins` was not marked `daemon=True`, so changing it only
reinitialized the scheduler. `pantsd` resolves plugins into its own
process, and `PluginResolver.resolve` derives each resolve's constraints
from that process's `sys.path`, so a surviving daemon keeps constraining
later resolves to the versions it resolved first. Fingerprinting the
option for the daemon replaces the process instead, so the next resolve
starts from a clean `sys.path`.

### Validation

Built from source at 2.34.0.dev3 and run against the private monorepo
where this was first hit, which pins `plugins = ["botocore==1.43.73"]`.
Same repo, same sequence, bumping the pin to `botocore==1.43.74`, with
only the patch differing.

Without the patch the daemon is reused (pid 2314187 on both runs) and
the resolve fails:

```
pip: ERROR: ResolutionImpossible
pip:      The user requested botocore==1.43.74
pip:      The user requested (constraint) botocore==1.43.73
```

With the patch the daemon is replaced (pid 2320168 becomes 2320670) and
the same run exits 0.

Supporting checks on that build:

- Repeating a value does not restart the daemon, so this does not
over-invalidate.
- `pants test src/python/pants/pantsd/pantsd_integration_test.py` passes
in 3m35s.
- `cargo test -p pantsd` passes, including `test_find_pantsd`, which
launches a daemon and rediscovers it through `fingerprint_compute`.
- `pants test src/python/pants/option/global_options_test.py
src/python/pants/option/options_bootstrapper_test.py` passes.

The Python and Rust option lists are asserted to match on every
invocation, so the second commit is what lets the first one run at all.
Further detail on the resolve chain is in the comment below.

### LLM assistance

Claude Code (Opus 5) did the code archaeology through
`PluginResolver.resolve` and `PantsDaemonCore.prepare`, wrote the patch
and the test, built the engine from source, and ran the valid... (continued)

5 of 5 new or added lines in 1 file covered. (100.0%)

95490 of 102794 relevant lines covered (92.89%)

3.66 hits per line