DuckDB ADBC Extension Roadmap

Completed

• Connection pooling — concurrent reads against an attached ADBC database each lease their own connection from a per-catalog pool (ADBC connections are not safe for concurrent statement execution).
• Transaction support — BEGIN / COMMIT / ROLLBACK on attached databases map to ADBC autocommit + commit/rollback on a pinned write connection. Drivers that cannot disable autocommit fail loudly rather than silently committing. (Caveat: in an explicit transaction, reads use pooled autocommit connections and do not yet observe the transaction's own uncommitted writes.)
• Streaming writes — CREATE TABLE … AS / INSERT INTO … stream into the driver's bulk-ingest with backpressure, so memory stays flat regardless of input size.
• Predicate & projection pushdown with per-driver dialect — filter pushdown emits the right positional-parameter style per driver (?, $1, @p1) and the right identifier quoting ("…" or `…`). Unsupported filters are dropped (re-applied locally) instead of failing the query.
• Arrow → DuckDB type mapping — attached-table column types come from DuckDB's Arrow type mapping (timestamps with unit/timezone, lists, structs, decimals) instead of a lossy hand-rolled parser.
• Read-only enforcement — ATTACH … (READ_ONLY) rejects writes.
• Driver connectivity examples & CI — seven drivers (SQLite, PostgreSQL, MySQL, Arrow Flight SQL, DataFusion, Trino, SQL Server) are documented in the README and exercised end-to-end in the driver-tests CI workflow against live service containers.

Near Term (weeks)

1. Partition parallelism for table scans — Snowflake can produce result data in partitions; DuckDB could consume that data in parallel.

2. Read-your-writes in transactions — route reads through the transaction's write connection while an explicit transaction is active, so in-transaction reads observe uncommitted writes.

Medium Term

1. INSERT INTO fallback — Support INSERT INTO for drivers that lack ADBC bulk copy. Flight SQL is one.

2. Adopt DuckDB's built-in Arrow-scan machinery — would give scan parallelism and cancellation for free and shed hand-rolled scan code; the work is re-integrating pushdown, statistics, progress, and the schema-vs-data type guard onto it.

Longer Term

All items below apply to ADBC databases created via ATTACH:

1. UPDATE/DELETE without row IDs — The Airport extension demonstrates a possible approach using a custom optimizer pass with replacement LogicalOperators.

2. DDL support — Implement CREATE TABLE and DROP TABLE.

3. Query passthrough — When the DuckDB parser detects that a query operates entirely on an attached ADBC database, send it directly to the foreign server. This would enable server-side aggregation.

---

ADBC Feature Gaps

The following ADBC-spec capabilities would let the extension do more without per-driver heuristics. ADBC's primary focus is connectivity rather than federation, so some may stay out of scope.

Gap	Description	Status
Type mapping	A mapping from Arrow types to driver-native types for table creation (DuckDB → Arrow → driver types).	Reads use DuckDB's Arrow→type mapping; driver-native mapping for DDL is still a gap.
Identifier escaping	A driver-specific way to escape table/column names.	Worked around per driver (" vs `); a driver-reported quote char would remove the heuristic.
Parameter binding introspection	A way for drivers to advertise positional-parameter support and syntax (? vs $1 vs @p1).	Worked around per driver by name; native introspection would remove the lookup table.
Row ID support	DuckDB's UPDATE/DELETE rely on a row_id pseudocolumn. Drivers could indicate an equivalent.	Open — blocks UPDATE/DELETE.
Column constraint metadata	Drivers could expose NOT NULL, UNIQUE, and primary keys with column metadata.	Open.

This roadmap is subject to change. If you'd like to accelerate any of it please contact Query.Farm.