TL;DR: Today we are releasing DuckDB 2.0.1, the first patch release of our 2.0 LTS line. This release ships bugfixes, performance improvements, and security patches. In addition, a capability is added to define custom extension repositories with only their hostname. Today we're releasing DuckDB 2.0.1, the first patch release in the 2.0 LTS (Cyanoptera) line. This planned release arrives one month after 2.0.0 and contains bugfixes, performance improvements, and security patches. In addition, it provides a more convenient way of registering custom extension repositories. You can obtain a copy from the installation page. For a detailed overview of the changes, see the release notes on GitHub. Read on for some highlights. Changes That May Require Action A few fixes in this release change existing behavior. In each case, the previous behavior was incorrect or a security risk, so we chose to fix it in 2.0.1 rather than wait for the next major release. If you rely on any of the behaviors below, review them before upgrading. Stricter SECRET Scope Matching A SECRET scoped to a host or bucket could also match other hosts or buckets with the same prefix, so it could be sent somewhere it wasn't meant for (#28684). A URL now matches such a scope only if it is identical to the scope or continues with a /. For example, the scope s3://bucket- no longer matches s3://bucket-1/. Action: check your host- and bucket-scoped SECRETs and make sure each scope uses the full host or bucket name. If a URL has an explicit port, the scope must include it too. UPDATE ... FROM and MERGE Again Apply Every Match In DuckDB 2.0.0, if several source rows matched the same target row, MERGE raised an error and UPDATE ... FROM updated the target row once, using one of the matches. Enforcing this behavior serialized large updates, causing a significant performance regression. DuckDB 2.0.1 restores the 1.x behavior for both statements (#28701): the target row is updated once per match, so the final value can be non-deterministic, and the row is counted and returned (with RETURNING) once per match. Action: review your UPDATE ... FROM and MERGE statements. To keep updates deterministic, make sure each target row is matched by at most one source row, or that all matching source rows have identical values for the updated columns. We expect a fix that keeps both correctness and performance in DuckDB 2.1. Improvement: Registering Extension Repositories by Host DuckDB 2.0.0 introduced custom extension repositories, which host extensions signed with the publisher's own keys. In DuckDB 2.0.1, you can register such a repository by naming the publisher's host (#28719). For example, you can register duckdb.org as a repository: CREATE EXTENSION REPOSITORY duckdb.org; Then, core extensions are reachable through it: INSTALL AND LOAD spatial FROM duckdb.org; Under the hood, the host acts as an extra redirect. DuckDB first reads https://duckdb.org/.well-known/duckdb-extension-repo.json over verified HTTPS. This file contains no keys or extensions: it only states where the host's repository lives, in this case, https://extensions.duckdb.org. DuckDB then follows this pointer, reads the repository's own .well-known/duckdb-extension-repo.json file, and pins the signing keys listed there. Only the owner of a host can publish files under its .well-known path, so serving this file is proof of control over the host, and no central registry is involved. The redirect also decouples the repository's name from its location: the name can be your organization's domain, while the extensions are served from a CDN, an S3 bucket, or GitHub Pages. The CREATE EXTENSION REPOSITORY statement returns the SHA-256 fingerprint of each pinned key, so you can compare them against values published through another channel. The power of this feature lies in the fact that your organization can do the same: publish a .well-known/duckdb-extension-repo.json file on your own domain, and your users can install your extensions with the same succinct syntax: INSTALL AND LOAD extension FROM your_domain; To get started, see the new “Publishing an Extension Repository” page. Unlike adding a repository by its prefix, adding one by its host does not require an opt-in. To support this, the allow_extension_repositories setting has new values: hosts (the new default) allows adding repositories by host, allowed also allows adding them by prefix, hosts_only only adds and trusts repositories resolved from their host, and forbidden disables custom repositories. The previous default value, undecided, is no longer accepted. If only extensions signed by the DuckDB team should be loaded in your environment, set allow_extension_repositories to forbidden. For details, see the “Securing Extensions” page. Important Fixes Silent Data Loss on Affected Linux Kernels Some Linux kernel versions have a bug (CVE-2026-97945) that can cause memory freed lazily with MADV_FREE to be silently zeroed after it has been reused. DuckDB 2.0.1 avoids this issue: its bundled allocator no longer frees memory this way (#28267). The patch works by setting the bundled jemalloc's muzzy_decay_ms to 0. If you cannot upgrade right away, you should set the DUCKDB_JE_MALLOC_CONF environment variable to muzzy_decay_ms:0, which achieves the same effect. Reading Databases Written by DuckDB 1.x DuckDB 2.0.0 could miss rows with NULL values when filtering tables in database files written by DuckDB 1.x (#28355). DuckDB 2.0.1 fixes this, along with several issues affecting views, macros, and other objects stored by 1.x versions. Security Fixes Configuring enable_external_access=false should prevent access to external state but would still allow CREATE SECRET, which can result in accessing external state. This is now fixed (28637) by refusing