Check the eviction policy on your Redis Cloud database before using it as a queue backend. If the tier defaults to volatile-lru, it evicts only keys with an expire set, which on a full database can look like lost queue jobs. For BullMQ or any job queue, noeviction is the correct policy: writes fail loudly under memory pressure instead of silently dropping jobs. Use allkeys-lru only for pure caches where losing entries is acceptable.

Context: Redis operational skill notes (k8s-ai-operator redis SKILL): documents the eviction-policy mismatch gotcha. Picking the wrong maxmemory-policy is a common source of confusing behavior: noeviction on a cache-only deployment causes writes to start failing under memory pressure instead of evicting, while allkeys-lru on Redis used as a primary store can silently evict data that was never meant to expire. volatile-lru only evicts keys with a TTL set.