Batchlog endpoint candidates should be picked randomly, not sorted by proximity. I'll be lazy and just copy-paste some lines from IRC:
[20:23:27] rbranson: is there an issue where batch logs tend to get written to a subset of the nodes?
[20:28:04] rbranson: I mean all the write batches are going thru 10% of the nodes
[20:28:16] rbranson: it means writes won't scale linearly w/the cluster size
Attaching a trivial patch.
|Status||Open [ 1 ]||Patch Available [ 10002 ]|
|Status||Patch Available [ 10002 ]||Resolved [ 5 ]|
|Resolution||Fixed [ 1 ]|