Modak

Tier-aware data federation between Postgres and Apache Iceberg. Modak knows which tier holds what, and every query stays real time and consistent.

Recent rows live in Postgres, history lives in the lake, and plain SQL works against the whole timeline: SELECT, INSERT, UPDATE, and DELETE reach any row, wherever it lives.

Status: beta

Modak has not cut a stable release. Interfaces can still change. See the note in Production deployment before running it anywhere that matters.

Each table chooses how the two systems share its rows:

• A tiered table keeps only its recent partitions in Postgres and moves the rest to Iceberg. This fits append-mostly, time-series data like telemetry and event logs, where old rows should stop taking up Postgres space.
• A mirrored table keeps its full copy in Postgres while CDC trails every change into the lake. It can optionally shed heap history it no longer needs hot. This fits reference tables that must stay fully writable in Postgres but should also be visible to the lakehouse.

Either way, a thin open seam stitches the tiers: a monotonic cut-line, a pinned lake snapshot, and a PK-keyed correction delta merged on read. Every query sees a consistent point-in-time view with no duplicates and no gaps.

Why Modak

Both tiers stay real, independently usable systems:

• A real Postgres you run OLTP on. Modak is an extension plus a schema, not a fork or a proxy.
• A real Iceberg warehouse any engine can read. Spark, Trino, and DuckDB see standard Iceberg tables with no Modak anywhere.

Modak owns only the glue: routing, tiering, mirroring, and the consistency protocol that makes a two-tier table read like one table.

What deploys where

Piece	What it is	Runs
modak extension	Planner hook (transparent reads), write routers, read pins	Inside your Postgres
modak.* schema	Plain catalog tables, the coordination contract	Inside your database
Worker (modak-worker.jar)	Tiering, CDC mirroring, compaction, maintenance, CLI	Alongside Postgres (VM, k8s, ...)
Console (modak-console.jar)	The worker plus an embedded web console	Optional, instead of the worker

Postgres itself, the S3-compatible object store, and the optional Iceberg REST catalog are yours. Modak never hosts them. See Production deployment for the full boundary.

Start here

• Quickstart: a full local stack and a scripted walkthrough in about ten minutes.
• Concepts: the five ideas the rest of the docs assume. Table modes, tier key, cut-line, pinned snapshot, delta.
• Choosing a mode: which mode fits your workload, decided from the shape of the data.
• The contract: the mode-by-mode matrix of what you can read, write, and correct, and which surface does it.
• Architecture: how the extension, the worker, and the catalog cooperate without ever calling each other.