[KAFKA-4207] Partitions stopped after a rapid restart of a broker

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 0.9.0.1, 0.10.0.1
• Component/s: controller
• Assignee: Unassigned
• Reporter: Dustin Cote
• Reviewer: Jun Rao
• Created: 22/Sep/16 17:22
• Updated: 26/Sep/16 23:04
• Resolved: 23/Sep/16 22:16

Description

Environment:
4 Kafka brokers
10,000 topics with one partition each, replication factor 3
Partitions with 4KB data each
No data being produced or consumed

Scenario:
Initiate controlled shutdown on one broker
Interrupt controlled shutdown prior completion with a SIGKILL
Start a new broker with the same broker ID as broker that was just killed immediately

Symptoms:
After starting the new broker, the other three brokers in the cluster will see under replicated partitions forever for some partitions that are hosted on the broker that was killed and restarted

Cause:
Today, the controller sends a StopReplica command for each replica hosted on a broker that has initiated a controlled shutdown. For a large number of replicas this can take awhile. When the broker that is doing the controlled shutdown is killed, the StopReplica commands are queued up even though the request queue to the broker is cleared. When the broker comes back online, the StopReplica commands that were queued, get sent to the broker that just started up.

CC: junrao since he's familiar with the scenario seen here

Issue Links

• duplicates: KAFKA-1342 Slow controlled shutdowns can result in stale shutdown requests