[GEODE-10171] AbstractRedisData version being incremented for no-op operations can lead to delta not being applied on secondary

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.15.0
• Fix Version/s: 1.15.0
• Component/s: redis
• Labels: pull-request-available redis-triage
• Assignee: Kristen
• Reporter: Donal Evans
• Created: 25/Mar/22 18:15
• Updated: 22/Jun/22 20:46
• Resolved: 12/Apr/22 16:57

Description

For SADD, which may or may not modify the data in the region depending on whether the member being added is already present in the set, the version in AbstractRedisData is updated regardless of whether a Delta is sent to the secondary. This can lead to the version on the primary wrapping around to be equal to the version on the secondary, which means that if a delta is sent, it will not be applied on the secondary, leading to potential data loss.

Below is a test to show this behaviour:

@Test
public void deltaVersionOnPrimary_shouldNotUpdate_ifNoDeltaSent() {
  String originalMember = "fixedMemberName";
  String key = clusterStartUp.getKeyOnServer("tag", 1);
  
  // Version of primary = 0
  // Version of secondary = 0
  jedis.sadd(key, originalMember);

  // No changes are made to the set, since adding an already existing member doesn't modify it
  // Version of primary wraps around back to -1
  // Version of secondary doesn't change because we don't send a delta
  for (int i = 0; i < 255; ++i) {
    jedis.sadd(key, originalMember);
  }

  String newMember = "aNewMemberName";
  // Version of primary = 0
  // Version of secondary = 0, delta is not applied
  jedis.sadd(key, newMember);

  assertThat(jedis.smembers(key)).containsExactlyInAnyOrder(originalMember, newMember);

  clusterStartUp.crashVM(1); // kill primary server

  assertThat(jedis.smembers(key)).containsExactlyInAnyOrder(originalMember, newMember);
} 

The example here uses SADD, but potentially, any command that can be a no-op and uses a versioned Delta is at risk of hitting this. All commands should be checked to ensure they're not vulnerable to this bug.

Issue Links

• links to: GitHub Pull Request #7496
• links to: GitHub Pull Request #7527