Details
-
Bug
-
Status: Closed
-
Minor
-
Resolution: Fixed
-
0.98.2
-
None
-
Reviewed
Description
AgeOfLastAppliedOp in MetricsSink.java is to indicate the time an edit sat in the 'replication queue' before it got replicated(aka applied)
/** * Set the age of the last applied operation * * @param timestamp The timestamp of the last operation applied. * @return the age that was set */ public long setAgeOfLastAppliedOp(long timestamp) { lastTimestampForAge = timestamp; long age = System.currentTimeMillis() - lastTimestampForAge; rms.setGauge(SINK_AGE_OF_LAST_APPLIED_OP, age); return age; }
In the following scenario:
1) at 7:00am a sink op is applied, and the SINK_AGE_OF_LAST_APPLIED_OP is
set for example 100ms;
2) and then NO new Sink op occur.
3) when a refreshAgeOfLastAppliedOp() is invoked at 8:00am. Instead of
return the 100ms, the AgeOfLastAppliedOp become 1hour + 100ms,
It was because that refreshAgeOfLastAppliedOp() get invoked periodically by getStats().
proposed fix:
--- hbase-server/src/main/java/org/apache/hadoop/hbase/replication/regionserver/MetricsSink.java +++ hbase-server/src/main/java/org/apache/hadoop/hbase/replication/regionserver/MetricsSink.java @@ -35,6 +35,7 @@ public class MetricsSink { private MetricsReplicationSource rms; private long lastTimestampForAge = System.currentTimeMillis(); + private long age = 0; public MetricsSink() { rms = CompatibilitySingletonFactory.getInstance(MetricsReplicationSource.class); @@ -47,8 +48,12 @@ public class MetricsSink { * @return the age that was set */ public long setAgeOfLastAppliedOp(long timestamp) { - lastTimestampForAge = timestamp; - long age = System.currentTimeMillis() - lastTimestampForAge; + if (lastTimestampForAge != timestamp) { + lastTimestampForAge = timestamp; + this.age = System.currentTimeMillis() - lastTimestampForAge; + } else { + this.age = 0; + } rms.setGauge(SINK_AGE_OF_LAST_APPLIED_OP, age); return age; }
detail discussion in dev@hbase
Attachments
Attachments
Issue Links
- is related to
-
-
- Closed
-