Would it be possible to SIGINT/SIGTERM processes instead of SIGKILLING them?
nikhilm opened this issue · comments
It seems like nodes are sent a sigkill right away at termination (via Process.destroyForcefully
. This makes it a little tricky to use things like in-process sampling profilers because the process is killed before the profile can be written.
It would be nicer if the node had its stdin closed/have a sigint and then was given a few seconds to quit before being forcibly killed.
It is pretty easy for the maelstrom implementation in a language to watch for stdin closure and break out of the Run()
.
Sure, you could patch this in! I should warn that like... at some point I intend to use Maelstrom for testing process crashes, not just network partitions, and when that happens it's definitely gonna be kill -9
--losing data is kind of the point! ;-)
Sorry for not responding to this. I was eventually able to skip using a profiler entirely for my specific problem, so I did not continue with this.
I agree that workloads killing processes would be cool!