forked from goodboy/tractor
1
0
Fork 0
tractor/newsfragments/245.feature.rst

14 lines
776 B
ReStructuredText

Change the core message loop to handle task and actor-runtime cancel
requests immediately instead of scheduling them as is done for rpc-task
requests.
In order to obtain more reliable teardown mechanics for (complex) actor
trees it's important that we specially treat cancel requests as having
higher priority. Previously, it was possible that task cancel requests
could actually also themselves be cancelled if a "actor-runtime" cancel
request was received (can happen during messy multi actor crashes that
propagate). Instead cancels now block the msg loop until serviced and
a response is relayed back to the requester. This also allows for
improved debugger support since we have determinism guarantees about
which processes must wait before hard killing their children.