Open EmilRex opened 2 weeks ago
After looking at this issue with fresh eyes, I think everything is working as expected. The time increase is due to the overhead of calling the Prefect API for each task run, and the order of execution is the same if the @task
decorator is or isn't there. I think it looks concurrent because the Python-only version is so much faster. I'm going to close this, but feel free to reopen or respond if there's something I missed!
Reopening because we've identified that the new engine creates task runs sequentially in async flows. See this example for reproduction:
import asyncio
from prefect import flow, task, get_run_logger
@task
async def waiter(n: int = 60):
get_run_logger().info(f"Waiting for {n} seconds")
await asyncio.sleep(n)
@flow
async def main(n: int = 100):
waiters = [waiter() for _ in range(n)]
await asyncio.gather(*waiters)
if __name__ == "__main__":
asyncio.run(main())
This is caused by the new engine's use of SyncPrefectClient
to send create task run API calls. The resolution of this issue will involve making those calls non-blocking in async contexts.
First check
Bug summary
In the example below, with
@task
commented out,python fibonacci.py 100
completes in a second or two. With@task
uncommented it takes a minute or two, and appears to run sequentially. I would expect it to run concurrently.Reproduction
Error
Versions
Additional context
No response