[HBASE-17387] Reduce the overhead of exception report in RegionActionResult for multi()

• Type: Bug
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 1.4.0, 1.3.1, 2.0.0
• Hadoop Flags: Reviewed
• Assignee: Ted Yu
• Reporter: Ted Yu
• Created: 29/Dec/16 00:04
• Updated: 24/Sep/18 15:19
• Resolved: 31/Dec/16 16:54

Description

For RSRpcServices#doNonAtomicRegionMutation() :

    for (ClientProtos.Action action: actions.getActionList()) {
...
      } catch (IOException ie) {
        rpcServer.getMetrics().exception(ie);
        resultOrExceptionBuilder = ResultOrException.newBuilder().
          setException(ResponseConverter.buildException(ie));
      }
      if (resultOrExceptionBuilder != null) {
        // Propagate index.
        resultOrExceptionBuilder.setIndex(action.getIndex());
        builder.addResultOrException(resultOrExceptionBuilder.build());
      }

The exceptions are added to builder in the for loop.
The ClientProtos.ResultOrException.Builder instance is created within the for loop.

For large multi call, this may incur non-trivial overhead for garbage collector if there're many exceptions.
e.g. Here was sample debug log showing the actions in a batch:

2016-12-23 04:21:56,263 DEBUG org.apache.hadoop.hbase.regionserver.RSRpcServices: NonAtomicRegionMutation batch summary: numAppends=0, numDeletes=11, numGets=0, numIncrements=15638, numPuts=15627, numServiceCalls=0, serializedSize=3871713, user=hbase/pob3.G.COM (auth:KERBEROS), client=null

Note the large number of increments in the batch.
When the increments encounter conflict at server side, the overhead of stringified exceptions in response is considerable.

Issue Links

• is related to: HBASE-17408 Introduce per request limit by number of mutations