The flexibility of the cloud, without its downsides

Your data warehouse.
Your data center.

An Iceberg lakehouse, a semantic layer, hybrid search and an MCP server through which AI agents query your data — under the same permissions as your analysts. Query elastically, store openly, operate it yourself. No credit burn, no egress fees, no data leaving the building.

Book a live demo Calculate TCO

Runs on Kubernetes via Helm chart or as a Docker Compose stack · S3-compatible storage · local embedding model, no cloud key

€0: for credits, egress and data sharing
1: installation for lakehouse, governance, search and AI
MCP: agents query your data under your permissions
100%: your hardware, your network, your keys

The platform

Four building blocks you would otherwise buy separately — in one installation

Building an open lakehouse yourself means Iceberg, a catalog, a query engine, an identity layer, a permission model and an interface — from different projects, each with its own operational load. lavalake ships those layers as one product, behind a single reverse proxy, with one permission model.

Getting data in

Four ways into the lakehouse

Databases, files, documents and MCP servers — each with an assistant, not a config file. Everything loads into Apache Iceberg, an open format.

PostgreSQL, MySQL and MS SQL Server: read-only, credentials AES-256-GCM encrypted, TLS per source
CSV and file import with an assistant for types, renames, null tokens and dedup
Documents (PDF, Word, text) are chunked, embedded and searchable — parsing in its own process
Fileshare crawler: point at a directory instead of uploading
Ingestion from registered MCP servers, SAP-MCP for instance
Change data capture and continuous ingestionIn development

Governance

Nobody gets around the permissions

There is no service account that may do anything. Every statement runs under the exchanged token of the signed-in user — the AI agent's included. What someone may not see, they do not see in a search result either.

Delegated identity through Keycloak Standard Token Exchange (RFC 8693)
Row-access and column-masking policies enforced in Trino, not in the interface
WORM audit log as a hash chain: after-the-fact edits are detectable
Lineage and access history per table — where the data came from, who read it
Search results are re-fetched under your token, so policies apply there too

MCP & AI

The agent gets your rights, not everyone's

An MCP server is part of the platform. The built-in assistant works with the same delegated rights you have, through curated tools behind a gate → audit → run chain. Mutating and destructive steps ask first.

Tools for catalog, schema, queries, the semantic layer, dashboards and roles
Every statement is classified as read, mutating or destructive and cleared accordingly
The model is your choice: own API key, self-hosted model, or per-user Claude sign-in
The assistant can build the semantic model out of your data itself
Natural language to SQL against curated metric definitionsIn development

Analysis

Search it, query it, put it on a tile

The semantic layer, hybrid search and dashboards are part of the platform. The embedding model runs locally on CPU, which is why semantic search needs no cloud key.

Hybrid search: meaning and exact words, fused into one ranking
Local embedding model (nomic-embed-text) — no key, no request leaving the building
Semantic layer: curated metrics and multi-table semantic views
SQL worksheet over Trino, data browser from catalog to table
Dashboards from queries and metrics — each tile computes per viewer under their identity

Architecture

Storage and compute separated — except behind your firewall

lavalake adopts the architectural principle of modern cloud warehouses: data sits as open Iceberg tables on S3-compatible storage, and the query engine reaches it independently. The difference: both stand in your data center, and your tables stay readable in an open format — with Spark, Flink or Trino too.

Access

Console (SPA) · SQL worksheet · dashboards · AI assistant · MCP server — all behind one reverse proxy

Identity & governance

Keycloak OIDC · token exchange per RFC 8693 · row-access and masking policies · WORM audit · lineage

Compute

Trino, every statement under the caller's token · elastic warehouses with auto-suspendIn development

Catalog & table format

Apache Iceberg with Nessie: ACID · time travel · branching · schema evolution · open to Spark, Flink, Trino

Storage & index

S3-compatible (MinIO, Ceph, NetApp StorageGRID, Dell ECS) · PostgreSQL with pgvector as the vector index

A Docker Compose stack on one machine for evaluation, a Helm chart for Kubernetes in production. High availability without single points of failure, and operation without an internet connection, are on the roadmap.

If you already know a cloud warehouse

The concepts stay, the bill changes. What is still missing says so.

