Seamlessly unite AI, analytical and OLTP workloads with pgEdge ColdFront
Fully writable hot and cold storage through the same SQL for up to 90% lower storage costs.
No application changes. No vendor lock-in.
No trade–offs. One table name. Three workload types.
pgEdge ColdFront unites three workload types that have historically required separate systems: OLTP transactions at full PostgreSQL speed, analytical queries at DuckDB columnar speed, and AI workloads with access to the complete data history, all through the same table(s) name and the same SQL.
ColdFront keeps active data in native PostgreSQL and moves the rest to Apache Iceberg on S3-compatible object storage, automatically and transparently. Applications keep using SELECT, INSERT, UPDATE, and DELETE against the same table name. ColdFront handles which rows live where. No separate analytics system, no ETL pipeline, no glue code.
Consider a GDPR deletion request against five-year-old archived data. With ColdFront, this is one SQL statement - not a restore-to-hot, delete, re-archive, and re-verify cycle. The cold tier is writable by default.
Up to 90%
lower storage cost on cold data vs. SSD-backed PostgreSQL*
Up to 22x
faster analytical queries on cold Parquet data via DuckDB
4.4M rows/sec
validated throughput across 6-node Spock mesh, 60M rows
0 sidecars
DuckDB runs in-process, no separate daemons or cluster managers
The only directly writable cold tier on unmodified, open source Postgres.
Most tiering solutions ask you to trade your PostgreSQL distribution, your ability to write to archived data, or your ability to self-host. ColdFront doesn't.
Fully Writable Cold Tier
Most tiering products make archived data read-only. ColdFront's cold tier supports UPDATE and DELETE through the same table name, so a GDPR deletion request or data correction is a single SQL statement, not a restore-delete-rearchive ops project.
Actually Open. No Lock-In.
Runs on stock upstream PostgreSQL from standard community packages. Cold data is Apache Iceberg (Parquet on any S3-compatible store), readable by any Iceberg-capable engine with no ColdFront dependency and no proprietary format at any layer.
In-Process Analytical Speed
DuckDB runs as a library call inside the PostgreSQL process. No separate daemon, no Arrow Flight RPC, no sidecar. Cold-tier analytical queries run up to 22x faster than row-based storage on the same data.
Distributed Cold Writes
On a pgEdge Spock mesh, every node can write to cold data concurrently. The bakery protocol, model-checked in TLA+, serializes Iceberg commits across nodes with no 409 conflicts and no application-level retry. Validated throughput at 4.4M rows/sec across a 6-node mesh with 60M rows.
Zero Application Changes
Tiering is a property of the deployment, not the application. The SQL before and after ColdFront is identical. No refactoring, no ORM changes, no new data access patterns.
Minimal Operational Footprint
The full ColdFront stack adds two in-process extensions, one small Rust catalog binary (Lakekeeper, ~20 MB), and object storage you probably already have. No cluster managers, no consensus groups, no minimum node count.
Your full data history, available to every AI workload.
AI agents and ML pipelines need access to the complete dataset, not just the recent working set. ColdFront makes that history queryable at analytical speed through standard PostgreSQL with no separate system required.
In decoupled mode, PostgreSQL becomes a stateless compute front-end over Iceberg. AI agents connect via standard PostgreSQL drivers and query terabytes of history without data duplication. Agent governance platforms generate exactly the workload pgEdge ColdFront was built for: append-heavy traces, mandatory retention, infrequent reads, and occasional corrections. ColdFront's writable cold tier closes the correction loop when a provenance trace leads back to a stale source record. Combined with the pgEdge Agentic AI Toolkit (MCP Server, RAG Server, Vectorizer, and Docloader), ColdFront turns PostgreSQL into a complete data backbone for agentic AI in regulated environments.
Runs on stock PostgreSQL. No forks, no sidecars.
ColdFront is a small, composable stack built on open components.| Component | WHAT IT IS |
|---|---|
| PostgreSQL 16/17/18 | Stock upstream, installed from standard community packages. No fork, no proprietary distribution. |
| pg_duckdb | DuckDB running in-process inside PostgreSQL. Handles all Iceberg 1/O as a library call, no separate daemon. |
| coldfront extension | Small C extension that rewrites DML to hit the correct tier. Mirrors schema DDL across both tiers. |
| Lakekeeper | Apache Iceberg REST catalog. Single Rust binary (~20 MB). Apache 2.0 licensed. |
| Archiver / Partitioner | Static Go binaries (~9 MB). Run via cron. No daemon, no runtime dependencies. |
| Any S3-compatible store | AWS S3, MinIO, GCS, Azure Blob, SeaweedFS, or any S3-compatible endpoint. |
Open Source. Your Infrastructure. Your Data.
Enterprise-ready with a subscription.The coldfront extension and archiver are fully open source under the PostgreSQL License. pgEdge Enterprise Postgres adds what production deployments require.| Community PostgreSQL + ColdFront | pgEdge Enterprise Postgres + ColdFront |
|---|---|
|
|
Ready to give it a try?
pgEdge is the only Enterprise-grade Postgres platform that works everywhere for AI, high availability and more.