Performance Issue: Global ReentrantLock in bubus Limits Concurrency #21
Description
I’ve noticed that when running browser-use within a FastAPI application or using the provided parallel_browser.py example, tasks do not scale concurrently as expected.
When initiating multiple agent instances simultaneously, the execution time increases significantly rather than staying relatively constant.
Single task: ~10 seconds.
3 concurrent tasks: ~25 seconds.
10 concurrent tasks: Performance degrades further.
After investigating the bubus library, I found that the event processing logic uses a global reentrant lock (_global_eventbus_lock). This lock ensures that only one event across all EventBus contexts can be processed at any given time.
# Always acquire the global lock (it's re-entrant across tasks)
async with _get_global_lock():
# Process the event
await self.process_event(event, timeout=timeout)
# Mark task as done only if we got it from the queue
if from_queue:
self.event_queue.task_done()
Was this global lock intended to prevent a specific race condition in shared resources, or could it be moved to a per-instance lock?
If the goal is to allow multiple agents to run in true parallel, we should consider:
Removing the global lock if it's not protecting shared state.
Scoping the lock to the instance so that different contexts don't block each other.