Author here. The queue had been growing since June — 6.9M delayed jobs nobody was reading. The counterintuitive part was fix order: UNLINK the key before deploying the worker, because the first poll would have migrated all of them in one atomic Lua call and held Redis for minutes. Happy to answer questions about the triage or the Horizon setup.