You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
WriteAnswers grows without bound on nodes with no waiting clients #36
TRequestProcessor::ProcessCommitted pushes a TAnswer into WriteAnswers for every applied entry, on every node (raft.cpp#L634-L644). The only place that removes entries from the queue is ProcessWaiting, and only while completing a local waiting write (raft.cpp#L654-L661). A follower has no waiting writes (writes are forwarded to the leader).
Scenario
A follower applies a long stream of writes. Each applied entry adds a TAnswer that holds the TCommandResponse (and whatever the RSM returned, for example the SQL result text), and nothing ever pops it. Memory grows linearly with the number of applied entries, without bound. The same happens on a leader for entries that no local client is waiting for (for example forwarded writes), until some later local waiting write pops past them.
Expected
Replies are kept only for writes that are pending on this node, keyed by log index and term, and are dropped when the request is answered, when leadership changes, or when the client disconnects. No state grows with the number of applied entries on a node that has no waiting clients.
PR #31 (closed, not merged) has the fix in e117c18 (write replies for local pending writes only, keyed by index and term; pending requests fail on leadership change and on disconnect). Tests that cover it: test_follower_application_does_not_retain_unused_replies, test_replaced_write_never_receives_another_commands_reply, test_stepdown_rejects_write_without_waiting_for_application, test_disconnect_*. In the split plan it comes after the FIFO fix for ProcessWaiting. Since the PR is closed, the fix would come as a separate small PR.
Activity