Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
kvserver: rebalance ranges to minimize QPS delta among stores
This commit fixes the regression(s) introduced by cockroachdb#65379 where we observed replica thrashing in various workloads (cockroachdb#70396 and cockroachdb#71244). The following is a description of the differences between the QPS based rebalancing scheme used in the previous implementation of the store rebalancer (release-21.2 and before). ** lease rebalancing ** *** release 21.2 and before *** QPS based lease rebalancing in CRDB 21.2 considers the overall cluster level average QPS and computes underfull and overfull thresholds based off of this average. For each range that the local store has a lease for, the store rebalancer goroutine checks whether transferring said range's lease away will bring the local store's QPS below the underfull threshold. If so, it ignores the range and moves on to the next one. Otherwise, it iterates through the stores of all the non-leaseholder voting replicas (in ascending order of their QPS) and checks whether it would be reasonable to transfer the lease away to such a store. It ensures that the receiving store would not become overfull after the lease transfer. It checks that the receiving store doesn't have a replica that's lagging behind the current leaseholder. It checks that the receiving store is not in violation of lease preferences. Finally, it ensures that the lease is not on the local store because of access locality considerations (i.e. because of follow-the-workload). All of this was bespoke logic that lived in the store rebalancer (using none of the Allocator's machinery). *** master and this commit *** In cockroachdb#65379, we moved this decision making into the Allocator by adding a new mode in `Allocator.TransferLeaseTarget` that tries to determine whether transferring the lease to another voting replica would reduce the qps delta between the hottest and the coldest stores in the replica set. This commit adds some padding to this logic by ensuring that the qps difference between the store relinquishing the lease and the store receiving the lease is at least 200qps. Furthermore, it ensures that the store receiving the lease won't become significantly hotter than the current leaseholder. ** replica rebalancing ** *** release 21.2 and before *** QPS replica rebalancing in CRDB <=21.2 works similarly to the lease rebalancing logic. We first compute a cluster level QPS average, overfull and underfull thresholds. Based on these thresholds we try to move replicas away from overfull stores and onto stores that are underfull, all while ensuring that the receiving stores would not become overfull after the rebalance. A critical assumption that the store rebalancer made (and still does, in the approach implemented by this commit) is that follower replicas serve the same traffic as the leaseholder. *** master and this commit *** The approach implemented by cockroachdb#65379 and refined by this commit tries to leverage machinery in the Allocator that makes rebalancing decisions that converge load based statistics per equivalence class. Previously, this machinery was only used for range count based replica rebalancing (performed by the `replicateQueue`) but not for qps-based rebalancing. This commit implements a similar approach to what we do now for lease rebalancing, which is to determine whether a rebalance action would reduce the qps delta between the hottest and the coldest store in the equivalence class. This commit adds some safeguards around this logic by ensuring that the store relinquishing the replica and the store receiving it differ by at least 200 qps. Furthermore, it ensures that the replica rebalance would not significantly switch the relative dispositions of the two stores. An important thing to note with the 21.2 implementation of the store rebalancer is that it was making all of its decisions based on cluster-level QPS averages. This behaves poorly in heterogenously sized / loaded clusters where some localities are designed to receive more traffic than others. In such clusters, heavily loaded localities can always be considered "overfull". This usually means that all stores in such localities would be above the "overfull" threshold in the cluster. The logic described above would effectively not do anything since there are no underfull stores to move replicas to. Release note (performance improvement): A set of bugs that rendered QPS-based lease and replica rebalancing in CRDB 21.2 and prior ineffective under heterogenously loaded cluster localities has been fixed. Additionally a limitation which prevent CRDB from effectively alleviating extreme QPS hotspots from nodes has also been fixed.
- Loading branch information