Search Shortcut cmd + k | ctrl + k
trino_parity

Native Trino-compatible scalar functions (trino_*) for the cases where DuckDB's built-ins diverge from Trino on Unicode and byte-level inputs — Trino's simple per-code-point case mapping, code-point reverse, Java-whitespace trim, NFC normalize, and raw-bytes xxhash64/sha512/hmac_sha256

Maintainer(s): apatrida

Installing and Loading

INSTALL trino_parity FROM community;
LOAD trino_parity;

Example

INSTALL trino_parity FROM community;
LOAD trino_parity;

-- Trino's upper()/lower() are simple per-code-point mappings
-- (Character.toUpperCase(int)), so German sharp-S is unchanged —
-- DuckDB's built-in upper() gives 'STRAẞE' (U+1E9E):
SELECT trino_upper('straße');     -- 'STRAßE'

-- Unicode 16 data includes case pairs absent from older ICU snapshots:
SELECT trino_lower(chr(11311)) = chr(11359) AS glagolitic_unicode16; -- true

-- Java whitespace trim: tab / LF / CR / FF / VT are stripped
-- (DuckDB's built-in trim() only strips spaces):
SELECT trino_trim(E'\tx\n');       -- 'x'

-- Code-point reverse (DuckDB's built-in reverse is grapheme-aware):
SELECT trino_reverse('a😀b');     -- 'b😀a'

-- The catalog of functions this extension provides:
SELECT trino_name, arg_count, category FROM trino_meta();  -- 10 rows

About trino_parity

trino_parity provides native trino_<name>(...) scalar functions for the specific cases where DuckDB's built-ins diverge from Trino's documented behaviour on Unicode and byte-level inputs.

Version 0.5.0 vendors ICU 76.1 / Unicode 16.0, fixing newer case mappings and NFC normalization data missing from the previous Unicode 13 snapshot. The compatibility target is Trino 483 on JDK 25. Matching version labels alone does not guarantee parity with other Trino/JDK combinations.

It is designed to be loaded server-side by anything that pushes Trino-shaped predicates down to DuckDB. If a Trino connector naively pushes WHERE lower(name) = 'apple' down to DuckDB, the filter runs with different Unicode semantics in each engine, and rows visible to Trino can silently disappear from the result. This extension supplies only the functions where that actually happens — where DuckDB's built-in genuinely diverges and a native C++ reimplementation is required to match Trino. For the large majority of Trino functions DuckDB's built-in is already equivalent (or a trivial rename / operator / one-line rewrite), so a caller emits those directly against DuckDB — no extension needed.

Concrete divergences fixed

Function DuckDB built-in Trino spec trino_* result
upper('ß') (U+00DF) 'ẞ' (U+1E9E, utf8proc special case) 'ß' unchanged (simple per-code-point mapping) matches Trino
reverse('cafe'+U+0301) grapheme-aware code-point-only matches Trino
trim (tab/LF/CR/FF/VT) not stripped Java Character.isWhitespace strips them matches Trino
xxhash64(varbinary) no direct built-in XXH64, big-endian bytes matches Trino

Function inventory

trino_meta() catalogs the ten native functions the extension provides:

  • String (ICU): trino_lower/1, trino_upper/1, trino_reverse/1, trino_trim/1, trino_ltrim/1, trino_rtrim/1, trino_normalize/1 — simple per-code-point case mapping (Trino's Character.toUpperCase(int) model — no SpecialCasing expansions or final-sigma rule), code-point reverse, and Character.isWhitespace trim, via statically-linked ICU 76.1 vendored into the binary (so Unicode behaviour is independent of the host DuckDB build).
  • Hash (vendored): trino_xxhash64/1, trino_sha512/1, trino_hmac_sha256/2 — over raw VARBINARY, self-contained (no dependency on the crypto / hashfuncs community extensions). Empty HMAC keys are rejected with Empty key, matching Trino; empty messages and nonempty binary keys remain supported.

The ten-function catalog and three-column trino_meta() schema are unchanged. Validation against pinned OpenJDK 25 covers all 1,112,064 valid Unicode scalar values and 99,825 NFC conformance column checks, with zero mismatches. See the Unicode 16 validation report for provenance, commands, and the limits of that validation.

Aligned Trino functions (e.g. length, abs, year, substring, regexp_extract) are intentionally NOT shipped — a caller emits them as bare DuckDB SQL. The full per-function Trino↔DuckDB mapping and the empirical audits that justify each decision are in the repository's docs/.

Source and full divergence catalog: https://github.com/brikk/duckdb-trino-parity-extension

Added Functions

function_name function_type description comment examples
trino_hmac_sha256 scalar NULL NULL  
trino_lower scalar NULL NULL  
trino_ltrim scalar NULL NULL  
trino_meta table_macro NULL NULL  
trino_normalize scalar NULL NULL  
trino_reverse scalar NULL NULL  
trino_rtrim scalar NULL NULL  
trino_sha512 scalar NULL NULL  
trino_trim scalar NULL NULL  
trino_upper scalar NULL NULL  
trino_xxhash64 scalar NULL NULL  

Overloaded Functions

This extension does not add any function overloads.

Added Types

This extension does not add any types.

Added Settings

This extension does not add any settings.