Bryan Beaudreault created HBASE-26713:
-----------------------------------------
Summary: Increments submitted by 1.x clients will be stored with
timestamp 0 on 2.x+ clusters
Key: HBASE-26713
URL: https://issues.apache.org/jira/browse/HBASE-26713
Project: HBase
Issue Type: Bug
Reporter: Bryan Beaudreault
We discovered this in our work on upgrading from ~1.2.0 to 2.4.6. A 1.2.0
client will submit an Increment, without specifying a timestamp. This should be
translated as LATEST_TIMESTAMP, but an hbase 2 server will store it with a
timestamp of 0.
This is because the QualifierValue timestamp proto field is optional, and is
not set in branch-1 MutationProtos:
[https://github.com/apache/hbase/blob/branch-1/hbase-client/src/main/java/org/apache/hadoop/hbase/protobuf/RequestConverter.java#L218-L219.]
In branch-1, the timestamp is actually ignored entirely:
[https://github.com/apache/hbase/blob/branch-1/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java#L8435-L8456.]
So this is not an issue there.
As of HBASE-18546, the timestamp is no longer ignored, but the usage of
QualiferValue.getTimestamp() in deserialization is not wrapped in a
.hasTimestamp() call:
[https://github.com/apache/hbase/blob/7d8dc3524981a04e82ff236bb3dc18d36ebc82b4/hbase-client/src/main/java/org/apache/hadoop/hbase/shaded/protobuf/ProtobufUtil.java#L886-L909].
The default value for non-present optionals is defined by the underlying
primitive, so for longs that is 0. Since the Increment's timestamp is now
respected by HRegion, it causes the erroneous 0 value to be stored:
[https://github.com/apache/hbase/blob/master/hbase-server/src/main/java/org/apache/hadoop/hbase/regionserver/HRegion.java#L4180]
(updateLatestStamp only overwrites the cell timestamp if it's
LATEST_TIMESTAMP, which in this case it has been deserialized as 0)
I think this might only apply to clients who do not use a KeyValueCodec.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
