A Laravel tagged-cache flush can trigger “BUSY Redis is busy running a script” when one atomic Lua script performs too much work. Confirm the slow script first, then reduce the keys and tags processed together instead of treating the error as a network failure.
Last updated: November 16, 2026.
redis-cli SLOWLOG GET 20
redis-cli CONFIG GET busy-reply-threshold
redis-cli --latency
# Inspect tag cardinality with application-specific key names.
redis-cli ZCARD "your-cache-prefix:tag:example"Run diagnostics against the affected Redis instance. Avoid printing production values that contain tenant identifiers or secrets, and do not delete keys until you understand the application’s tag layout.
Why other Redis clients also fail
Redis executes a Lua script atomically. The Redis BUSY error documentation explains that once a script exceeds busy-reply-threshold, five seconds by default, Redis rejects most other commands until the script completes. Queue workers, sessions and locks can therefore fail even when they use different logical databases on the same server.
A current Laravel framework report shows how tagged-cache cleanup can multiply work across many keys and tags. The exact behavior is version-sensitive, so record the Laravel, Redis client and Redis server versions before changing code.
Reduce the atomic work
Remove tags that do not provide useful invalidation boundaries. Prefer a small stable tag set, versioned cache namespaces or narrower invalidation jobs over thousands of tags in one flush. If the cache is operationally important, separate cache data from sessions, locks and queues so a large invalidation cannot block every workload.
Use SLOWLOG duration and the application cache prefix to connect a slow EVAL or EVALSHA call to the flush that started it. Record the number of affected keys and tags so the same workload can be reproduced outside production.
Do not raise busy-reply-threshold as the primary fix; that only delays the warning while the server remains blocked. Redis notes that SCRIPT KILL cannot stop a script after it has written data. Test the redesigned invalidation with production-like key counts, monitor SLOWLOG and latency, and confirm that each batch completes well below the threshold.