A data platform that fits in one file.

havn loads your data with Python, models it with plain SQL and serves it through a web UI. Everything runs on DuckDB, and the whole warehouse is a single file on a machine you control.

pip install havn

Getting started

havn is Norwegian and Danish for harbour. It is made for teams whose data fits on one server, which is most of them, and who would rather not rent a cluster to find that out.

Recorded at real speed (typing sped up): the sample project runs from Home, then a query reads the gold table it just built.

havn is Norwegian and Danish for harbour. It is made for teams whose data fits on one server, which is most of them, and who would rather not rent a cluster to find that out.

1 file: holds the whole warehouse
1 package: no services to run beside it
14 connectors: databases, SaaS APIs, files
~3 seconds: to load and build the sample project

A model is a SQL file.

One line of @config at the top, then a SELECT. havn reads the FROM and JOIN clauses to work out the order to build things in. There is no templating language and no compile step.

Each run rebuilds only what changed: the SQL itself, anything upstream of it, or new data loaded since the last build. Assertions run right after the build, and a failing one stops the models below it.

transform/gold/tonnes_by_port.sql

@config materialized=table, schema=gold
@description Shipments and tonnage per port.

SELECT s.port, COUNT(*) AS shipments,
       round(SUM(s.tonnes), 1) AS tonnes
FROM silver.shipments s
GROUP BY s.port
ORDER BY tonnes DESC

# a second batch of shipments lands
$ havn run ingest/shipments.py
  done shipments.py (4ms)

# the new rows reach gold, no --force needed
$ havn transform
  skip  bronze.shipments
  done  silver.shipments (75 rows, 2ms)
  done  gold.tonnes_by_port (4 rows, 5ms)

$ havn query "SELECT * FROM gold.tonnes_by_port"
  Bergen      21   1239.6
  Stavanger   18   1078.2
  Ålesund     18    946.8
  Tromsø      18    815.4

What comes with it

Everything below ships in the one package, and each part works from both the command line and the web UI.

Loading data	Python scripts with a database connection ready to use, plus connectors for Postgres, MySQL, Snowflake, BigQuery, Redshift, Databricks, Stripe, HubSpot, Shopify, Google Sheets, S3, CSV, REST and webhooks.	`havn connect` `havn run`
Modelling	SQL or Python models as views, tables, incremental tables (merge, append, delete+insert, microbatch) or SCD2 snapshots. Graph selectors such as `+gold.orders` and `tag:daily`.	`havn transform`
Testing	Inline assertions, YAML contracts, unit tests with fixture rows, and a bind pass that catches unknown columns and type errors before anything runs.	`havn check` `havn test`
Shipping	Environments with their own warehouse file. Deploy plans that show what will rebuild, a pre-deploy snapshot, and rollback when a deploy fails.	`havn deploy prod --plan`
Reviewing	A warehouse per git branch that reads unchanged models from production. Row-level diffs of what a change does to each table, column lineage, and generated GitHub Actions that post the diff on pull requests.	`havn branch diff` `havn ci generate`
Running	Jobs defined in YAML, cron schedules, retries, a file watcher, and Slack or webhook alerts. Live models refresh within seconds of new rows from webhooks or Postgres CDC.	`havn jobs run` `havn live`
Performance	The plan and timings of every model build, alerts when a model gets slower than usual, and advice such as which tables to make incremental. Metrics for Prometheus, OpenTelemetry and OpenLineage.	`havn perf`
Exploring	A SQL editor with autocomplete, a table browser, the model graph, notebooks, dashboards and a metrics layer. Dashboards can be shared by link, embedded, or sent by email and Slack on a schedule.	`havn serve` `havn reports`
Access	Users with admin, editor and viewer roles, row-level security, column masking that follows the data into downstream models, and an audit log. Scripts run by restricted users only see what those users may see.	`havn serve --auth` `havn rls`
Recovery	Verified backups and restore points for the whole warehouse.	`havn backup` `havn rewind`
AI assistants	An MCP server, and context files for Claude Code, Cursor and Copilot in every new project. Questions answered from your defined metrics, with the SQL shown, and model edits from agents checked in a scratch build before you apply them.	`havn mcp` `havn ask`

For the parts that are still missing or rough, see What havn supports.

havn Query page: the table list by schema on the left, a SQL query over silver.shipments, and four result rows grouped by port — Data › Query: write SQL against any layer and inspect the result.

havn Ask page: the question 'How many tonnes did each port land?' answered from the tonnes_landed metric, with a bar chart and a table by port — Data › Ask: a question answered from a defined metric, marked verified, with the SQL one click away.

havn Governance page previewing a query as the editor ingrid: a row policy limits her to Bergen and Ålesund, and two derived models inherit it — Settings › Governance: what a user sees once row policies, and the ones inherited downstream, apply.

havn Performance page: slowest models with build-time trends, an advice card about an ORDER BY in a model others read, and the run's critical path — Observe › Performance: build times per model, advice with the evidence, and what set the run's length.

A published havn dashboard, Port operations: three numbers, tonnes by port as a bar chart, and each port's share of the latest batch — A published dashboard: a read-only link that works on a phone, or a scheduled email.

Who it is for

Teams that have outgrown spreadsheets and CSV exports but do not need, or want to pay for, a cloud warehouse sized for data they will never have.

A single DuckDB file handles hundreds of millions of rows on an ordinary server. You can back it up by copying it and open it with any DuckDB client.

Where it stops

One machine does the work. There is no distributed compute.
The DuckDB file allows one writer at a time. Reads run alongside it.
Data well beyond a single disk belongs in a bigger system.

Running it

havn Core

On your own hardware

Install it from PyPI or run the Docker image on a laptop, a server or your own cloud account. It is free to use inside your company, including in production. The source is published under the Business Source License 1.1, and each release becomes Apache 2.0 after four years.

havn Cloud · in development

Hosted in Norway

A managed instance on Norwegian infrastructure run by a non-US company, with a fixed monthly price. Follow the repository for news.

Try it on the sample project

Python 3.10 or newer is the only requirement. The sample loads earthquake data from the USGS feed.

pip install havn
havn init my-project and cd my-project
havn jobs run full-refresh to load and build everything
havn serve and open localhost:3000

With Docker instead: docker run -v "$(pwd)":/project -p 3000:3000 ghcr.io/chraltro/havn