Duckdb Adbc

duckdb-adbc

The adbc extension lets DuckDB (v1.4.5+ or v1.5.4+) query any database that has an ADBC (Arrow Database Connectivity) driver — Snowflake, Databricks, BigQuery, PostgreSQL, MySQL, and more — directly from SQL, using zero-copy Arrow data transfer instead of legacy row-based APIs like ODBC/JDBC.

Repo: https://github.com/columnar-tech/duckdb-adbc-client

Prerequisites

1. Install the extension

From within a DuckDB session (or a .sql file / -c invocation):

INSTALL adbc FROM community;
LOAD adbc;

If DuckDB reports it can't find the extension, confirm the DuckDB version is 1.4.5+ (or 1.5.4+) — the extension requires a recent build.

2. Install dbc

ADBC drivers themselves are installed and managed with dbc, a small CLI from Columnar. Check whether it's already available with dbc --help; if not, install it with whichever of these fits the environment:

# shell (macOS/Linux)
curl -LsSf https://dbc.columnar.tech/install.sh | sh

# Homebrew
brew install columnar-tech/tap/dbc

# uv
uv tool install dbc

# pipx
pipx install dbc

# PowerShell (Windows)
powershell -ExecutionPolicy ByPass -c irm https://dbc.columnar.tech/install.ps1 | iex

# winget (Windows)
winget install dbc

3. Install an ADBC driver for the target database

dbc install sqlite
dbc install postgresql
dbc install snowflake
# ...etc.

If a query fails with a driver-not-found error, install the appropriate driver with dbc install <system> before retrying.

Connecting to a database

The extension connects via ADBC connection profiles — TOML files describing a driver and its options — referenced with a profile:// URI.

Create a profile

# mydb.toml
profile_version = 1
driver = "sqlite"

[Options]
uri = "./games.sqlite"

For remote/authenticated systems, add the relevant options, e.g.:

profile_version = 1
driver = "postgresql"

[Options]
uri = "postgresql://user:pwd@localhost:5432/app"

Save the profile to a discoverable location

# Linux
mv mydb.toml ~/.config/adbc/profiles/

# macOS
mv mydb.toml ~/Library/Application Support/ADBC/Profiles/

# Windows
move "mydb.toml" "%LOCALAPPDATA%\ADBC\Profiles\"

Once saved, reference the profile as profile://mydb (filename minus .toml) from SQL, or point directly at a path.

Prefer profiles over embedding credentials in ad-hoc SQL — they keep secrets out of query text and can be reused across sessions.

Common workflows

Ad-hoc read via read_adbc

Best for one-off queries, or when pushing a projection/predicate down to the remote system (the extension does not automatically push down projections/predicates for attached tables — use read_adbc directly for that).

INSTALL adbc FROM community;
LOAD adbc;

SELECT * FROM read_adbc('profile://mydb', 'SELECT * FROM games');

Persistent connection via ATTACH

Best when the remote database should behave like a local DuckDB catalog — supports catalog lookups, SELECT, INSERT, COPY, and CREATE TABLE AS (CTAS).

ATTACH 'profile://mydb' AS mydb (TYPE adbc);
USE mydb.main;

SHOW ALL TABLES;
SELECT * FROM games;
INSERT INTO games (SELECT 6, 'Battleship', 'Clifford Von Wickler', 1931, 7, 2, 2, 12.99);

-- Move data into or out of local DuckDB tables
CREATE TABLE memory.inventors AS (SELECT id, inventor FROM games);
CREATE TABLE game_inventors(id, inventor) AS (SELECT * FROM memory.inventors);

For systems with non-default schema/table delimiters (e.g. SQL Server's [schema].[table]):

ATTACH 'profile://mydb' AS mydb (TYPE adbc, DELIMITER '[]');

Arbitrary DDL/DML via adbc_execute

For statements that don't fit the SELECT/INSERT/COPY/CTAS surface (e.g. DROP TABLE, CREATE INDEX):

CALL adbc_execute('profile://mydb', 'DROP TABLE games');

Refreshing cached metadata

DuckDB caches schema/table metadata from attached ADBC databases locally. After a remote schema change, clear it:

CALL adbc_clear_cache();

Mixing ADBC reads and writes in one statement

Disabled by default to avoid concurrency bugs. To allow it:

SET adbc_mix_reads_writes = true;

To materialize input rows before INSERT/CTAS (also helps avoid concurrency issues when mixing reads/writes):

SET adbc_materialize_insert_rows = true;

Tuning connection pooling and insert batching

-- Max pooled connections per attached database (default 50)
SET adbc_connection_pool_size = 100;

-- Chunks buffered in memory before an ADBC bulk insert (default 1000, i.e. ~2M rows @ 2048 rows/chunk)
SET adbc_insert_buffer_size = 10000;

Limitations

• Autocommit only. Every statement takes effect immediately; there is no manual transaction control.
• No automatic pushdown. Predicate/projection pushdown isn't performed for attached (ATTACH) tables — use read_adbc with a scoped SQL query when pushdown matters.
• No DuckDB-to-DuckDB. The extension doesn't currently support connecting to another DuckDB (or Quack) instance via ADBC.
• No intra-process concurrency. Concurrent ADBC operations within a single process aren't supported.
• Mixed reads/writes need an opt-in. See adbc_mix_reads_writes above.

Troubleshooting

• Extension won't load / INSTALL adbc FROM community fails. Confirm the DuckDB version is 1.4.5+ or 1.5.4+.
• Driver not found / failed to load. Install it with dbc install <system>; verify the driver value in the profile matches the name dbc uses.
• Profile not found (profile://name fails). Confirm the .toml file is saved in the correct per-OS ADBC profile directory, or reference it by absolute path instead of name.
• Authentication errors. Double-check the uri and any credential fields under [Options] in the profile. For databases needing certificates, tokens, or extra parameters, consult the driver's own documentation via dbc.
• Unexpected "Not implemented" error on INSERT. Likely a mixed read/write statement — set adbc_mix_reads_writes = true (and consider adbc_materialize_insert_rows = true).
• Stale results after an external schema change. Run CALL adbc_clear_cache();.