Skip to:

  1. Skip to Jira Navigation
  2. Skip to Side Navigation
  3. Skip to Main Content

Node inconsistency seen in PS during update on compressed column table

Description

While testing PXC-5.7 and PXC-8.0 on a 3 node cluster setup, it is observed that if a certain combinations of DMLs (INSERT,UPDATE,DELETE,REPLACE) are executed on a prepared metadata containing both partition and non-partition tables with frequent COMMIT and ROLLBACK, the nodes gets into an inconsistent state leading to:

shut down of node 2 and node 3 with node1 being alive (on 5.7)
shutdown of node 1 and node 2(primary) and node 3 being alive (on 8.0)

Attaching the conf file for starting the nodes and the SQL file which can be run from primary node to repeat the problem.

Update: The problem exists in PS as well. Please refer to comments section

Environment

None

Attachments

4

Smart Checklist

Activity

Show:

Venkatesh Prasad March 11, 2021 at 1:26 PM

Without partition, for UPDATE query, the server logs the uncompressed value in binary log

With partition, it looks like it logs compressed values to the binlog. 

Venkatesh Prasad March 11, 2021 at 11:01 AM
Edited

Few more observations: This bug happens only when 

  1. COLUMN_FORMAT COMPRESSED is specified in the query.

  2. Data in the BLOB column is at least of length 96 characters.

  3. The table is partitioned

Venkatesh Prasad March 11, 2021 at 10:33 AM

This issue is not specific to PXC as it is present in normal async replication in PS also.

Here is the simplified test case.

I think it is better if this bug is moved under PS project.

Duplicate

Details

Assignee

Reporter

Time tracking

1d 3h 45m logged

Priority

Smart Checklist

Created March 11, 2021 at 7:00 AM
Updated March 6, 2024 at 10:40 AM
Resolved May 4, 2021 at 6:20 PM