[CASSANDRA-13899] Fix buffer length comparison when decompressing in netty-based streaming

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 4.0-alpha1, 4.0
• Severity: Normal
• Assignee: Jason Brown
• Reporter: Paulo Motta

Authors

Jason Brown

Reviewers

Paulo Motta

• Created: 25/Sep/17 10:16
• Updated: 15/May/20 08:00
• Resolved: 04/Oct/17 07:05

Description

Streaming a single partition with ~100K rows fails with the following exception:

ERROR [Stream-Deserializer-/127.0.0.1:35149-a92e5e12] 2017-09-21 04:03:41,237 StreamSession.java:617 - [Stream #c2e5b640-9eab-11e7-99c0-e9864ca8da8e] Streaming error occurred on session with peer 127.0.0.1
org.apache.cassandra.streaming.StreamReceiveException: java.lang.RuntimeException: Last written key DecoratedKey(-1000328290821038380) >= current key DecoratedKey(-1055007227842125139)  writing into /home/paulo/.ccm/test/node2/data0/stresscql/typestest-482ac7b09e8d11e787cf85d073c
8e037/na-1-big-Data.db
        at org.apache.cassandra.streaming.messages.IncomingFileMessage$1.deserialize(IncomingFileMessage.java:63) ~[main/:na]
        at org.apache.cassandra.streaming.messages.IncomingFileMessage$1.deserialize(IncomingFileMessage.java:41) ~[main/:na]
        at org.apache.cassandra.streaming.messages.StreamMessage.deserialize(StreamMessage.java:55) ~[main/:na]
        at org.apache.cassandra.streaming.async.StreamingInboundHandler$StreamDeserializingTask.run(StreamingInboundHandler.java:178) ~[main/:na]
        at java.lang.Thread.run(Thread.java:745) [na:1.8.0_121]

Reproduction steps:

• Create CCM cluster with 2 nodes
• Start only first node, disable hinted handoff
• Run stress with the attached yaml: tools/bin/cassandra-stress "user profile=largepartition.yaml n=10K ops(insert=1) no-warmup -node whitelist 127.0.0.1 -mode native cql3 compression=lz4 -rate threads=4 -insert visits=FIXED(100K) revisit=FIXED(100K)"
• Start second node, run repair on stresscql table - the exception above will be thrown.

I investigated briefly and haven't found anything suspicious. This seems to be related to CASSANDRA-12229 as I tested the steps above in a branch without that and the repair completed successfully. I haven't tested with a smaller number of rows per partition to see at which point it starts to be a problem.

We should probably add a regression dtest to stream large partitions to catch similar problems in the future.

Issue Links

• is related to: CASSANDRA-10520 Compressed writer and reader should support non-compressed data.