Integrations

Modak has two integration surfaces. External engines read the consistent two-tier view through the seam protocol, and external producers write through Stream Load. Everything in this section builds on one of the two.

Reads: the seam protocol

The worker keeps the seam state in plain catalog tables, so any engine that can reach Postgres and the lake can produce the same consistent view the extension does. That matters most on managed Postgres (RDS, Aurora, Cloud SQL), where the extension cannot be installed and a connector becomes the read path for tiered data.

Connectors share modak-connector, a small engine-agnostic library covering the pin lifecycle and the atomic seam capture over JDBC. The engine half stays thin: scan the heap above the cut-line, scan the lake at the pinned snapshot, merge the delta.

Connector	Status
Spark	Available
Trino	Planned
DuckDB	Planned
Athena	Planned

Writes: Stream Load

Stream Load is the ingestion counterpart. The modak-load library gives any JVM process labeled exactly-once micro-batches, and the worker exposes the same path over HTTP for everything else. Kafka Connect, Flink, and Spark Structured Streaming sinks built on it are planned.

Anything not listed can still read the lake directly. Modak writes standard Iceberg, and a plain snapshot read is consistent, just bounded at the cut-line rather than merged with the heap.