// the take
A CDC pipeline that streams changes from MySQL and PostgreSQL (with MongoDB, SQL Server and Oracle marked experimental) into ClickHouse, using Debezium to read the source replication log. It ships as either a single lightweight process or a full Kafka Connect deployment, aimed at teams running ClickHouse as an analytics layer next to a transactional database who need ongoing replication rather than one-off batch loads.
Builds on Debezium for the actual binlog/WAL parsing instead of rolling its own CDC layer, so the hardest part — correctly reading a replication stream — is someone else's battle-tested code. It handles automatic DDL/schema-change propagation and checksum-based table comparison, which most CDC tools leave entirely to the user. Two deployment modes (single lightweight process vs. Kafka-based) mean you can start simple and only take on a Kafka Connect cluster when you actually need the throughput or decoupling. It's backed by Altinity, the company behind a commercial ClickHouse distribution, and runs in production inside Altinity.Cloud rather than being an unmaintained side project.
Writes land in ClickHouse's ReplacingMergeTree, which deduplicates asynchronously via background merges — queries against unmerged parts can return duplicate or stale rows unless you add FINAL or manage merges yourself. MongoDB, Oracle and SQL Server sources are explicitly experimental; the well-tested path is MySQL and PostgreSQL only, despite the project description listing MongoDB as a first-class source. The Kafka deployment mode trades the simplicity of the lightweight process for a full Kafka Connect plus schema-registry stack, so you're really choosing between two different operational models rather than scaling up one. At 347 stars for a project with this much surface area (two deployment topologies, five source databases), there's a thinner trail of third-party production war stories than more established CDC tools like Debezium itself or Airbyte.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →