Bug #47613
openpg autoscaler: don't always treat target ratios as a weight
0%
Description
https://tracker.ceph.com/issues/43947 has an undesireable bootstrapping effect that if an admin sets the target ratio of just one pool to a small value, e.g. 0.1, on a new cluster, the ceph will immediately scale that up to 1.0 and therefore consume all of the PGs for a cluster. (Once they add target ratios to other pools, then those PGs will eventually be merged, but this is all a complex waste of IO, imho).
IMHO it would be more intuitive if we treat target ratios as weights only if their sum is >1. If their sum is <1, we should use them as-is.
Updated by Neha Ojha over 3 years ago
I can see how this is problematic. Wondering if you've measured the performance impact of the PG merging operation in your cluster, just trying to evaluate how bad it is. Your proposal does solve this problem but can be confusing in the sense that we are changing the meaning of target_size_ratio here.