[13.x] Queue workers crash-loop on Redis Cluster / ElastiCache Serverless: cross-slot MGET in getPausedQueues() → TypeError in PhpRedisConnection::mget()

[dan-zonalusa]

Laravel Version

13.25.0

PHP Version

8.4

Database Driver & Version

No response

Description

After upgrading to v13.25.0, every queue:work poll iteration throws:

TypeError: array_map(): Argument #2 ($array) must be of type array, false given
at vendor/laravel/framework/src/Illuminate/Redis/Connections/PhpRedisConnection.php:67

The exception is thrown inside Worker::getNextJob() before a job is popped, is caught by its
catch (Throwable) block, the worker sleeps 1s and retries — and fails again, forever. The net
effect is that queue workers stop consuming jobs almost entirely (we saw this take down all
production queues; cache/queue backend is AWS ElastiCache Serverless Valkey 7.2, which
enforces Redis Cluster semantics).

The chain:

1. [13.x] Add a global pause switch for queues #61126 (shipped in v13.25.0, "Add a global pause switch for queues") changed
   QueueManager::getPausedQueues() to fetch a global pause key alongside the per-queue keys:

   $keys = array_map(fn ($queue) => "illuminate:queue:paused:{$connection}:{$queue}", $queues);
   
   $states = $this->app['cache']
       ->store()
       ->many(
           array_merge(['illuminate:queues:paused'], $keys)
       );

2. On a Redis cache store this becomes a multi-key MGET. On any cluster-semantics Redis
   (Redis Cluster, ElastiCache Serverless, Valkey serverless), illuminate:queues:paused and
   illuminate:queue:paused:{connection}:{queue} hash to different slots, so the server replies
   CROSSSLOT Keys in request don't hash to the same slot — on every single worker loop.

3. phpredis signals command error replies by returning false (no exception), so
   PhpRedisConnection::mget() does array_map($closure, false) → TypeError. The real
   CROSSSLOT error is swallowed (getLastError() is never surfaced), which made this
   painful to diagnose.

Notes:

• [13.x] Optimize Worker queue pause check #60109 (v13.24.0) introduced the many() batching for the pause check. That could already
  cross-slot for a worker polling multiple queues (--queue=a,b), but the unconditional global
  key in v13.25.0 makes even single-queue workers fail 100% of the time.
• RedisStore::many() special-cases PredisClusterConnection and falls back to manyAlias()
  (per-key gets), but there is no equivalent handling for phpredis — and none is possible via
  connection type when the endpoint enforces cluster semantics behind a single proxy endpoint
  (ElastiCache Serverless is used with the regular non-cluster phpredis client).

Steps To Reproduce

1. Point CACHE_DRIVER=redis and QUEUE_CONNECTION=redis at any cluster-semantics Redis
   (Redis Cluster, or AWS ElastiCache Serverless Redis/Valkey) using phpredis.
2. Run php artisan queue:work.
3. Every poll iteration logs the TypeError above and no jobs are consumed.

Minimal reproduction without a worker:

Cache::many(['illuminate:queues:paused', 'illuminate:queue:paused:redis:{default}']);
// TypeError: array_map(): Argument #2 ($array) must be of type array, false given

Truncated trace:

#0 .../Redis/Connections/PhpRedisConnection.php(67): array_map(Object(Closure), false)
#1 .../Cache/RedisStore.php(112): PhpRedisConnection->mget(Array)
#2 .../Cache/Repository.php(167): RedisStore->many(Array)
#3 .../Queue/QueueManager.php(335): Repository->many(Array)
#4 .../Queue/Worker.php(489): QueueManager->getPausedQueues('redis', Array)
#5 .../Queue/Worker.php(450): Worker->getPausedQueues('redis', Array)
#6 .../Queue/Worker.php(243): Worker->getNextJob(Object(RedisQueue), '{default}')

Expected behavior / suggested fixes

Any of these would have prevented the outage, and they're complementary:

1. Fail open: a failure while checking pause state should not prevent job consumption —
   wrap the pause lookup so workers keep processing when the cache read fails.
2. Cross-slot-safe lookup: fetch the global key separately from the per-queue keys (or fall
   back to per-key gets à la manyAlias() when MGET fails), so cluster deployments work.
3. Don't swallow the Redis error: PhpRedisConnection::mget() should detect false and
   throw a RedisException including getLastError() instead of a bare TypeError from
   array_map().

Workarounds

• Pin laravel/framework to ~13.24.0 (sufficient for single-queue workers), or
• Queue::withoutInterruptionPolling() in a service provider (disables pause and restart polling).