Ergonomic, hermetic, user-friendly software build system for Python, Go, Java, Kotlin, Protobuf, Scala, Shell. Pants lets you fearlessly scale up your codebase!

Pants

in particular, any time we fork with a Scheduler already instantiated

I think the other one (in goal_runner) wasn’t actually creating a scheduler yet before forking, so that should be it iirc

<https://github.com/pantsbuild/pants/blob/master/src/python/pants/bin/goal_runner.py#L181> -&gt; <https://github.com/pantsbuild/pants/blob/master/src/python/pants/pantsd/pants_daemon.py#L186>

ok. but in the <https://github.com/pantsbuild/pants/blob/master/src/python/pants/pantsd/pants_daemon.py#L186> case, the scheduler isn't supposed to have been created?

...ok, i see the threads in goal_runner. looking.

yeah afaict it forks before calling _init_graph()

but it passed in `self._daemon_graph_helper` ?

I believe that’s only used in the case of a running daemon for graph reuse

defaults to None here: <https://github.com/pantsbuild/pants/blob/master/src/python/pants/bin/local_pants_runner.py#L20> and isnt overridden via <https://github.com/pantsbuild/pants/blob/master/src/python/pants/bin/pants_runner.py#L46>

<@U054MQUPL> : so yea, definitely agreeing now that a separate entrypoint for the daemon would be amazing

at some point we should consider forcing the daemon mode

it is a highly-impressive-but-difficult-to-grok situation with the forking now

and then the idea of a decoupled native ng client/launcher bin vs daemon entrypoint would make more sense

well, but as you said: `fork-exec` would be clearer, even if it still happened the same way as now

er, if the first fork still happened in the same place, i mean

<@U054MQUPL> : it looks like we do create a scheduler before the first fork; here: <https://github.com/pantsbuild/pants/blob/master/src/python/pants/pantsd/subsystem/pants_daemon_launcher.py#L130>

so maybe i add a generic `Service.post_fork`?

could potentially move _setup_services() into `PantsDaemon` and do that post-fork?

iirc, the one bit there we would want to preserve is launching watchman pre-fork so we can fail the parent pants run if it won’t launch for wahtever reason

but the rest seems like it could easily move across the fork line

<@U054MQUPL> : what's the entry point post fork there? it looks like all of `pants_daemon_launcher` runs in the parent process

so then something in `src/python/pants/pantsd/pants_daemon.py` ?

<https://github.com/pantsbuild/pants/blob/master/src/python/pants/pantsd/pants_daemon.py#L186>

...maybe in `post_fork_child`, for self._services, post_fork ?