Proprietary table format: Apache Iceberg with Nessie (open, versioned)
Web interface for queries: Console with SQL worksheet and data browser
Roles and row filters: Policies in Trino, under delegated identity
Access history and provenance: WORM audit log, lineage and access history
Full-text and vector search: Hybrid search with a local embedding model
The vendor's AI features: MCP server · the model of your choice
Virtual warehouse: Elastic compute clustersIn development
Streams and scheduled tasks: Continuous ingest, CDC and a schedulerIn development
Credit billing: A fixed subscription

In practice

What the platform actually does today

Attaching a source

Databases are registered in the console: type, host, credentials. Credentials are stored AES-256-GCM encrypted, access is read-only, and TLS can be enforced per source. You then pick the tables to bring over as Iceberg.

Read the docs

Console · Data sources

Register a source
  Type        postgres | mysql | mssql
  Host        erp-db.dc.internal:5432
  Database    sales
  Login       reader / ********     (stored encrypted)
  TLS         enforced

Bring tables over
  sales.public.orders     ->  lakehouse.raw.orders
  sales.public.line_items ->  lakehouse.raw.line_items

Other paths: CSV import with a modeling assistant,
documents (PDF/Word/text), fileshare crawler, MCP ingestion.

Cost

The same analytics. A fraction of the invoice.

Cloud warehouses bill by consumption — which is exactly what makes data work expensive as soon as it succeeds. lavalake inverts the model: you pay for the platform, not for every query. Move the slider to your current monthly spend.

Your current cloud warehouse cost / month€45,000

€8k€250k

No credits

A fixed subscription per cluster instead of consumption units. A badly written query costs compute time, but not budget.

No egress

Data never leaves your network, so transfer charges for exports, replication and data sharing between sites simply disappear.

Hardware you already own

lavalake runs on your existing Kubernetes and your S3-compatible storage. Capacity you have already written off keeps earning.

Planning you can budget

Cost follows cluster size and term, not usage behavior. Your teams are free to explore the data again.

Monthly comparison

60% cheaper

Cloud warehouse today€45,000

lavalake on-premises€17,900

lavalake subscription: €9,900
Hardware amortization: €4,500
Operations & administration: €3,500
Egress and credit charges: €0

Savings over 3 years

€975,600

A model calculation based on the values you set. For a defensible figure we will work through your actual workload.

Request a TCO analysis

Where this comes from

The calculation above is a model built on your numbers. Two independent sources suggest the order of magnitude holds:

roughly half
Andreessen Horowitz estimates that repatriating cloud workloads cuts total cost of ownership to about half. An estimate rather than a survey, and calculated for software companies at scale.
a16z, The Cost of Cloud, a Trillion Dollar Paradox, 2021
$2 million a year
37signals published the breakdown of its cloud exit: $3.2 million in annual spend, then roughly $2 million less per year after $700K of hardware. A single case with real invoices — and not a data warehouse.
37signals, Cloud Exit, 2024–2025

Sovereignty

The reason many projects are never allowed into the cloud

Production data, patient records, design files, case files: in many organizations the modern warehouse does not fail on technology, it fails on sign-off. lavalake solves that at the root — the platform comes to the data, not the other way around.

Data sovereignty without the fine print

Data never leaves your data center — not even for telemetry
The vendor has no access to content or metadata
The embedding model runs locally; semantic search needs no cloud key
Operation without an internet connection, with a self-hosted language model as a required optionIn development

Provable, not asserted

WORM audit log as a hash chain, over users and agents alike
Lineage from source to reported figure, access history per table
Every query under the identity of whoever is signed in — no shared service account
Access, deletion, retention and classification as platform featuresIn development

Operations

One Helm chart for Kubernetes, one Docker Compose stack for evaluation
Everything behind one reverse proxy; Trino and storage are unreachable from outside
Egress guard against SSRF, document parsing in its own process
High availability, metrics, traces and alerting; secrets through VaultIn development

This is not a niche

85%: consider Germany too dependent on US cloud providers
71%: buy cloud from the US — only 8% would choose to
91%: would prefer German providers; 53% actually use one

Bitkom Cloud Report 2026, representative survey of 603 German companies with 20+ employees, June 2026

GDPR through location control
Delegated identity (RFC 8693)
WORM audit
Open table format
Kubernetes-native
Open source at the core

Use cases

Built for data that is not allowed to leave the building

Request a reference architecture

Manufacturing

Machine data and ERP in one analysis

Production telemetry meets SAP orders — without mirroring sensor data into a public cloud. Scrap analysis runs on the same cluster as financial controlling.

