duckdb/duckdb

Other in-tree extensions

A short tour of the smaller in-tree extensions: autocomplete, TPC-H, TPC-DS, jemalloc, delta, demo_capi, and loader.

autocomplete

extension/autocomplete/ powers the SQL autocomplete used by the DuckDB CLI shell. It is a self-contained suggestion engine that knows about:

SQL keywords (driven by the same data the parser uses).
Identifiers visible in the catalog.
Built-in and registered function names.
Common SQL patterns ("after SELECT, suggest column expressions then FROM").

The CLI calls into it via the C API to populate linenoise-style completion lists. See extension/autocomplete/CMakeLists.txt and the tools/cli-shell page.

tpch

extension/tpch/ registers the TPC-H schema and a data generator:

INSTALL tpch;
LOAD tpch;
CALL dbgen(sf=1);          -- generate scale factor 1
SELECT * FROM tpch_queries(); -- list canonical queries
PRAGMA tpch(1);            -- run query #1

It bundles the official TPC-H reference query texts (extension/tpch/dbgen/) and a port of the TPC-H data generator. Used by BUILD_TPCH=1 make for benchmark builds and by the test suite for regression checks (test/sql/tpch/).

tpcds

extension/tpcds/ is the TPC-DS counterpart. It provides:

INSTALL tpcds;
LOAD tpcds;
CALL dsdgen(sf=1);
PRAGMA tpcds(1);

The vendored sources live in extension/tpcds/dsdgen/. Several queries are skipped or rewritten where SQL features have moved between TPC-DS revisions; the differences are documented in extension/tpcds/README.md.

Both tpch and tpcds populate the schema as base tables, so subsequent queries are run against real DuckDB tables.

jemalloc

extension/jemalloc/ ships a vendored jemalloc allocator. When loaded, it replaces the default allocator. On Linux, jemalloc generally improves throughput for write-heavy workloads and reduces fragmentation in long-running processes; on macOS the system allocator is usually fine.

Build with:

DUCKDB_EXTENSIONS=jemalloc make
# or
BUILD_JEMALLOC=1 make

Implementation: a thin wrapper that replaces Allocator::DefaultAllocator (src/common/allocator.cpp) with the jemalloc-backed implementation at extension load.

delta

extension/delta/ is an in-tree wrapper around the delta-kernel-rs Rust library. It exposes a read_delta table function that reads Delta Lake tables (transaction-aware reads with snapshot isolation, deletion vectors, schema evolution).

Most of the implementation lives in the delta-kernel-rs Rust sources fetched at build time; the in-tree code in extension/delta/ is the DuckDB-side glue (table function bind, scan, schema inference). Building requires Rust:

DUCKDB_EXTENSIONS=delta make

demo_capi

extension/demo_capi/ is a documentation-only example: a minimal extension implemented entirely through the C API. It is what extension/README.md references for "how do I write an out-of-tree extension?" It is not loaded by default and exists purely as a working sample.

loader

extension/loader/ is a build-time-only helper that emits the C ABI entry point an extension needs in order to be loadable as a shared library. It does not register any SQL functions. Other extensions depend on it during their build so they all share the same ABI shim.