In #83851, we prevent sending Raft snapshots to followers that have an overloaded engine:
|
repl.mu.RLock() |
|
_, destPaused := repl.mu.pausedFollowers[id] |
|
repl.mu.RUnlock() |
|
if ioThresh := repl.store.ioOverloadedStores.Load()[repDesc.StoreID]; ioThresh != nil && destPaused { |
|
// If the destination is paused, be more hesitant to send snapshots. The destination being |
|
// paused implies that we have recently checked that it's not required for quorum, and that |
|
// we wish to conserve I/O on that store, which sending a snapshot counteracts. So hold back on |
|
// the snapshot as well. |
|
err := errors.Errorf("skipping snapshot; %s is overloaded: %s", repDesc, ioThresh) |
|
repl.reportSnapshotStatus(ctx, repDesc.ReplicaID, err) |
|
return false, err |
|
} |
However, we don't do this when sending snapshots to new learners:
|
if err := r.sendSnapshot(ctx, rDesc, kvserverpb.SnapshotRequest_INITIAL, priority); err != nil { |
|
return nil, err |
|
} |
We probably should. Erroring out here may be sufficient.
Jira issue: CRDB-18266
Epic CRDB-15069