#10601 Centralize all fork+exec in the scheduler. Issue created by jsirois Currently we fork+exec in 2 places once the scheduler is resident in memory: 1. pants/src/rust/engine/process_execution/src/local.rs Lines 244 to 253 in</pantsbuild/pants/commit/05c3fd2d78814338428871ab3553985caef61e46|05c3fd2> | fn stream<'a, 'b>( | | ----------------------------------------------------------------- | | &'a mut self, | | req: &Process, | | ) -> Result<BoxStream<'b, Result<ChildOutput, String>>, String> { | | self | | .inner | | .stdin(Stdio::null()) | | .stdout(Stdio::piped()) | | .stderr(Stdio::piped()) | | .spawn() | 2. pants/src/rust/engine/process_execution/src/nailgun/nailgun_pool.rs Lines 285 to 307 in</pantsbuild/pants/commit/05c3fd2d78814338428871ab3553985caef61e46|05c3fd2> | impl NailgunProcess { | | ---------------------------------------------------------------------- | | fn start_new( | | name: NailgunProcessName, | | startup_options: Process, | | workdir_path: &PathBuf, | | nailgun_server_fingerprint: NailgunProcessFingerprint, | | build_id: String, | | ) -> Result<NailgunProcess, String> { | | let cmd = startup_options.argv[0].clone(); | | // TODO: This is an expensive operation, and thus we info! it. | | // If it becomes annoying, we can downgrade the logging to just debug! | | info!( | | "Starting new nailgun server with cmd: {:?}, args {:?}, in cwd {:?}", | | cmd, | | &startup_options.argv[1..], | | &workdir_path | | ); | | let handle = std:processCommand:new(&cmd) | | .args(&startup_options.argv[1..]) | | .stdout(Stdio::piped()) | | .stderr(Stdio::piped()) | | .current_dir(&workdir_path) | | .spawn(); | This needs to be centralized to continue to confidently support a fix for ETXTBSY when launching user processes that have the binary as part of their input digest (#10507). For now we are getting by by relying on the fact that the second place above - in the

process_execution::nailgun

module - is not used / dead code in practice. pantsbuild/pants