[ https://issues.apache.org/jira/browse/HBASE-11368?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14181168#comment-14181168 ]
Qiang Tian commented on HBASE-11368: ------------------------------------ initial YCSB test: Env: --- hadoop 2.2.0 YCSB 1.0.4(Andrew's branch) 3 nodes, 1 master, 2 RS //ignore cluster details since just to evaluate the new lock Steps: --- Followed Andrew's steps(see http://search-hadoop.com/m/DHED4hl7pC/) the seed table has 3 CFs, pre-split to 20 regions load 1 million rows to CF 'f1', using workloada run 3 iterations for workloadc and workloada respectively. the parameter in each run: bq. -p columnfamily=f1 -p operationcount=1000000 -s -threads 10 Results: --- 0.98.5: workload c: [READ], AverageLatency(us), 496.225811 [READ], AverageLatency(us), 510.206831 [READ], AverageLatency(us), 501.256123 workload a: [READ], AverageLatency(us), 676.4527555821747 [READ], AverageLatency(us), 622.5544771452717 [READ], AverageLatency(us), 628.1365657163067 0.98.5+patch: workload c: [READ], AverageLatency(us), 536.334437 [READ], AverageLatency(us), 508.405555 [READ], AverageLatency(us), 491.416182 workload a: [READ], AverageLatency(us), 640.3625218319231 [READ], AverageLatency(us), 642.9719823488798 [READ], AverageLatency(us), 631.7491770928287 looks little performance penalty. I also ran PE in the cluster, since the test table has only 1 CF, the new lock is actually not used. interestingly, with the patch the performance is even a bit better... > Multi-column family BulkLoad fails if compactions go on too long > ---------------------------------------------------------------- > > Key: HBASE-11368 > URL: https://issues.apache.org/jira/browse/HBASE-11368 > Project: HBase > Issue Type: Bug > Reporter: stack > Assignee: Qiang Tian > Attachments: hbase-11368-0.98.5.patch > > > Compactions take a read lock. If a multi-column family region, before bulk > loading, we want to take a write lock on the region. If the compaction takes > too long, the bulk load fails. > Various recipes include: > + Making smaller regions (lame) > + [~victorunique] suggests major compacting just before bulk loading over in > HBASE-10882 as a work around. > Does the compaction need a read lock for that long? Does the bulk load need > a full write lock when multiple column families? Can we fail more gracefully > at least? -- This message was sent by Atlassian JIRA (v6.3.4#6332)