Index traits using multiple child processes. #95
Conversation
Run trait importing in parallel.
…he parent to immediately exit with a non-zero status.
# Conflicts: # src/CommandLineRunner.php
| * @param array $toProcess | ||
| * @return array | ||
| */ | ||
| public function partition(Config $config, array $toProcess): array { |
There was a problem hiding this comment.
pulling out these methods makes this much easier to understand.
| use BambooHR\Guardrail\Socket; | ||
| use BambooHR\Guardrail\SocketBuffer; | ||
|
|
||
| abstract class ChildProcess { |
There was a problem hiding this comment.
I like this organization of classes, with this one handling the actual i/o blocking.
| if ($status != 0) { | ||
| call_user_func($serverReadCallBack, $this->connections[$childPid], "Child died with non-zero status"); | ||
| echo "Child died with non-zero status!\n"; | ||
| exit($status); |
There was a problem hiding this comment.
If any child crashes or fails, make sure the parent fails. should prevent cihub reeporting success when something goes wrong, I think? Might solve some of the previous issues we've had.
| echo "Error: Unexpected error reading from socket\n"; | ||
| exit(1); | ||
| } | ||
| if ($msg === false || (trim($msg) !== "" && self::CLOSE_CONNECTION == $this->dispatchMessage($socket,$msg))) { |
There was a problem hiding this comment.
$msg === false is unreachable here.
| function handleClientMessage(\Socket $socket, string $message, string ...$params): int { | ||
| assert($message == "TRAITED"); | ||
| $this->output->outputExtraVerbose("Updating class ".$params[0]. " with ".strlen($params[1])." bytes of data\n"); | ||
| $class=unserialize(base64_decode($params[1])); |
There was a problem hiding this comment.
this area is the crux of being able to handle traits in forked processes. The children do the work, but do not update the symbol table (which would be futile, since they don't share the memory space of the parent), but instead re-emit the entire structure of the class after the trait has been applied, and the parent replaces the structure of the class in the symbol table. Keeps the children stateless, only accepting input, and emitting output. The parent is the one that keeps track of state.
There was a problem hiding this comment.
I love that you spend so much time understanding the code you're asked to review. Thank you.
Greatly speeds up trait importing on high core count machines. Also, cleans up the IPC code some.