[HBASE-28595] Losing exception from scan RPC can lead to partial results - ASF JIRA

XML

Word

Printable

JSON

Details

Type: Bug
Status: Resolved
Priority: Critical
Resolution: Fixed
Affects Version/s: None
Fix Version/s: 2.4.18, 2.7.0, 3.0.0-beta-2, 2.6.1, 2.5.9
Component/s: regionserver, Scanners
Labels:
- pull-request-available

Hadoop Flags:

Reviewed

Description

This was discovered in Apache Impala using HBase 2.2 based branch hbase client and server. It is not clear yet whether other branches are also affected.

The issue happens if the server side of the scan throws an exception and closes the scanner, but at the same time, the client gets an rpc connection closed error and doesn't process the exception sent by the server. Client then thinks it got a network error, which leads to retrying the RPC instead of opening a new scanner. But then when the client retry reaches the server, the server returns an empty ScanResponse instead of an error, leading to closing the scanner on client side without returning any error.

A few pointers to critical parts:
region server:
1st call throws exception leading to closing (but not deleting) scanner:
https://github.com/apache/hbase/blob/0c8607a35008b7dca15e9daaec41ec362d159d67/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RSRpcServices.java#L3539
2nd call (retry of 1st) returns empty results:
https://github.com/apache/hbase/blob/0c8607a35008b7dca15e9daaec41ec362d159d67/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/RSRpcServices.java#L3403

client:
some exceptions are handled as non-retriable at RPC level and are only handled through opening a new scanner:
https://github.com/apache/hbase/blob/0c8607a35008b7dca15e9daaec41ec362d159d67/hbase-client/src/main/java/org/apache/hadoop/hbase/client/ScannerCallable.java#L214
https://github.com/apache/hbase/blob/0c8607a35008b7dca15e9daaec41ec362d159d67/hbase-client/src/main/java/org/apache/hadoop/hbase/client/ClientScanner.java#L367

This mechanism in the client only works if it gets the exception from the server. If there are connection issues during the RPC then the client won't really know the state of the server.