ClickHouse

Overview

ClickHouse is a column-oriented database for analytics, and the thing that sets it apart from the other warehouses here is that the engine itself is open source under Apache 2.0. You can run it yourself with no licence fee, or buy it as a managed cloud service, and the same engine is underneath both.

Its reputation is speed. Values for each column are stored sequentially, so analytical queries read only the columns they need, and the documentation's own example processes 100 million rows in 92 milliseconds.

The self-hosted shape is the classic one: MergeTree-family tables on local disks with asynchronous multi-master replication, which means you size and shard the cluster yourself.

ClickHouse Cloud changes that architecture. Storage moves to object storage, separated from compute, giving effectively unlimited capacity without manual sharding, with automatic scaling and idling to zero when quiet.

Cloud also adds compute-compute separation, so read and write workloads run on independent compute groups over the same data, which is the pattern used to stop ingestion and querying competing.

Concurrency is worth checking against your needs: queries per second are unlimited but each replica caps at 1,000 concurrent queries, so you scale concurrency by adding replicas.

For lake work it reads Iceberg, Delta Lake, Hudi and Paimon through table functions, with write-back documented, and can surface external catalogues, including Glue, Unity Catalog, Polaris and OneLake, as databases.

Cloud tiers gate more than support. Basic cannot scale at all, private networking starts at Scale, and HIPAA and PCI are Enterprise-only, with HIPAA needing a business associate agreement.

The AI direction is agents: an Ask AI assistant in the console, a remote MCP server so external agents can query through scoped read-only access, plus HNSW vector indexes in the engine itself.

Key features and capabilities

The same headings are used for every cloud data warehouses entry, so two tools can be read side by side.

Architecture and storage

• Self-hosted, this is MergeTree-family columnar storage on local disks with asynchronous multi-master replication
• In Cloud, storage sits on object storage and is separated from compute, with automatic scaling and idling
• Compute-compute separation runs read and write workloads on independent compute over shared data
• Cloud isolates tenants at the network, compute and storage layers

Performance and scaling

• Vertical autoscaling on Scale and Enterprise; Basic cannot scale and horizontal scaling is manual
• 1,000 concurrent queries per replica, raised by adding replicas
• Approximate calculation functions trade precision for speed
• Materialised views and projections are core engine features

SQL and workloads

• SQL with GROUP BY, ORDER BY, joins and window functions
• Table functions query lake data in place, including iceberg, deltaLake, hudi and paimon
• A DataLakeCatalog database engine surfaces external catalogues as ClickHouse databases
• Kafka ingestion and ClickPipes are available, ClickPipes priced separately

Open formats and sharing

• The engine is Apache 2.0, so self-hosting is a first-class option rather than a downgrade
• Reads Iceberg, Delta Lake, Hudi and Paimon, with write-back documented
• Catalogues include AWS Glue, Unity Catalog, Iceberg REST, Polaris, Lakekeeper, BigLake and OneLake
• Cloud egress is charged, around $0.115 a GB to the internet from AWS US East

Security and governance

• Role-based access control with users, roles, row policies, settings profiles and quotas
• Row policies filter which rows a user or role can see
• SOC 2 and ISO 27001 on all Cloud tiers; HIPAA and PCI on Enterprise only
• Private networking from Scale upward, with customer-managed keys and SAML on Enterprise
• Column-level grants are not documented on the access-rights page

AI and machine learning

• Vector similarity indexes in the engine, HNSW only, with quantisation and pre- or post-filtering
• An Ask AI agent in the Cloud console turns questions into queries, charts and summaries
• A remote MCP server lets external agents inspect schemas and run scoped read-only queries
• ClickHouse Agents, built on Claude, are in public beta alongside AI notebooks

Where it runs and what it costs

• Self-hosted anywhere under Apache 2.0, with no licence fee
• Cloud on AWS, Google Cloud and Azure across 39 regions, some on request
• Bring your own cloud is available on a custom quote
• Metered per minute in 8 GB memory increments, plus storage

Pricing

Usage-basedFree self-hosted; Cloud from about $0.39 a compute unit

Self-hosting is free under Apache 2.0. Cloud is metered per minute, one compute unit being 8 GiB of memory and 2 vCPU. In AWS US East the published Enterprise rate is $0.39 per compute-unit hour with storage at $25.30 a terabyte a month; Basic and Scale unit rates are not published. The billing docs work examples through, a Basic single replica from about $40 to $160 a month and a two-replica Scale service at $437. Egress is extra. ClickPipes adds $0.20 a unit hour plus $0.04 a GB. A trial gives $300 of credit over 30 days with no card.

Vendor pricing page →

Demos and videos

• How to get started with ClickHouse Tutorial · YouTube
• An intro to ClickHouse Cloud Product demo · YouTube

About ClickHouse

The open-source project started in 2009 and was released under Apache 2.0 in 2016; ClickHouse, Inc. was incorporated in Delaware in 2021 to commercialise it, co-founded by Aaron Katz, Alexey Milovidov and Yury Izrailevsky. It is private and heavily venture-funded, most recently a $400m Series D led by Dragoneer announced in January 2026, alongside the acquisition of Langfuse and the launch of a native Postgres service. The company reports more than 4,000 customers and over $250m in annual recurring revenue. The split is clean: the database is Apache 2.0, and ClickHouse Cloud is the commercial managed service.

Founded 2021 · San Francisco Bay Area, California · clickhouse.com