Tony Wu created HADOOP-12482:
--------------------------------

             Summary: Race condition in JMX cache update
                 Key: HADOOP-12482
                 URL: https://issues.apache.org/jira/browse/HADOOP-12482
             Project: Hadoop Common
          Issue Type: Bug
    Affects Versions: 2.7.1
            Reporter: Tony Wu
            Assignee: Tony Wu


updateJmxCache() was in HADOOP-11301. However the patch introduced a race 
condition. In updateJmxCache() function in MetricsSourceAdapter.java:

{code:java}
  private void updateJmxCache() {
    boolean getAllMetrics = false;
    synchronized (this) {
      if (Time.now() - jmxCacheTS >= jmxCacheTTL) {
        // temporarilly advance the expiry while updating the cache
        jmxCacheTS = Time.now() + jmxCacheTTL;
        if (lastRecs == null) {
          getAllMetrics = true;
        }
      } else {
        return;
      }

      if (getAllMetrics) {
        MetricsCollectorImpl builder = new MetricsCollectorImpl();
        getMetrics(builder, true);
      }

      updateAttrCache();
      if (getAllMetrics) {
        updateInfoCache();
      }
      jmxCacheTS = Time.now();
      lastRecs = null; // in case regular interval update is not running
    }
  }
{code}

Notice that getAllMetrics is set to true when:
# jmxCacheTTL has passed
# lastRecs == null

lastRecs is set to null in the same function, but gets reassigned by 
getMetrics().

However getMetrics() can be called from a different thread:
# MetricsSystemImpl.onTimerEvent()
# MetricsSystemImpl.publishMetricsNow()

Consider the following sequence:
# updateJmxCache() is called by getMBeanInfo() from a thread getting cached 
info. 
** lastRecs is set to null.
# metrics sources is updated with new value/field.
# getMetrics() is called by publishMetricsNow() or onTimerEvent() from a 
different thread getting the latest metrics. 
** lastRecs is updated (!= null).
# jmxCacheTTL passed.
# updateJmxCache() is called again via getMBeanInfo().
** However because lastRecs is already updated (!= null), getAllMetrics will 
not be set to true. So updateInfoCache() is not called and getMBeanInfo() 
returns the old cached info.

We ran into this issue on a cluster where a new metric did not get published 
until much later.

The case can be made worse by a periodic call to getMetrics() (driven by an 
external program or script). In such case getMBeanInfo() may never be able to 
retrieve the new record.

The desired behavior should be that updateJmxCache() will guarantee to call 
updateInfoCache() once after jmxCacheTTL, if lastRecs has been set to null by 
updateJmxCache() itself.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to