DuckCon #7
DuckLake on Hetzner: Your own data lakehouse for under €15 a month
Floyd Berndsen · DuckCon #7
6
sections
8
slides
01 / Intro
- Floyd shares how he discovered Hetzner as a very cheap way to run your own data lakehouse.
- The talk focuses on making that deployment easy to reproduce.
- Framing: run a full DuckLake stack for under €15 a month.
02 / Who
- Floyd Berndsen: ~8 years in data, and much longer in software.
- Newly appointed Head of Data at All Options (options trading); team is hiring.
- Works with DuckDB and DuckLake as a side quest / consultant (Berndsen.io).
More detail
All Options is an options-trading firm; Floyd started the Head of Data role the week of the talk.
03 / Why
- Cost sensitivity: Hetzner is chronically cheaper, roughly a third of AWS pricing.
- Strict data requirements: German-based provider, keeps data in the EU, less exposed to the US Cloud Act.
- Great for learning self-hosting and owning your own stack, closer to raw Linux.
- If you have no such requirements, MotherDuck makes DuckDB easier and more ergonomic (not affiliated).
04 / How
- A deployment script (github.com/berndsen-io/ducklake-hetzner) spins up the stack in about 3 minutes, testing included.
- You bring your own Hetzner account and API key; everything else is infrastructure and config as code.
- DuckDB 1.5 is the query engine and runs locally (not hosted yet).
- Two required backend components: PostgreSQL 16 as the DuckLake metadata catalog, and Hetzner Object Storage (S3) for data files.
- The PostgreSQL catalog runs on a single VPS on Ubuntu 24.04.
More detail
Architecture: DuckDB 1.5 reads/writes metadata to PostgreSQL 16 (metadata catalog) on a Hetzner VPS, and reads/writes data to Hetzner Object Storage / S3 (data files).
05 / Caveats
- Role-based access control is only partly possible; a separate ducklake-guard repo helps, but there is no true IAM-level control (Hetzner has no near-term plans to add it).
- Hetzner S3 is 'S3 compatible' but missing features and lacking AWS/Scaleway-level uptime guarantees; reputation on Reddit is mixed, though workable in practice.
- Price hikes in April and June raised the CX33 (4 vCPU, 8GB RAM) cost by 70%, from €4.99 to €8.49; still cheaper, but the gap is narrowing.
- Production needs extra steps: Floyd provides a production checklist to support that.
More detail
CX33 (4 vCPU, 8GB RAM) pricing: €4.99 before April 1st, €6.49 on April 1st, €8.49 on June 15th (+70%).
06 / Resources
- A ~10-minute blog article, 'DuckLake on Hetzner', covers the setup in depth (QR code on the slide).
- More content, repos and articles are available at Berndsen.io.
- Floyd invites feedback and change requests on the GitHub repos.
Key Takeaways
- You can run a complete DuckLake data lakehouse on Hetzner for under €15 a month, roughly a third of comparable AWS cost.
- DuckLake needs just two backend pieces: a PostgreSQL metadata catalog and S3-style object storage; DuckDB is the query engine.
- An infrastructure-as-code deployment script provisions the whole stack in about 3 minutes given a Hetzner account and API key.
- Hetzner is attractive for EU data residency and Cloud Act concerns since it is German-based.
- Trade-offs: limited RBAC/IAM, incomplete S3 compatibility, and weaker uptime guarantees than AWS/Scaleway.
- Recent April and June price hikes raised the reference VPS cost by 70%, so the price advantage is shrinking.
- If you lack strict cost or data-residency requirements, MotherDuck offers a more ergonomic managed alternative.