Discovered and heavily worked around in #165 as well as well documented in #113.
The test_debugger.py suite currently is marked up with a @has_nested_actors for the test scripts which do include nested proc trees and these tests are being skipped in CI (despite seemingly nearly working when running local on linux), as well as the test_multi_nested_subactors_error_through_nurseries() test which was never re-written to include ctrl-c cases since we know it gets totally rekt by the zombie reaping machinery.
The detailed lingering TODO from #165:
[ ] we still have an unhandled case that can cause lockups when there is an intermediate child actor (a child of the root which also has children) that has a child in debug mode:
the intermediate can receive SIGINT but does not have the shield_sigint() handler set because it, itself is not in debug mode and thus accepts the signal and cancels its nurseries (and thus children who maybe are in debug mode)
the children of the intermediary are cancelled while still using the debugger and so aren't able to actually cancel since control is given to the user and won't resume until c or q cmds are sent to the repl
the intermediary that's also a parent's nursery will engage the hard kill machinery which will then blow out the IPC connections thus making it impossible for the children to unlock the TTY / debugger lock
the root get's stuck never being able to acquire the debug lock and also still has the shield_sigint() handler enabled making it impossible to raise a KBI and let the hard proc kill (zombie lord) teardown complete.
https://github.com/goodboy/tractor/issues/145, which is currently related the main unexpected outcome: proc/zombie reaping machinery destroying a leaf-actor who's parent doesn't know it's in debug (and thus neither does the root actor).
Discovered and heavily worked around in #165 as well as well documented in #113.
The
test_debugger.pysuite currently is marked up with a@has_nested_actorsfor the test scripts which do include nested proc trees and these tests are being skipped in CI (despite seemingly nearly working when running local on linux), as well as thetest_multi_nested_subactors_error_through_nurseries()test which was never re-written to includectrl-ccases since we know it gets totally rekt by the zombie reaping machinery.The detailed lingering TODO from #165:
SIGINTbut does not have theshield_sigint()handler set because it, itself is not in debug mode and thus accepts the signal and cancels its nurseries (and thus children who maybe are in debug mode)corqcmds are sent to the replshield_sigint()handler enabled making it impossible to raise a KBI and let the hard proc kill (zombie lord) teardown complete.Other related issues: