Penca — The operational layer for your lakehouse

The problem

Your data lives in the lakehouse. Production can't.

Your lakehouse is the cheap, open, permanent home for your data. It's where analytics, ML, and AI all run. But you can't serve production from it, so your live writes go to Postgres, then CDC copies them into the lake. The copy lags, the pipeline pages you, and you store everything twice.

Production runs on Postgres, so your experiments have to as well. Testing a new model, validating a migration, giving an agent a sandbox: each one means cloning the full Postgres dataset and its environment. That's slow and expensive, so you run only a handful at a time.

When something breaks, you can't see who changed what. Debugging takes hours, rolling back means manual surgery. Audit and restore end up as custom projects.

                                                                                               one storage bill
  Postgres / Spanner / MySQL  ────  CDC ~~~ replication lag ~~~  ────►  your lakehouse  ────►  consistent reads
                                                                                               simple ops

Why now

Agents make it worse.

The OLTP/OLAP split has been painful for a decade. Weekly refreshes worked when experiments took quarters. Now AI agents iterate by the minute. They branch aggressively, read data that was written ten seconds ago, and write data with no human in the loop to catch the mistake. Same problem, less patience.

The solution

Bottomless storage. Branching. Versioning.

Built around all three from day one. Nothing bolted on.

Bottomless object storage

Bring your own bucket and only pay for compute. Or let us host the storage if you'd rather skip the setup. Bottomless capacity, open formats, zero lock-in.

Branch like Git, on real data

Fork live production to isolated branches with zero data copy and zero shared compute in minutes. Run experiments in parallel without setup or a queue. Throw the branch away, or promote it.

Row-level versioning

Every mutation appended to an immutable log with author and timestamp. Every consistent state of the database is auditable and recoverable at any point in time. It’s Git for data.

No lock-in

Every interface is an open standard.

SQL, gRPC, Iceberg — all industry standards your team already uses. Your tools work. Your data is portable. Nothing proprietary to migrate off.

SQL

Standard SQL via JDBC, ODBC, or ADBC. Whatever’s in your BI tool already works.

gRPC

Programmatic access via standard gRPC clients in any language.

Iceberg

Iceberg tables in object storage at rest. Read your data with DuckDB, Polars, pandas, Spark.

Query your data with SQL, gRPC, or your favorite analytics tool:

# Standard SQL via JDBC, ODBC, ADBC
SELECT transaction_id, customer_id, region, price
FROM checkout_events;

# gRPC API in any language
client.read_data(
    table_name="checkout_events",
    columns=["transaction_id", "customer_id", "region", "price"],
)

# Or read directly from your lakehouse, no Penca required
catalog.load_table("penca.checkout_events").scan(
    selected_fields=("transaction_id", "customer_id", "region", "price")
).to_pandas()

Shape what’s next

Apply to be a design partner.

We want a handful of teams running production on a split system today, and feeling the pain daily. You get direct access to the engineering team and a voice in the roadmap. We get sharp signal from people who live the problem.

Ready to delete your CDC pipeline?

Send a note about your team: what you’re running today, what you’re trying to unlock. We'll reply within a few days.

[email protected]