You are viewing an old version of this page. View the current version.

Compare with Current View Page History

Version 1 Next »

Status

Current stateUnder Discussion

Discussion thread

JIRA: Unable to render Jira issues macro, execution error.

Released: 

Motivation

KIP-708 was implemented some time ago in Kafka. This allows brokers and consumers to communicate about the rack (or AWS Availability Zone) they're located in. Reading from a local broker can save money in bandwidth and improve latency for your consumers.

Flink Kafka consumers currently cannot easily use rack awareness if they're deployed across multiple racks or availability zones, because they have no control over which rack the Task Manager they'll be assigned to may be in. 

This improvement proposes that a Kafka Consumer could be configured with a callback that could be run when it's being configured on the task manager, that will set the appropriate value at runtime if a value is provided. 

Public Interfaces

KafkaSourceBuilder will get a new method setRackId that will take a callback as an argument. Setting this value will result in the value of the `client.rack` property for the Kafka client on each Task Manager. Note that to see any behavior change or benefit from this setting, the Kafka brokers the client is connecting to must be configured to assign replicas with the RackAwareReplicaSelector and have their own broker rack IDs set correspondingly to the client rack IDs. 

https://kafka.apache.org/documentation/#consumerconfigs_client.rack

Proposed Changes

The callback for determining rack ID at run time will need to be passed from the KafkaSourceBuilder down to the KafkaPartitionSplitReader. When the KafkaPartitionSplitReader is instantiated, the callback will be run and the result added to the config properties passed to the KafkaConsumer, similar to how Client ID is handled now.


Compatibility, Deprecation, and Migration Plan

There shouldn't be any impact on existing users. This is new, non-required functionality.

Test Plan

Testing is going to include verifying that the properties are include the expected result for client.rack, and also that an Exception thrown by the callback does the right thing.

Rejected Alternatives

This can't be provided through KafkaSourceBuilder.setProperties() because Task Managers may sit in different racks or data centers, and thus must be configured at runtime.

  • No labels