[jira] [Created] (HBASE-25677) Server+table counters on each scan #nextRaw invocation becomes a bottleneck when heavy load

["Michael Stack (Jira)" <ji...@apache.org>]

Michael Stack created HBASE-25677:
-------------------------------------

             Summary: Server+table counters on each scan #nextRaw invocation becomes a bottleneck when heavy load
                 Key: HBASE-25677
                 URL: https://issues.apache.org/jira/browse/HBASE-25677
             Project: HBase
          Issue Type: Sub-task
          Components: metrics
    Affects Versions: 2.3.2
            Reporter: Michael Stack
            Assignee: Michael Stack


On a heavily loaded server mostly doing reads/scan, I saw that 90+% of handlers were BLOCKED in this fashion in thread dumps:

{code}
"RpcServer.default.FPBQ.Fifo.handler=117,queue=17,port=16020" #161 daemon prio=5 os_prio=0 tid=0x00007f748757f000 nid=0x73e9 waiting for monitor entry [0x00007f74783e0000]
  java.lang.Thread.State: BLOCKED (on object monitor)
       at java.util.concurrent.ConcurrentHashMap.computeIfAbsent(ConcurrentHashMap.java:1674)
       - waiting to lock <0x00007f7647e3cc38> (a java.util.concurrent.ConcurrentHashMap$Node)
       at org.apache.hadoop.hbase.regionserver.MetricsTableQueryMeterImpl.getOrCreateTableMeter(MetricsTableQueryMeterImpl.java:80)
       at org.apache.hadoop.hbase.regionserver.MetricsTableQueryMeterImpl.updateTableReadQueryMeter(MetricsTableQueryMeterImpl.java:90)
       at org.apache.hadoop.hbase.regionserver.RegionServerTableMetrics.updateTableReadQueryMeter(RegionServerTableMetrics.java:89)
       at org.apache.hadoop.hbase.regionserver.MetricsRegionServer.updateReadQueryMeter(MetricsRegionServer.java:274)
       at org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl.nextRaw(HRegion.java:6742)
       at org.apache.hadoop.hbase.regionserver.RSRpcServices.scan(RSRpcServices.java:3319)
       - locked <0x00007f896c0165a0> (a org.apache.hadoop.hbase.regionserver.HRegion$RegionScannerImpl)
       at org.apache.hadoop.hbase.regionserver.RSRpcServices.scan(RSRpcServices.java:3566)
       at org.apache.hadoop.hbase.shaded.protobuf.generated.ClientProtos$ClientService$2.callBlockingMethod(ClientProtos.java:44858)
       at org.apache.hadoop.hbase.ipc.RpcServer.call(RpcServer.java:393)
       at org.apache.hadoop.hbase.ipc.CallRunner.run(CallRunner.java:133)
       at org.apache.hadoop.hbase.ipc.RpcExecutor$Handler.run(RpcExecutor.java:338)
       at org.apache.hadoop.hbase.ipc.RpcExecutor$Handler.run(RpcExecutor.java:318)

{code}

It kept up for good periods of time.

I saw it to a leser extent on other servers, with less load.

These RS had 400+ Regions a good few of which were serving out scan reads; the server was doing ~1M hits a second. In this scenario, I saw the above bottleneck.

Looking at it, it came in w/ when the parent issue feature was added. There are these read counts and then there were also write counts. The write counts are mostly batch-based. Let me do same thing here for the read.... update the central server+table count after scan is done rather than per invocation of #nextRaw.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)