There have been several releases for Hbase 1.2 Which release are you using ?
The images you sent didn't go through the mailing list. Please consider using third party site for delivery. Have you taken a look at what the server hosting hbase:meta was doing during this period of time ? Thanks On Tue, Aug 7, 2018 at 8:04 AM Mike Freyberger <mfreyber...@appnexus.com> wrote: > Kafka Dev, > > > > I’d love some help investigating a slow Hbase mutator. > > > > The cluster is Hbase 1.2 and cluster has 22 region servers. The region > servers are pretty big: 24 cores, 126 GB RAM. > > > > The cluster has 2 tables, each only have 1 column family. Both tables have > the same pre splits. > > > > Each table is pre split into 400 regions. The split keys are all 2 bytes > and evenly divide the key space. > > > > The keys are 13 bytes. The key is formed by concatenating: > > 1 byte kafka partition > > 8 byte random int > > 4 byte timestamp (second level granularity) > > The workload is 100% write for now. There are about 1M writes per second > with a total data volume of .6GB per second. > > > > I find that my application is spending the majority of its CPU time > (71.7%) calling org.apache.hadoop.hbase.client.BufferedMutatorImpl.mutate > (), which is in turn spending most of its time calling > org.apache.hadoop.hbase.client.ConnectionManager$HConnectionImplementation.locateRegion > (). > > > > Attached are two images showing the performance of my application. The > first is an overview showing that my application is spending a lot of in > mutate. The next is a deep dive into the functions that mutate is calling > internally. > > I am very surprised to see this function taking so long. My intuition is > that all this needs to do is: > 1) Determine which region the Mutation belongs in > 2) Append the Mutation to a queue for async write to HBase. > > > Any thoughts, comments of suggestions from the community would be much > appreciated! I’m really hoping to improve the performance profile here so > that my CPU can be freed up. > > > > Thanks, > > > > Mike Freyberger >