Cloud Parts

Databend

A cloud data warehouse written in Rust that keeps every table as Parquet blocks and Git-like snapshots in object storage, so stateless compute clusters can start, scale and stop without moving data. Open core: Apache 2.0, with paid features under the Elastic License 2.0.

What it is
Data warehouses,Real-time analytics databases
Type
Open source
Made by
Databend Labs
License
Apache 2.0 for the core; enterprise features (src/query/ee, src/meta/ee, BendSave) are under the Elastic License 2.0 and need a paid license key. Both ship in the same binary
Facts checked
11 October 2026
Website Source code

What it is

Databend splits into two programs. databend-meta is a Raft cluster that holds the catalog: table schemas and where their files live, users, roles and grants, and which query nodes belong to which cluster. databend-query nodes are stateless: they parse and plan SQL with a rule- and cost-based optimizer, then run it on a vectorized push-pull pipeline engine across one or more nodes. All table data sits in object storage (S3, Azure Blob Storage, Google Cloud Storage, Alibaba OSS or Tencent COS), which Databend reaches through Apache OpenDAL, a storage library that began inside Databend. The default storage engine, Fuse, works like Git: each write creates immutable Parquet blocks, groups them into segments with min/max and bloom filter indexes for pruning, and commits a new snapshot that points at the segments. Old snapshots stay readable, so SELECT … AT (SNAPSHOT => …), AT (TIMESTAMP => …) and UNDROP TABLE work within a retention period of 24 hours by default.

Clients connect with the BendSQL shell, drivers for Go, Python, Node.js, Java and Rust, or an HTTP API, and the server also listens on a MySQL-compatible port. Besides its own tables, Databend can query staged files (Parquet, CSV, TSV, NDJSON, ORC, Avro) without loading them, read and insert into Apache Iceberg tables through REST, AWS Glue, S3 Tables or Hive Metastore catalogs, and, in nightly builds since July 2026, read and write Apache Paimon tables. Its read-only Delta Lake engine was removed in September 2026. In 2026 Databend Labs began pitching it as a warehouse for AI agents, with vector and full-text search, geospatial functions and sandboxed Python UDFs.

Databend Labs builds it and sells two products on top: Databend Cloud, a managed service on AWS that bills warehouses per second ($1 an hour for an XSmall warehouse on the Personal edition) and storage at $23 per TB a month, and a self-hosted Enterprise edition. The license has been split since April 2023: features in the enterprise directories, such as masking policies, storage encryption, streams, Python UDFs, virtual columns and BendSave backups, are under the Elastic License 2.0 and need a license key. Versions are still numbered 1.2.x, and the project ships nightly builds about once a week (v1.2.952-nightly on 11 October 2026).

Why teams pick it

  • Compute and storage are fully separate: query nodes keep no data, so a cluster can be resized, replaced or suspended without rebalancing, and Databend Cloud suspends idle warehouses after five minutes and bills running ones per second
  • Snapshot-based Fuse tables give time travel by snapshot, timestamp or offset, UNDROP TABLE, and cheap pruning from segment and block statistics, all on plain object storage
  • One engine covers SQL analytics, semi-structured VARIANT data, vector and full-text search, geospatial queries, querying staged files in place, and Iceberg catalogs, so many teams need fewer separate systems
  • Easy to start: one Docker image runs meta and query together, and the Apache 2.0 core already includes distributed execution, transactions, time travel, RBAC and spilling of large queries

Watch out for

  • Open core with a hard line: masking policies, storage encryption, materialized views, streams and CDC, Python UDFs, virtual and computed columns, ATTACH TABLE, BendSave and automatic storage cleanup are Enterprise features under the Elastic License 2.0, which also forbids offering Databend as a managed service
  • Fast-moving and still numbered 1.2.x: the last GitHub release not marked pre-release is v1.2.881 (April 2026), and features come and go between builds; the Delta Lake engine was removed in September 2026 and experimental table branches were added in January 2026, removed in March and reintroduced in May
  • Every cache miss is an object-store request: Databend’s own engineers put an S3 GET at tens to hundreds of milliseconds, 1,000 to 5,000 times slower than local NVMe, so it suits large scans rather than point lookups or high-rate small writes
  • One company, Databend Labs, leads development and also sells the cloud service and the Enterprise license

Alternatives

What all data warehouses have in common →
ClickHouseOpen sourceAn open-source column-oriented database for real-time analytics, fast enough to serve dashboards over billions of rows.Google BigQueryGoogle CloudGoogle Cloud’s serverless data warehouse: no clusters, and you pay per TiB scanned or per slot-hour.MotherDuckCommercialA serverless cloud data warehouse built on DuckDB, where one query can run partly on your laptop and partly in the cloud.SnowflakeCommercialA fully managed cloud data warehouse that separates storage from compute, with virtual warehouses billed per second.

How it fits with other technologies

Snapshots, time travel, querying staged Parquet and UNDROP (run on the datafuselabs/databend Docker image, v1.2.881) · sql

-- docker run -p 8000:8000 datafuselabs/databend   then connect with bendsql
CREATE TABLE orders (id INT, amount DECIMAL(10, 2));
INSERT INTO orders VALUES (1, 9.50), (2, 20.00);
INSERT INTO orders VALUES (3, 4.25);

-- Each write committed a snapshot that points at segments of Parquet blocks
SELECT snapshot_id, segment_count, block_count, row_count
FROM fuse_snapshot('default', 'orders');
-- 01a12dfb461f75298c70a5fbdd9e37d4 | 2 | 2 | 3
-- 01a12dfb45f17d2ebec8d13e0995cf02 | 1 | 1 | 2

DELETE FROM orders WHERE id = 2;
SELECT count(*), sum(amount) FROM orders;   -- 2 | 13.75

-- Time travel: read the table as it was at an older snapshot
SELECT count(*), sum(amount)
FROM orders AT (SNAPSHOT => '01a12dfb461f75298c70a5fbdd9e37d4');   -- 3 | 33.75

-- Unload to Parquet in the user stage, then query the files without loading them
COPY INTO @~/export/ FROM orders FILE_FORMAT = (TYPE = PARQUET);   -- 2 rows unloaded
SELECT id, amount FROM @~/export/ (FILE_FORMAT => 'parquet') ORDER BY id;
-- 1 | 9.50
-- 3 | 4.25

DROP TABLE orders;
UNDROP TABLE orders;   -- within the retention period (24 hours by default)
SELECT count(*) FROM orders;   -- 2

Community, Enterprise and Cloud#

What each edition includes, from Databend’s editions and pricing pages, checked 2026-10-11.

CommunityEnterpriseDatabend Cloud
Who runs itYou: Docker, Kubernetes or bare metalYou, with Databend Labs supportDatabend Labs, on AWS
LicenseApache 2.0 core, plus free Elastic License 2.0 partsElastic License 2.0 features unlocked by a paid key (15-day trials on request)Commercial service
EngineDistributed SQL, transactions, time travel, RBAC, Iceberg and Hive queries, spillAdds materialized views, streams and CDC, Python UDFs, virtual and computed columnsAs Enterprise
Security and recoveryBasic audit, password policies, network allow listsAdds masking policies, storage encryption, full audit trail, BendSave backups and Fail-Safe recoveryAs Enterprise
PriceFree; you pay for servers and object storageQuotePer second of running warehouse ($1 to $512 an hour on Personal, 1.5 times that on Business) plus $23 per TB a month of storage

See it in a real stack

No architecture uses Databend yet. Drop it into the playground and wire it up yourself.Open playground →
Where it came from2020–2026

Databend began as FuseQuery: its first commit, “hello FuseQuery”, landed on 10 October 2020 from the GitHub user BohuTANG. Four days later the code moved from Apache 2.0 to the AGPL, and in February 2021 it went back to Apache 2.0. Databend Labs, the company behind it, was founded in 2021, and in September 2021 the project was renamed from Datafuse to Databend.

In December 2021 Xuanwo began rewriting Databend’s storage access layer as dal2. In February 2022 it left the repository as a standalone project called OpenDAL, which entered the Apache Incubator in 2023. Databend 1.0 shipped on 5 March 2023, the second anniversary of Databend Labs, adding UPDATE, ALTER TABLE, a DECIMAL type, a cost-based optimizer with histograms, querying files in stages and a query result cache. The release post described Databend as a cloud-native warehouse in Rust with storage-compute separation and an object-storage-oriented design, and said users were writing about 700 TB a day to object storage and analyzing it with Databend.

Weeks later, in April 2023, Databend added an enterprise license: code under src/query/ee and src/meta/ee became Elastic License 2.0 and needs a key, while the rest stayed Apache 2.0. In 2026 Databend Labs repositioned the product as a “data agent ready” warehouse built around vector and full-text search and sandboxed Python UDFs, and added geospatial functions and an R-tree index that February and March. Later that year it added Paimon read and write support and dropped the Delta Lake engine to shed a duplicate copy of Arrow and DataFusion from its dependencies.

Why it’s built this way

Separate storage from stateless compute
In a shared-nothing warehouse, adding nodes means reshuffling terabytes and a failed node takes its data with it. With data in S3, which became strongly consistent in 2020, compute nodes can be added, replaced or stopped without moving anything.
Git-like snapshots over immutable Parquet blocks
Object storage has no in-place updates or atomic rename, so each write adds new files and a new snapshot. That design makes time travel, UNDROP and pruning from per-segment statistics fall out of the storage format.
An asynchronous push-pull pipeline engine in Rust
An S3 read takes tens to hundreds of milliseconds, so an engine that blocks on I/O would sit idle. Databend overlaps I/O with computation and uses backpressure so fast and slow stages don’t overload or starve each other.
Elastic License 2.0 for paid features
Databend Labs’ FAQ says it found ELv2 “simple, short and clear” compared with the Business Source License, with three limits: no hosted service, no bypassing the license key, and keep the notices.

Timeline

  1. 2020

    First commit, “hello FuseQuery”, on 10 October

  2. 2021

    Back to Apache 2.0 in February; renamed from Datafuse to Databend in September

  3. 2022

    The storage layer leaves the repository in February as OpenDAL

  4. 2023

    Databend 1.0 on 5 March; Elastic License 2.0 enterprise features added in April

  5. 2026

    Paimon read and write support merged in July

  6. 2026

    Delta Lake table engine removed in September

Sources25 · facts checked 11 October 2026

Improve this page on GitHub