CREATE SECRET with a Permission Error and an error message indicating Creating secrets is disabled when enable_external_access is false. We also fixed a remote path traversal issue (#28684) that occurred when DuckDB was configured with enable_external_access=false. Additionally, this release pulls in a fix (#27791) for the Brotli decoder (CVE-2026-111497). Other Fixes Here are some other important fixes from the DuckDB 2.0.1 release, organized by category: Correctness #27514 – Fix IS NULL pruning of Parquet pages with constant string statistics #27923 – Fix hour, minute, and second rounding TIMESTAMP_NS values up into the next second #27779 – Fix LIMIT being dropped above a join with an aggregate over an empty table #28698 – Fix wrong results when a lateral join reads from a CTE #28210 – Fix random functions returning the same values in processes cloned with fork() Storage and Transactions #27764 – Keep the rows of bulk inserts invisible to other transactions until commit #26814 – Fix lost index entries when an automatic checkpoint runs after WAL replay #28190 – Fix rows getting lost after a restart that follows a reverted commit Crashes and Internal Errors #28293 – Fix crash in list_select and list_where on empty lists #27596 – Fix window operators hanging when they hit the memory limit #27936 – Fix deadlocks in materialized CTEs Performance #28646 – Fix aggregates such as quantile_cont and arg_max running out of memory on large GROUP BY queries #27760 – Push down filters to VIEWs used in subqueries, which can make queries orders of magnitude faster #27326 – Use the Parquet page index to skip pages, e.g., in files written by Spark #28193 – Improve join ordering for range joins Miscellaneous #27685 – Add the RETURN_COLUMN_TYPES option to COPY ... TO #27969 – Add the parquet_can_have_nan setting for Parquet files whose statistics do not account for NaN values #28011 – Ship the Windows CLI as ZIP archives for WinGet Extensions and Clients As is customary for DuckDB releases, while DuckDB 2.0.1 itself is a patch release, extensions and clients ship new features along with their fixes. Here are some highlights: The azure extension no longer deletes live data files stored at abfss://container@account paths when running CHECKPOINT on DuckLake (duckdb-azure#211). The experimental glue extension attaches an AWS Glue Data Catalog as a DuckDB database, so you can read and write its Hive tables stored as Parquet, CSV, JSON, or Avro files on S3. The iceberg extension ships many correctness and performance fixes. For example, TLS now stays enabled when using vended path-style S3 credentials (duckdb-iceberg#1527). Note that iceberg_rewrite_data_files now returns Spark's output column names (duckdb-iceberg#1597). pg_duckdb, the PostgreSQL extension that embeds DuckDB, now publishes binaries for all supported PostgreSQL versions on its GitHub releases page. It also ships a number of bugfixes and improves compatibility with other popular PostgreSQL extensions such as Citus and pgvector. In the postgres extension, queries that only touch a single attached PostgreSQL database are now executed entirely in PostgreSQL. If you would like to disable this, run SET pg_enable_remote_pushdown = false before attaching the database (duckdb-postgres#643). The Python client received fixes for several hangs and crashes at exit, including on Python 3.14 (duckdb-python#709, duckdb-python#719). Fetching TIMESTAMPTZ values no longer requires the pytz package (duckdb-python#710). The aws and glue extensions are now available in DuckDB-Wasm, and the vortex extension is now available on Windows. The wasm_mvp platform was discontinued. Core extensions are no longer built for this platform. The remaning Wasm platforms are wasm_eh for slim deployments and wasm_threads for enabling threads and SIMD operations. End of Life for DuckDB 1.4 LTS Community support for DuckDB 1.4 LTS (Andium) ended on November 17, 2026. If you are still using it, we recommend upgrading to DuckDB 2.0 LTS, which is supported until at least October 2027. For details, see the release calendar. Conclusion This post was a short summary of the changes in v2.0.1. As usual, you can find the full release notes on GitHub. We would like to thank our contributors for providing detailed issue reports and patches. We are also grateful to the following people whose reports helped us make DuckDB and its extensions more robust and secure: @TristanInSec for reporting out-of-bounds reads and a string-length buffer sizing overflow in DICT_FSST decompression, and for reporting missing validation of Arrow string offsets. @zerovulnlabs for reporting missing page-size validation when reading uncompressed Parquet V2 pages. zz lin for reporting a problem in the MySQL extension. Stay tuned for future DuckDB releases, including DuckDB 2.1.0 in the first half of 2027! Changes That May Require Action Stricter SECRET Scope Matching UPDATE ... FROM and MERGE Again Apply Every Match Improvement: Registering Extension Repositories by Host Important Fixes Silent Data Loss on Affected Linux Kernels Reading Databases Written by DuckDB 1.x Security Fixes Other Fixes Correctness Storage and Transactions Crashes and Internal Errors Performance Miscellaneous Extensions and Clients End of Life for DuckDB 1.4 LTS Conclusion Recent posts DuckDB User Group Workshop on AI-Powered Review Analysis with Hugging Face Local Analytics Doesn’t Need a Database Server Ducks in a Column
AI NEWS JOURNAL
DuckDBAI NEWS / JOURNAL
Announcing DuckDB 2.0.1 LTS
Today we are releasing DuckDB 2.0.1, the first patch release of our 2.0 LTS line. This release ships bugfixes, performance improvements, and security patches. In addition, a capability is added to define custom extension repositories with only their hostname.