- Size MAX_ACCOUNTS_PER_WORKER to ~1.2x the fair share so the fleet fills each
worker evenly at startup and the rebalancer barely sheds; without this, three
coordinators racing for the whole fleet shed/reclaim in a thrash that reset
reclaimed accounts' first-poll clocks and dragged the sync-lag ramp.
- Sample a "dead" count (owned runners whose work loops have stopped) to
distinguish a real zombie-runner bug from a slow ramp. Validated: dead=0
throughout; 400-account run holds lag ~48s under a 52s budget across 13 steady
samples, persistent sockets == hot, RSS flat.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lukasandClaude Opus 4.8
ee4dd99ad1