Scaling HBase (nosql store) to handle massive loads at Pinterest by Jeremy Carol
2 likes2,172 views
The document discusses HBase operations on EC2, focusing on deployment strategies, schema design, and production support. A bug related to hostname resolution in the EC2 environment caused issues with region server coordination, which was resolved by enabling reverse DNS. The document concludes with a note on improving reverse DNS functionality within the Hadoop setup.
Scaling HBase (nosql store) to handle massive loads at Pinterest by Jeremy Carol
- 1. HBase Operations on EC2 Jeremy Carroll Big Data Gurus Pinterest Engineering
- 2. Overview • Deployment Strategies for EC2 • Validating Design • Production Support Pinterest Engineering
- 3. Powered by HBase Pinterest Engineering
- 4. Lets Deploy • First Question Asked • Rack Locality? • Cloud Concepts Pinterest Engineering
- 9. Does This Work? • Schema Design • Hot Spots • Load Testing • Tools Pinterest Engineering
- 10. Does This Work? Pinterest Engineering
- 13. Production • Monitoring • Alerting • Health Pinterest Engineering
- 20. Snapshots & DNS HBASE-8473 17:10 <jeremy_carroll> jmhsieh: I think I found the root cuase. All my region servers reach the barrier, but it does not continue. 17:11 <jeremy_carroll> jmhsieh: All RS have this in their logs: DEBUG org.apache.hadoop.hbase.procedure.Subprocedure: Subprocedure 'backup1' coordinator notified of 'acquire', waiting on 'reached' or 'abort' from coordinator. 17:11 <jeremy_carroll> jmhsieh: Then the coordinator (Master) never sends anything. They just sit until the timeout. 17:12 <jeremy_carroll> jmhsieh: So basically 'reached' is never obtained. Then abort it set, and it fails. ... 17:24 <jeremy_carroll> jmhsieh: Found the bug. The hostnames dont match the master due to DNS resolution 17:25 <jeremy_carroll> jmhsieh: The barrier aquired is putting in the local hostname from the regionservers. In EC2 (Where reverse DNS does not work well), the master hands the internal name to the client. 17:26 <jeremy_carroll> jmhsieh: So it's waiting for something like 'ip-10-155-208-202.ec2.internal, 60020,1367366580066' zNode to show up, but instead 'hbasemetaclustera-d1b0a484,60020,1367366580066,' is being inserted. Barrier is not reached 17:27 <jeremy_carroll> jmhsieh: Reason being in our environment the master does not have a reverse DNS entry. So we get stuff like this on RegionServer startup in our logs. 17:27 <jeremy_carroll> jmhsieh: 2013-05-01 00:03:00,614 INFO org.apache.hadoop.hbase.regionserver.HRegionServer: Master passed us hostname to use. Was=hbasemetaclustera-d1b0a484, Now=ip-10-155-208-202.ec2.internal 17:54 <jeremy_carroll> jmhsieh: That was it. Verified. Now that Reverse DNS is working, snapshots are working. Now how to figure out how to get Reverse DNS working on Route53. I wished there was something like 'slave.host.name' inside of Hadoop for this. Looking at source code. Pinterest Engineering