In Prefect 3.x, SIGKILL is not how normal in-process task cancellation works. Short version: You get

CancelledError

when the task is being cancelled cooperatively inside the Python process. You get

SIGKILL

only when Prefect is cancelling an external process and that process does not exit after a grace period. So the distinction is mostly: * same Python process →

CancelledError

* separate process / worker infrastructure →

SIGTERM

first, then

SIGKILL

if it refuses to stop ### When Prefect uses

CancelledError

This is the normal case for tasks executed by the task runner in the same process as the flow: * async tasks * sync tasks running in-process * timeout/cancellation handled cooperatively by the engine In these cases, Prefect raises cancellation exceptions in Python rather than killing the process. ### When Prefect uses

SIGKILL

This happens when the flow run is executing in a separate OS process and Prefect is trying to stop that process. Typical sequence:

SIGTERM -> wait for grace period -> SIGKILL

So Prefect will send

SIGKILL

when: * the run is in a subprocess / worker-managed process * Prefect has already asked it to stop * the process does not exit within the grace period

SIGKILL

is a last resort. It means the process ignored or did not respond to graceful shutdown. ### Important nuance Usually Prefect sends SIGTERM first, not SIGKILL immediately. That gives the process a chance to: * handle cancellation * run cleanup * exit cleanly If it still hangs, then Prefect force-kills it with

SIGKILL

. ### Examples of when SIGKILL is more likely If your task/flow process is stuck in something like: * blocking native code * hung network call * deadlock * uninterruptible subprocess behavior then it may never respond cleanly to cancellation, so Prefect eventually kills the whole process. ### Practical rule If you are asking:

“Why did my task not get

CancelledError

, but instead the process died?”

That usually means the task was *not being cancelled in-process*; instead, the worker/infrastructure process running the flow was terminated. ### One more subtlety Prefect generally cancels at the flow run process level in worker/infrastructure execution. So if that process gets SIGKILL, your task code does not get a chance to catch

CancelledError

at all. If you want, I can also give you a matrix by execution mode like: * local flow + ThreadPool task runner * async flow/task * subprocess worker * Docker / Kubernetes worker and show exactly which ones get

CancelledError

vs signal termination.