mirror of
https://github.com/luxfi/consensus.git
synced 2026-09-24 00:26:58 +00:00
zeekay
b371706134
buildBlocksLocked held pendingBuildBlocks on a BuildBlock error and returned 'will retry
next tick', so the very next Notify re-called BuildBlock. Under single-proposer scheduling
a non-leader's BuildBlock ALWAYS fails ('not our proposer slot') until its window opens, so
the held demand became an unbounded hot loop — 4 of 5 nodes burning CPU rebuilding a slot
they can never win (observed alongside the mainnet rebuild storm). Mirror avalanchego: consume
the demand per attempt (decrement before the call), and on error clear any remaining queued
demand (it would fail identically this tick) and wait for a FRESH trigger — the node's slot
opening (proposervm WaitForEvent) or the elected leader's block advancing our tip. Liveness
preserved; log downgraded Error→Debug so a healthy fleet is quiet. Tests updated to assert the
new consume-on-error invariant (no spin, no accumulation, recovery via fresh Notify).