// the find
timeplus-io/proton
⚡ Fastest SQL ETL pipeline in a single C++ binary, built for stream processing, observability, analytics and AI/ML
Proton is a streaming SQL engine built on top of ClickHouse internals, packaged as a single C++ binary with no JVM or ZooKeeper dependency. It handles Kafka ingestion, tumble/hop/session windowing, incremental materialized views, and CDC in one process — think ksqlDB but without the Java tax. Aimed at teams who want real-time stream processing without operating a Flink cluster.
No JVM means it actually starts fast and runs on an AWS t2.nano — the contrast with Flink's resource appetite is real, not marketing. Built on ClickHouse's vectorized engine, so you get SIMD-accelerated aggregations and thousands of battle-tested SQL functions for free rather than reimplementing them. The RANDOM STREAM generator is a genuinely useful development feature: you can prototype a full pipeline locally without standing up Kafka. Incremental materialized views that persist queryable state inside Proton itself is the right architecture for dashboards that need both live and historical data.
Clustering is enterprise-only, which means the OSS version has no HA story — if the process dies, your pipeline dies. That's a hard stop for production use at any real scale. The feature gap between OSS and Enterprise is aggressive: webhooks, NATS, and the entire Redpanda Connect connector catalog are gated behind a sales call, so the connector matrix in the README is mostly not what you actually get. At 2,200 stars with a VC-backed company driving it, the risk of the open-source version being quietly deprioritized in favor of the commercial product is non-trivial — the OSS tier looks more like a lead-gen funnel than a standalone project. Watermark and late-arrival handling for out-of-order events is mentioned but not explained anywhere in the README or surface-level docs, which is precisely where stream processing gets hard.