databendlabs / openraft / 28652084830

87%

Build Type

push

github

Committed by drmingdrmer

Commit Message

fix: return an error instead of hanging on a dropped responder

# Summary

When a response channel closes while RaftCore is still running -- for example a
state machine drops an ApplyResponder without calling send() -- recv_msg used to
join RaftCore to fetch its stop error. A healthy core never finishes, so the
caller (e.g. client_write) hung forever. recv_msg now bounds that wait and
returns Fatal::Stopped when RaftCore is still running.

# Details

- wait_core_stopped detects a real stop by observing the metrics watch channel,
  without joining (and thus consuming) RaftCore. get_core_stop_error is only
  called once a stop is confirmed, so the bounded wait never cancels
  join_core_task mid-flight -- which would strand CoreState in Joining and panic
  a later caller through unreachable!().
- The wait is bounded by RECV_CORE_STOP_TIMEOUT (1s). A real stop surfaces almost
  immediately: RaftCore reports it in metrics and drops the metrics sender as its
  task ends, including on panic.

- Ref: #1816

Coverage Stats

15 of 19 new or added lines in 1 file covered. (78.95%)

5 existing lines in 3 files now uncovered.

17062 of 19607 relevant lines covered (87.02%)

159512.44 hits per line