🔒 Paid
Kafka Partition Skew
A single large tenant's traffic has spiked, and its messages are concentrated on one Kafka partition. Aggregate consumer-group utilization still looks moderate — most consumers are nearly idle — but the one partition that tenant's messages land on is falling critically behind, because only one consumer can ever read from a given partition at a time. Diagnose why the backlog keeps growing despite healthy-looking aggregate metrics, and restore throughput before downstream processing falls critically behind.
- Difficulty
- Staff
- Expected Time
- ~18 min
- Category
- Messaging & Streaming
Skills
- Diagnosing per-partition hot-key bottlenecks
- Partition key design and salting strategies
- Recognizing when scaling consumers does not help
- Reading aggregate vs. per-partition metrics correctly
System Architecture
- Producer Service
- Kafka Topic
- Hot Partition
- Consumer Group
- Downstream Processor
This is a Pro scenario. Upgrade to Pro to unlock it.
Upgrade to ProBack to Incidents