Currently the balancer ensures the { ns: 1, min: 1 } index on the config.tags collection many times each balancer round — once per sharded collection.
Each time this happens, the config servers are first fsync'd, which is expensive and causes a high volume of fsync messages in the config server logs (e.g. as observed on SERVER-16638).
The solution is to move lines 428 – 438 of s/balance.cpp to be before the start of the loop over sharded collections.
- is related to
-
SERVER-17151 Avoid fsync command in ConfigCoordinator
- Closed
- related to
-
SERVER-16638 remove fsync messages from config server logs
- Closed