[
https://issues.apache.org/jira/browse/HADOOP-18217?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17526693#comment-17526693
]
Catherinot Remi commented on HADOOP-18217:
------------------------------------------
added wtf.java to show what i have in mind before proposing a patch (patch
would also slightly rework try/catch/finally blocks to ensure the executor is
shutdown). If puts in place 2 differents safety nets : the delayed Runtime.halt
one + an uncaught throwable one. From the command line you can run 4 scenarios
: normal ending, System.exit ending, uncaugh Throwable ending, bad hook ending.
Still i would rather remove the executor stuff since being monothread only
makes timeout visible and logged, but does not makes all hooks to run. In
wtf.java for exemple, the 1st hook runs, timeout and cancel is called, but even
with that, calling cancel on the 2nd future just keep interrupting the 1st
(because future.cancel interrupt the only running thread, currently running
f1's code).
The uncaught handler safety net in the code is one I use to exit from JVM that
get OOM errors. Error not being Exception are often not catched/handled (to
trigger cleanup code and rethrow them) and often break asynchronous jobs (the
oom raised often make the synchronized/notify block of code to be skipped,
never unlocking the caller stuck inside the synchronized/wait block
counter-part). I can be dangerous though when used with fraemworks that just
spawn threads and let them die by crashing. Hopefully this is not a very common
pattern :)
For the patch, to minimize the changes and to be as safe as possible, I would
only keep the delayed halt safety net + slight try/catch/finally rework around
the exectutor shutdown calls.
The uncaught throwable safety net is just a proposal.
Removing the executor is just a proposal too, or change the future.cancel call
to keep calling it on a future until it is effectively cancelled (calling
cancel once is not enough has wtf.java demonstrate)
> shutdownhookmanager should not be multithreaded (deadlock possible)
> -------------------------------------------------------------------
>
> Key: HADOOP-18217
> URL: https://issues.apache.org/jira/browse/HADOOP-18217
> Project: Hadoop Common
> Issue Type: Bug
> Components: util
> Affects Versions: 2.10.1
> Environment: linux, windows, any version
> Reporter: Catherinot Remi
> Priority: Minor
> Attachments: wtf.java
>
>
> the ShutdownHookManager class uses an executor to run hooks to have a
> "timeout" notion around them. It does this using a single threaded executor.
> It can leads to deadlock leaving a never-shutting-down JVM with this
> execution flow:
> * JVM need to exit (only daemon threads remaining or someone called
> System.exit)
> * ShutdowHookManager kicks in
> * SHMngr executor start running some hooks
> * SHMngr executor thread kicks in and, as a side effect, run some code from
> one of the hook that calls System.exit (as a side effect from an external lib
> for example)
> * the executor thread is waiting for a lock because another thread already
> entered System.exit and has its internal lock, so the executor never returns.
> * SHMngr never returns
> * 1st call to System.exit never returns
> * JVM stuck
>
> using an executor with a single thread does "fake" timeouts (the task keeps
> running, you can interrupt it but until it stumble upon some piece of code
> that is interruptible (like an IO) it will keep running) especially since the
> executor is a single threaded one. So it has this bug for example :
> * caller submit 1st hook (bad one that would need 1 hour of runtime and that
> cannot be interrupted)
> * executor start 1st hook
> * caller of the future 1st hook result timeout
> * caller submit 2nd hook
> * bug : 1 hook still running, 2nd hook triggers a timeout but never got the
> chance to run anyway, so 1st faulty hook makes it impossible for any other
> hook to have a chance to run, so running hooks in a single separate thread
> does not allow to run other hooks in parallel to long ones.
>
> If we really really want to timeout the JVM shutdown, even accepting maybe
> dirty shutdown, it should rather handle the hooks inside the initial thread
> (not spawning new one(s) so not triggering the deadlock described on the 1st
> place) and if a timeout was configured, only spawn a single parallel daemon
> thread that sleeps the timeout delay, and then use Runtime.halt (which bypass
> the hook system so should not trigger the deadlock). If the normal
> System.exit ends before the timeout delay everything is fine. If the
> System.exit took to much time, the JVM is killed and so the reason why this
> multithreaded shutdown hook implementation was created is satisfied (avoding
> having hanging JVMs)
>
> Had the bug with both oracle and open jdk builds, all in 1.8 major version.
> hadoop 2.6 and 2.7 did not have the issue because they do not run hooks in
> another thread
>
> Another solution is of course to configure the timeout AND to have as many
> threads as needed to run the hooks so to have at least some gain to offset
> the pain of the dealock scenario
>
> EDIT: added some logs and reproduced the problem. in fact it is located after
> triggering all the hook entries and before shutting down the executor.
> Current code, after running the hooks, creates a new Configuration object and
> reads the configured timeout from it, applies this timeout to shutdown the
> executor. I sometimes run with a classloader doing remote classloading,
> Configuration loads its content using this classloader, so when shutting down
> the JVM and some network error occurs the classloader fails to load the
> ressources needed by Configuration. So the code crash before shutting down
> the executor and ends up inside the thread's default uncaught throwable
> handler, which was calling System.exit, so got stuck, so shutting down the
> executor never returned, so does the JVM.
> So, forget about the halt stuff (even if it is a last ressort very robust
> safety net). Still I'll do a small adjustement to the final executor shutdown
> code to be slightly more robust to even the strangest exceptions/errors it
> encounters.
--
This message was sent by Atlassian Jira
(v8.20.7#820007)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
