[FLINK-2188] Reading from big HBase Tables

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Not A Problem
• Fix Version/s: 0.9
• Assignee: Unassigned
• Reporter: Hilmi Yildirim
• Created: 09/Jun/15 09:39
• Updated: 12/Jun/15 09:31
• Resolved: 12/Jun/15 09:31

Description

I detected a bug in the reading from a big Hbase Table.

I used a cluster of 13 machines with 13 processing slots for each machine which results in a total number of processing slots of 169. Further, our cluster uses cdh5.4.1 and the HBase version is 1.0.0-cdh5.4.1. There is a Hbase Table with nearly 100. mio rows. I used Spark and Hive to count the number of rows and both results are identical (nearly 100 mio.).
Then, I used Flink to count the number of rows. For that I added the hbase-client 1.0.0-cdh5.4.1 Java API as dependency in maven and excluded the other hbase-client dependencies. The result of the job is nearly 102 mio. , 2 mio. rows more than the result of Spark and Hive. Moreover, I run the Flink job multiple times and sometimes the result fluctuates by +-5.