[CASSANDRA-15459] Short read protection doesn't work on group-by queries

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.11.8, 4.0-beta2, 4.0
• Component/s: Legacy/Coordination
• Labels: correctness
• Bug Category: Correctness - Consistency
• Severity: Normal
• Complexity: Normal
• Discovered By: Code Inspection
• Platform: All
• Since Version: 3.10

Source Control Link

https://github.com/apache/cassandra/commit/86a9261fd94707283e4ce149f88756099e22fcb6

Test and Documentation Plan

New in-JVM dtests added here.

• Assignee: Andres de la Peña
• Reporter: Zhao Yang

Authors

Andres de la Peña

Reviewers

Caleb Rackliffe

• Created: 18/Dec/19 14:13
• Updated: 27/May/21 15:53
• Resolved: 06/Aug/20 12:30

Description

DTest to reproduce: it affects all versions..

In a two-node cluster with RF = 2

Execute only on Node1:
* Insert pk=1 and ck=1 with timestamp 9
* Delete pk=0 and ck=0 with timestamp 10
* Insert pk=2 and ck=2 with timestamp 9

Execute only on Node2:
* Delete pk=1 and ck=1 with timestamp 10
* Insert pk=0 and ck=0 with timestamp 9
* Delete pk=2 and ck=2 with timestamp 10

Query: "SELECT pk, c FROM %s GROUP BY pk LIMIT 1"
* Expect no live data, but got [0, 0]

Note: for group-by queries, SRP should use "group counted" to calculate limits used for SRP query, rather than "row counted".

Issue Links

• relates to: CASSANDRA-16307 GROUP BY queries with paging can return deleted data
• links to: 3.11 PR
• links to: dtest PR
• links to: trunk PR