A cage in search of a bird: With streaming on an all-time high, Kafka trims latencies, adds security

A cage in search of a bird: With streaming on an all-time high, Kafka trims latencies, adds security

In a particularly timely release, distributed data streaming platform Apache Kafka –  used by companies including Netflix, Coursera and Slack – has unleashed version 2.5, aiming for a more stable, less laggy experience.

One of the problems the Kafka team tackled in the new release is the aggregation of multiple streams into a larger object, which was a bit of a stolid process before. To accommodate such use cases better, the tool’s DSL now comes with a new operator called cogroup(). Instead of creating state stores for each stream and a chain of ValueJoiners to keep them together, it uses a single state store for the new object. Besides the better representation, this approach is also said to improve performance, which is always helpful in data heavy scenarios.

Another enhancement is an attempt to reduce latencies for applications using multiple partitions on a single instance. Instead of letting the appropriate get function run through a number of stores when a stream instance is calling for a particular key, the platform now allows fetching the key from a single partition. The implementation of this was mainly made possible by the team’s work on incremental rebalance to work around node partitions being unavailable, since this helped to expose information on which key belongs to which partition. 

To make the platform a bit more secure, Kafka 2.5 is able to make use of TLS 1.3, though v1.2 is still the default. Earlier versions have been disabled due to security vulnerabilities. Should a setup still need those protocols for some reason, they can however be enabled via configuration options ssl.protocol and ssl.enabled.protocols. 

Resilience was another item on the 2.5 todo list, which is why default values for zookeeper.session.timeout.ms and replica.lag.time.max.ms have been increased from 6s to 18s and 10s to 30s respectively. 

When making the switch to the new version, users should be aware that UsePreviousTimeOnInvalidTimestamp has been replaced by UsePartitionTimeOnInvalidTimeStamp, and KafkaStreams.store(StoreQueryParameters) has taken over for KafkaStreams.store(String, QueryableStoreType), which means some adjustments in older code might be necessary. Additional details can be found in the release notes.

Apache Kafka  was initially developed at job networking site LinkedIn, which is now part of Microsoft. It is meant to help with handling real-time data feeds, though lately more and more machine learning engineers seem to have developed a taste for it as a building block for complex ML systems.