[FLINK-17355] Exactly once kafka checkpointing sensitive to single node failures

• Type: Bug
• Status: Open
• Priority: Not a Priority
• Resolution: Unresolved
• Affects Version/s: 1.10.0
• Component/s: Connectors / Kafka
• Labels: auto-deprioritized-major auto-deprioritized-minor
• Assignee: Unassigned
• Reporter: Teng Fei Liao
• Created: 23/Apr/20 19:37
• Updated: 15/Nov/21 22:38

Description

With exactly one semantics, when checkpointing, FlinkKafkaProducer creates a new KafkaProducer for each checkpoint. KafkaProducer#initTransactions can timeout if a kafka node becomes unavailable, even in the case of multiple brokers and in-sync replicas (see https://stackoverflow.com/questions/55955379/enabling-exactly-once-causes-streams-shutdown-due-to-timeout-while-initializing).

In non-flink cases, this might be fine since I imagine a KafkaProducer is not created very often. With flink however, this is called per checkpoint which means practically an HA kafka cluster isn't actually HA. This makes rolling a kafka node particularly painful even in intentional cases such as config changes or upgrades.

 

In our specific setup, these are our settings:

5 kafka nodes

Per topic, we have a replication factor = 3 and in-sync replicas = 2 and partitions = 3