Details
-
Bug
-
Resolution: Unresolved
-
Major
-
7.6.0
-
None
-
debian
-
Untriaged
-
0
-
Unknown
Description
There was a system run for testing the upgrade code path in planner with shard affinity. It was observed that the planner variance was more than 25% during the rebalance. Creating this ticket to investigate the issue further.
Build - 7.6.0-2024
Steps to repro
- Initialize a 8 node cluster with 2 kv:n1ql nodes and 6 index nodes
- Create a bucket with 2 scopes and 6 collections and load each keyspace with 10k docs
- Load one the collections with 100k docs so that there is skew in data for that keyspace
- Create around 13 indexes on each keyspace with num replica ranging from 2-5
- Set indexer.settings.rebalance.redistribute_indexes:True
- Start continuous mutations on all the keyspaces. 1000 mutations per keyspace
- Enable shard based rebalance and swap rebalance one indexer node
- swap rebalance two indexer nodes
- swap rebalance three indexer nodes
- disable shard based rebalance
- swap rebalance three indexer nodes
- swap rebalance the other three indexer nodes
- Step 7-12 is considered as one cycle. So this cycle was run 20 times