[
https://issues.apache.org/jira/browse/MESOS-8256?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Benjamin Mahler updated MESOS-8256:
-----------------------------------
Description:
Currently, libprocess uses a fixed number of worker threads. This means that
any code that blocks a worker thread and requires another worker thread to
unblock it can lead to deadlock if there are sufficiently many of these to
block all of the worker threads. The deadlock will occur without any logging of
it and we don't expose an endpoint for it either.
Our current approach to avoid this issue is to (1) forbid blocking a worker
thread, however there is a lot of blocking code using {{process::wait}} (the
alternative is to spawn a managed process) and other code still performs other
blocking (such as {{ZooKeeper}} or custom module code, this code could be fixed
to be non-blocking), and (2) set the worker thread pool minimum size to a known
safe value. (2) is brittle and we cannot determine the minimum safe number
easily as the code evolves, and as users run module code.
Ideally:
(1) We can indicate that the deadlock occurs via a log message and also
possibly an endpoint, or even crashing! Ideally, the user can see all of the
stack traces to know why the deadlock occurred.
(2) Libprocess could keep a dynamically sized worker pool. At the very least,
we could detect deadlock and spawn additional threads to get out of it,
removing these threads later. Perhaps a simpler approach is just to spawn
another temporary worker any time that a particular worker is going to block
(or inversely, decommission the worker that's about to block and have the newly
spawned worker replace it).
was:
Currently, libprocess uses a fixed number of worker threads. This means that
any code that blocks a worker thread and requires another worker thread to
unblock it can lead to deadlock if there are sufficiently many of these to
block all of the worker threads. The deadlock will occur without any logging of
it and we don't expose an endpoint for it either.
Our current approach to avoid this issue is to (1) forbid blocking a worker
thread, however there is a lot of blocking code using {{process::wait}} (the
alternative is to spawn a managed process) and other code still performs other
blocking (such as {{ZooKeeper}} or custom module code, this code could be fixed
to be non-blocking), and (2) set the worker thread pool minimum size to a known
safe value. (2) is brittle and we cannot determine the minimum safe number
easily as the code evolves, and as users run module code.
Ideally:
(1) We can indicate that the deadlock occurs via a log message and also
possibly an endpoint, or even crashing! Ideally, the user can see all of the
stack traces to know why the deadlock occurred.
(2) Libprocess could keep a dynamically sized worker pool. At the very least,
we could detect deadlock and spawn additional threads to get out of it,
removing these threads later. Perhaps a simpler approach is just to spawn
another worker any time that a particular worker is going to block.
> Libprocess can silently deadlock due to worker thread exhaustion.
> -----------------------------------------------------------------
>
> Key: MESOS-8256
> URL: https://issues.apache.org/jira/browse/MESOS-8256
> Project: Mesos
> Issue Type: Bug
> Components: libprocess
> Reporter: Benjamin Mahler
> Priority: Critical
>
> Currently, libprocess uses a fixed number of worker threads. This means that
> any code that blocks a worker thread and requires another worker thread to
> unblock it can lead to deadlock if there are sufficiently many of these to
> block all of the worker threads. The deadlock will occur without any logging
> of it and we don't expose an endpoint for it either.
> Our current approach to avoid this issue is to (1) forbid blocking a worker
> thread, however there is a lot of blocking code using {{process::wait}} (the
> alternative is to spawn a managed process) and other code still performs
> other blocking (such as {{ZooKeeper}} or custom module code, this code could
> be fixed to be non-blocking), and (2) set the worker thread pool minimum size
> to a known safe value. (2) is brittle and we cannot determine the minimum
> safe number easily as the code evolves, and as users run module code.
> Ideally:
> (1) We can indicate that the deadlock occurs via a log message and also
> possibly an endpoint, or even crashing! Ideally, the user can see all of the
> stack traces to know why the deadlock occurred.
> (2) Libprocess could keep a dynamically sized worker pool. At the very least,
> we could detect deadlock and spawn additional threads to get out of it,
> removing these threads later. Perhaps a simpler approach is just to spawn
> another temporary worker any time that a particular worker is going to block
> (or inversely, decommission the worker that's about to block and have the
> newly spawned worker replace it).
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
