Hazelcast Jet

Achieves predictable low latency under load, processing millions of events per second on a single node with cooperative multithreading, as benchmarked in the README.

Provides exactly-once processing guarantees without external dependencies like Zookeeper, using distributed in-memory storage and the Chandy-Lamport algorithm, as highlighted in the features.

Applications scale up or down automatically as nodes are added or removed, preserving computational state and ensuring continuous processing, as described in the architecture.

First-class support for event time processing with distributed watermarks, effectively managing out-of-order event data, which is crucial for real-time streams.

Out-of-the-box support for Kafka, Hadoop, JDBC, Elasticsearch, and more, simplifying integration with common data sources and sinks.