[SOLR-13248] Autoscaling based replica placement is broken out of the box

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Fixed
• Affects Version/s: 7.6, 7.7
• Fix Version/s: 7.7.1, 8.0, 9.0
• Component/s: AutoScaling
• Assignee: Shalin Shekhar Mangar
• Reporter: Shalin Shekhar Mangar
• Created: 14/Feb/19 12:36
• Updated: 08/Jun/19 15:00
• Resolved: 07/Mar/19 18:51

Description

SOLR-12739 made autoscaling as the default replica placement strategy. However in the absence of SOLR-12845, replicas can be placed without any regards for maxShardsPerNode causing multiple replicas of the same shard to be placed on the same node together. Also it was reported in SOLR-13247 that createNodeSet is not being respected as well.

SOLR-13159 was an early signal of the problem but it was not reproducible and there was a DNS problem in the cluster too so the root cause was not clear then.

I am creating this blocker issue because as it stands today, we cannot guarantee the layout of new collections. At a minimum, we should revert to using the legacy replica assignment policy or add default policies with SOLR-12845 and have createNodeSet work. Related but not mandatory would be to fix SOLR-12847 as well.

Issue Links

• incorporates: SOLR-13247 Incorrect shard placement during Collection creation
• incorporates: SOLR-13159 Autoscaling not distributing collection evenly
• is caused by: SOLR-12739 Make autoscaling policy based replica placement the default strategy for placing replicas
• is related to: SOLR-12944 Bugs around createNodeSet=EMPTY
• is related to: SOLR-12847 Cut over implementation of maxShardsPerNode to a collection policy
• is related to: SOLR-12845 Add a default cluster policy