[jira] [Updated] (HBASE-8747) Log when blocked longer than expected on a lock/synchronized block

["Andrew Purtell (JIRA)" <ji...@apache.org>]

     [ https://issues.apache.org/jira/browse/HBASE-8747?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Andrew Purtell updated HBASE-8747:
----------------------------------

    Affects Version/s: 0.94.9

This is great. I volunteer to port back into 0.94. Please let me know how/if I can help in addition.
                
> Log when blocked longer than expected on a lock/synchronized block
> ------------------------------------------------------------------
>
>                 Key: HBASE-8747
>                 URL: https://issues.apache.org/jira/browse/HBASE-8747
>             Project: HBase
>          Issue Type: Improvement
>          Components: Usability
>    Affects Versions: 0.95.2, 0.94.9
>            Reporter: stack
>            Assignee: stack
>            Priority: Critical
>
> Recently, on an internal support thread, the support folks asked that HBase surface when it is blocked up internally or an internal resource is highly contended.  I replied in essence that this would be 'too hard'.  Our Todd instead suggested:
> {quote}
> ...straw man proposal:
> - swap out all our use of sychronized (...) for ReentrantLock (probably most of our locking is already based on j.u.c locks anyway for rwlock purposes)
> - build a wrapper around ReentrantLock that looks something like:
> class AwesomeLock implements Lock {
>   AwesomeLock(Lock realLock, int expectedAcquisitionTime, String metricName) {
>     this.realLock = realLock;
>   }
>   public void lock() {
>     if (realLock.tryLock(expectedAcquisitionTime, MILLIS)) { return; }
>     Stopwatch sw = new Stopwatch();
>     // collect realLock.getOwner()
>     realLock.lock();
>     long elapsed = sw.elapsed() + expectedAcquisitionTime;
>     if (elapsed > expected * 2) { LOG.warn("Waited too long for lock which was held by owner " + owner); }
>     collect metrics about how long the lock took, the lock name etc
>   }
> }
> Basically the idea is that any possible point of contention should emit log messages and metrics when contended.
> {quote}
> Let me do the above over the next week or so.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira