Databend Feb–Mar Bimonthly Report: Continuous Data Development to Enhanced GIS Capabilities

From February to March 2026, Databend continued advancing SQL capabilities, spatial analysis, file formats, query execution pipelines, and Meta architecture evolution. Sandbox UDF, Geography Functions, R-Tree Spatial Index, TEXT format, enhanced

EXPLAIN PERF

, Meta KV compression, and more — continuous data development delivery defined the two most notable months.

Hi, Databend community!

Since we didn't publish a standalone February report last month, this issue combines February 2026 and March 2026 into a single bimonthly recap, covering from

v1.2.874-nightly

v1.2.891-nightly

The main themes across these two months are clear: Sandbox UDF and Geography Functions expanded SQL expressiveness; R-Tree Spatial Index, Spatial Join, and spatial pruning pipelines were progressively filled in; CSV/TSV capabilities continued to improve and converged into the TEXT format in March; the query engine kept pushing forward with new join, spill fixes,

EXPLAIN PERF

bendsql

Monthly Stats

From

v1.2.874-nightly

v1.2.891-nightly

• 18 nightly versions
• 2 patch versions

Breakdown:

• February: 10 nightlies
• March: 8 nightlies and 2 patches

By release note counts, the Databend main repository merged across these two months:

• 43 new features
• 49 bug fixes
• 50 refactors
• 2 CI / build improvements
• 3 documentation updates
• 17 other optimizations

These numbers reflect Databend advancing on three parallel tracks in Feb–Mar: feature expansion, execution stability, and foundational restructuring.

Highlights

Core New Features

• Sandbox UDF + LATERAL

  generate_series
  

  — Databend continues advancing SQL expressiveness, enabling more flexible query composition and a safer foundation for custom logic execution.
• Geography Functions +

  st_hilbert
  

  — Spatial analysis functions keep expanding, from Geography Functions to

  st_hilbert
  

  .
• Object governance enhancements — Tagging support extended to

  USER
  

  ,

  ROLE
  

  ,

  STREAM
  

  ,

  view
  

  ,

  udf
  

  , and

  procedure
  

  ; experimental table tags support added for FUSE table snapshots.
• Data pipeline enhancements —

  COPY INTO location
  

  now supports

  PARTITION BY
  

  ;

  COPY INTO
  

  supports Lance dataset. Import/export pipelines continue to fill out.
• Table Branches with independent Schema — Branched data management is now more clearly organized. Table Branch is still being further optimized.
• Table Tag is included in the latest version and ready to use.

Spatial Analysis: From Geography Functions to Spatial Index

• Geography Functions — February laid the groundwork for geospatial computation, setting the stage for indexing and pruning.
• R-Tree Spatial Index — March formally introduced R-Tree-based Spatial Index, landing spatial indexing capabilities.
• Spatial Statistics — Added spatial statistics to

  BlockMeta
  

  for geospatial range pruning.
• Spatial Join — Runtime Filter now supports Spatial Index Join, integrating spatial indexes into the query optimization and execution path.

File Formats & Data Exchange: From CSV/TSV Enhancements to TEXT

• CSV / TSV compatibility improvements — Multi-byte field delimiters supported; complex types can be encoded as JSON for more flexible heterogeneous data ingestion.

• read_file
  

  +

  COPY INTO location PARTITION BY
  

  — File read/write pipelines continue to strengthen, making data engineering scenarios more complete.
• TSV → TEXT — March formally converged TSV into TEXT and added TEXT file format parameters, making text-based data import definitions cleaner.
• Detail fixes —

  infer_schema()
  

  now supports TSV;

  FIELD_DELIMITER=''
  

  supports full-line reads; CRLF handling continues to improve.

Query Performance & Execution Engine: From New Join to EXPLAIN PERF

• Experimental new join enabled by default — A significant milestone in the execution pipeline in February, followed by ongoing correctness fixes for spill scenarios.

• EXPLAIN PERF
  

  — March added per-plan hardware performance counters, further enhancing execution plan analysis.
• Sort / Spill optimizations — Spilled sort blocks prefetch, source sort key reuse, spilled sort streams compression, and batch rank-limit sort continue to advance.
• Join and filter optimizations — Small bloom index read optimization, proactive memory reclamation after Hash Join completion, and ongoing Runtime Filter enhancements.

Meta & Operations: Ongoing Foundational Restructuring

• Meta architecture cleanup — HTTP admin API extracted into a standalone crate

  databend-meta-admin
  

  ; Meta config structure, CLI config, and crate boundaries continue to be cleaned up.
• Meta KV Compression — Introduced transparent zstd compression and typed serialization in

  v1.2.883-nightly
  

  , further optimizing metadata storage.
• Query-Meta compatibility — Query-meta version compatibility documentation filled in, providing clearer boundaries for future evolution.
• Connection and operational details — Meta value compression toggle, gRPC listener

  TCP_NODELAY
  

  , and other capabilities continue to be added.

Stability & Quality: Execution Correctness and Edge Case Fixes

• Execution correctness — Ongoing fixes for new join spill data loss; recursive CTE concurrent result issues resolved.
• Storage and metadata stability — raft-log concurrent chunk read race condition fixed; unnecessary S3 refresh for HTTP catalog attached tables resolved.
• SQL edge case fixes —

  UNION
  

  parenthesis preservation, correlated scalar subquery decorrelation, Unicode identifier and alias support continue to improve.
• Security and governance boundaries — Row access policy execution issue under direct

  UPDATE
  

  resolved.

Practical Demo: Data Development and Delivery with Databend

Putting these two months of updates into a real-world example makes Databend's value more tangible. A common data development workflow looks like this: build cleansing, aggregation, and models inside Databend, then deliver results to object storage partitioned for downstream batch jobs; for specific analysis or application scenarios, export results into a more consumption-friendly dataset format. This maps directly to two practical capabilities shipped in Feb–Mar:

• COPY INTO <location> PARTITION BY
  

• COPY INTO <location> FILE_FORMAT = (TYPE = LANCE)
  

Scenario 1: Partitioned Parquet Delivery by Date and Hour

Suppose you've built an order wide table in Databend and want to deliver results daily to object storage partitioned by

date/hour

You can use:

COPY INTO @delivery_stage/orders/
FROM (
  SELECT
    order_id,
    user_id,
    amount,
    status,
    created_at,
    to_date(created_at) AS dt
  FROM mart_orders
  WHERE created_at >= '2026-03-01'
)
PARTITION BY (
  'date=' || to_varchar(dt, 'YYYY-MM-DD')
  || '/hour=' || lpad(to_varchar(date_part('hour', created_at)), 2, '0')
)
FILE_FORMAT = (TYPE = PARQUET);

The benefits are straightforward:

• Partition directories are determined directly by SQL expressions — delivery paths are fully controlled
• Results land in directory structures like

  date=2026-03-01/hour=08/
  

• Downstream jobs don't need an extra partition rewrite step

Scenario 2: Delivering Analysis Results as a Lance Dataset

If downstream consumption prioritizes efficient reads, columnar access, or a specific dataset organization, the Lance format support added in March is more valuable.

For example, to export a feature result set directly as a Lance dataset:

COPY INTO @delivery_stage/user_features_lance/
FROM (
  SELECT
    number
  FROM numbers(10)
)
FILE_FORMAT = (TYPE = LANCE)
USE_RAW_PATH = TRUE
OVERWRITE = TRUE;

A few important notes:

• A Lance dataset is a complete dataset directory, not a single file output
• By default, Databend generates a

  query_id
  

  subdirectory under the target path
• Use

  USE_RAW_PATH = TRUE
  

  to write results directly to the specified directory
• Lance currently does not support use with

  PARTITION BY
  

Databend's delivery pipeline offers two clear modes:

• For general data lake delivery:

  PARTITION BY + PARQUET
  

• For specific dataset organization or high-efficiency consumption:

  LANCE
  

Ecosystem Updates

bendsql: 2 Releases in March

bendsql

• v0.33.5
  

  , released 2026-03-11

• v0.33.6
  

  , released 2026-03-15

These updates focused on:

• More readable error messages, with

  reqwest::Error
  

  source and source chain output added
• Request stability improvements, including heartbeat headers fix,

  response.bytes()
  

  error retry, and timeout adjustments
• Release and build pipeline improvements, including OIDC trusted publishing and abi3 wheel build fixes

Both releases lean toward CLI usability, error diagnostics, and engineering quality improvements.

Drivers and Clients

Beyond

bendsql

• databend-go
  

  released v0.9.2 on 2026-03-23, continuing Go driver usability improvements

• databend-jdbc
  

  shipped a set of exception handling improvements around 2026-03-20, including embedding

  query_id
  

  in

  SQLException
  

  messages for easier troubleshooting

• databend-sqlalchemy
  

  continued compatibility fixes and version upgrades between 2026-02-25 and 2026-02-28, including

  varchar
  

  type parsing fixes and

  get_sequence_names
  

  issue corrections

These updates don't show up as headline features the way the main repository does, but they're equally important for Databend's usability at the application integration layer.