TopolearnExtensions let DuckDB query data where it lives, which removes whole export and import steps from pipelines.
Much of DuckDB's functionality ships as extensions: INSTALL name; LOAD name;. Core extensions are signed, downloaded from the official repository and cached locally; many (like httpfs, json, parquet) autoload the first time you use a feature that needs them. SELECT * FROM duckdb_extensions() lists what is installed and loaded.
The most used ones:
httpfs: read and writehttps://,s3://,gcs://and Azure paths. Configure credentials with the secrets manager:CREATE SECRET (TYPE s3, PROVIDER credential_chain)picks up your AWS environment, or giveKEY_ID/SECRET/REGIONexplicitly. AddPERSISTENTto store it across sessions.postgres,mysql,sqlite:ATTACH 'dbname=shop host=...' AS pg (TYPE postgres)exposes the remote tables; you can join them with local Parquet and even write back.iceberganddelta: read lakehouse table formats.spatial(geometry types and GIS functions),fts(full-text search),excel,vss(vector similarity index).
Community extensions are installed with INSTALL name FROM community. Because extensions are native code, only load ones you trust; settings like allow_community_extensions and autoinstall_known_extensions let you lock this down in production.
Example
-- read Parquet straight from S3 using your AWS credential chain
INSTALL httpfs; LOAD httpfs;
CREATE SECRET aws (TYPE s3, PROVIDER credential_chain);
SELECT count(*) FROM 's3://my-bucket/events/2025/*.parquet';
-- join a live PostgreSQL table with a local file
INSTALL postgres; LOAD postgres;
ATTACH 'dbname=shop host=localhost user=app' AS pg (TYPE postgres, READ_ONLY);
SELECT c.country, sum(e.amount) AS total
FROM 'events.parquet' e
JOIN pg.public.customers c ON c.id = e.customer_id
GROUP BY ALL;Remember
INSTALLdownloads once;LOADactivates; many core extensions autoloadCREATE SECRETmanages S3/GCS/Azure credentialsATTACH ... (TYPE postgres|mysql|sqlite)queries other databases in place
Watch out for
- Pulling an entire remote PostgreSQL table into DuckDB when a filter could have been pushed to the source; check EXPLAIN.