Details
-
Bug
-
Resolution: Fixed
-
Major
-
4.1.2, 4.6.4, 5.0.0, 5.5.0
-
CentOS 7
E5-2680 v3 (48 vCPU)
64 GB RAM
Samsung PM863a SATA SSD
-
Triaged
-
Centos 64-bit
-
-
No
Description
Test scenario:
- 4 nodes
- 1 bucket, 1 replica, full eviction
- 1B items (~1KB), 5-10% resident ratio
- 15K ops/sec (90% read, 10% update), 10% cache miss ratio (before rebalance)
- Swap rebalance of one node (172.23.96.103 -> 172.23.96.104)
Swap rebalance causes high memory usage on the new node and, as a result, thousands of TMP OOM failures. Although the number of failures is not that big (<100K), TMP OOM errors obviously cause significant drops in the rate of the incoming GET and SET requests.
Graphs: http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=titan_510-1368_rebalance_a416