18 billion measurements / month

Financial services

Risk calculation with auditable provenance

Regulatory reporting and risk models read from versioned Iceberg tables. Time travel reproduces any reporting date, and lineage answers the supervisor's question.

Any reporting date in seconds

Healthcare

Clinical analysis inside your own network

Treatment and billing data stay in-house, with analysis and AI assistance running on top. Roles and masking make sure each department sees only its own slice.

Masking applies in the engine, not in the interface

Retail

One metric instead of four exports

Store data comes into the lakehouse from the upstream systems, and replenishment and marketing compute on the same curated metrics of the semantic layer instead of each on its own extract.

One metric definition for every report

Public sector

Making case management systems analyzable

Data from separate case management systems is brought together without leaving the legal framework. Operated in the state data center, analyzed by self-service.

Operated in the state data center

AI applications

Agents on company data

Through MCP, assistants query reviewed metrics rather than improvising over raw tables. Answers are traceable and every request is logged.

Every agent query in the audit log

Migration

Four weeks, and your cloud subscription is last winter's problem

Migration here does not mean rebuilding. Tables come over as Iceberg, roles and permissions move into Keycloak and the Trino policies. Both systems run in parallel until the results reconcile.

Week 1
Taking stock
We read your existing warehouse metadata: tables, views, roles, query history and cost distribution. The result is a migration plan with a defensible TCO calculation.
Week 2
Installation
lavalake is installed on your cluster via Helm chart and wired to your S3-compatible storage, your AD and your monitoring. No new hardware purchase required.
Week 3
Data and pipelines
Tables are taken over as Iceberg and queries are re-pointed at Trino. Both systems run in parallel until the results match row for row.
Week 4
Cutover
BI connections point at lavalake, roles and row filters are in place, reports are validated against the old figures. Then the cloud subscription gets canceled.

Pricing

A subscription, not a meter

The price follows cluster size and support tier — not the number of your queries, users or transferred gigabytes.

Core

free

One node, no time limit

For evaluation, development and small departmental warehouses. The engine's full feature set.

Iceberg lakehouse on your storage
SQL worksheet, import assistant and dashboards
MCP server included
Community support

Download

Team

€249

per month, up to 3 nodes

For a department's first production warehouse: scheduled pipelines, roles and support by ticket.

Up to 3 nodes and 10 users
Database sources, documents and the fileshare crawler
User and role management through Keycloak
Ticket support, answer within a business day

Start a free trial

Business

€749

per month, up to 8 nodes

For several departments on one platform: SSO, fine-grained permissions and committed response times.

Up to 8 nodes, unlimited users
Single sign-on over OIDC
Row-access and column-masking policies
WORM audit, lineage and access history
8/5 support with an SLA

Request a quote

Recommended

Enterprise

€2,499

per month, per cluster

For production use across several departments: round-the-clock support, migration help, and the enterprise items on the roadmap first.

Unlimited nodes
24/7 support with a European SLA
Migration support included
Elastic compute with auto-suspendIn development
High availability and backup orchestrationIn development

Request a quote

Sovereign

Custom

Regulated environments

For organizations under special obligations: separated operation, offline updates, source code escrow and on-site support.

Operation without an internet connectionIn development
Secrets through Vault or External SecretsIn development
Source code escrow
Named contact, on-site engagements
Help with your certifications

Schedule a call

All prices net, excluding VAT. There are no hosting costs — the platform runs on your infrastructure.

Let's run the numbers on your workload.

In 30 minutes you will see your architecture, your expected cost and a realistic migration plan. After that you decide in your own time — no sales pressure.

Technical demo on your own example queries
TCO comparison based on your query history
Proof of concept in your environment within two weeks

Your data warehouse.Your data center.