Skip to content

v0.6.17

@lhaig lhaig tagged this 20 Aug 20:36
Derived search/metadata indexing no longer runs inside the Raft apply path, so
opening a user's index handles cannot block mail delivery.

Raft's LastIndex no longer scans the entire transaction log to find its last
entry. On a log the size of production's this was ~20ms per call against
15.7us for a direct seek, and Raft calls it constantly.

Raft's own logs are no longer discarded. The consensus layer previously ran
completely silent, which made a delivery stall inside it impossible to explain
from the outside.

The delivery stage now reports a per-step breakdown when it is slow, and the
FSM reports slow applies and slow index work.

KNOWN ISSUE: the first delivery after a restart can still take ~70s. Normal
delivery is sub-second (median 0.59s) and warm delivery is ~63ms, so this
affects one message per restart. It is measured to be inside hashicorp/raft
between the write call and our FSM being invoked; see task B12.
Assets 2
Loading