db-gaussdb-sql

name	db-gaussdb-sql
description	Generate, review, and understand GaussDB / openGauss SQL. Use for GaussDB queries, DDL, DML, table design, and SQL rewrites where PostgreSQL compatibility, A/B/PG database compatibility modes, distributed deployments, or version-specific syntax matter.

GaussDB SQL

Generate GaussDB-compatible SQL from metadata-provided object and column names. Treat PostgreSQL 9.2 as the baseline syntax; GaussDB Kernel 50x and openGauss share this dialect. Write PostgreSQL-style SQL only — never Oracle or MySQL syntax, regardless of the database's compatibility mode.

Namespaces and identifiers

• Address objects in the connected database as schema.table; use public only when metadata or connection context selects it. There is no cross-database qualification.
• Fold unquoted identifiers to lowercase. Preserve mixed case, reserved words, and special characters with double quotes; do not use backticks or square brackets.
• Avoid schema names reserved by GaussDB, including dbe_*, pg_*, blockchain, snapshot, db4ai, and cstore.

Compatibility modes

• Each database runs in one compatibility mode: A (Oracle, the common GaussDB default), B (MySQL), or PG. The mode changes semantics, not accepted syntax — keep generating PostgreSQL-style SQL in every mode.
• In A mode, the empty string and NULL are the same value: use col IS NULL / col IS NOT NULL for emptiness checks. col = '' never matches, and inserting '' stores NULL.
• In A mode, string concatenation follows Oracle semantics: NULL || 'x' returns 'x', not NULL.
• Do not use Oracle-only functions or syntax (nvl, decode, sysdate, add_months, (+) outer joins) even in A mode; use PostgreSQL equivalents (coalesce, CASE, current_timestamp, interval arithmetic, ANSI joins).
• Prefer explicit casts (CAST(x AS type) or x::type) — A mode implicit conversions differ from vanilla PostgreSQL, and DATE behaves like a timestamp there.
• In B mode, date1 - date2 subtracts the dates as numbers (20240315 - 20240101 = 214), not as a day count. For a portable day difference in every mode use to_char(d1,'J')::int - to_char(d2,'J')::int; DATEDIFF exists only in B mode.
• In B mode, NULLs sort first ascending (opposite of PostgreSQL). When NULL placement matters — especially before LIMIT — always spell it out: ORDER BY col NULLS LAST (or NULLS FIRST).
• In B mode, extract(week ...) uses MySQL week numbering; use to_char(d,'IW') for ISO week in every mode.
• In B mode, boolean values render as '1'/'0' and aggregates over an empty set may not be NULL. For text output of booleans compare explicitly (CASE WHEN flag THEN ...) instead of relying on the rendering.

Queries and types

• Use PostgreSQL-style LIMIT and OFFSET, ANSI joins, window functions, and CTEs.
• Prefer TEXT for unbounded strings and VARCHAR(n) only when a length bound is meaningful.
• Use GaussDB-supported PostgreSQL types such as SMALLINT, INTEGER, BIGINT, NUMERIC, REAL, DOUBLE PRECISION, BOOLEAN, TEXT, BYTEA, DATE, TIMESTAMP, and TIMESTAMPTZ according to metadata.
• Treat PostgreSQL features newer than 9.2 as version-dependent: to_regclass, FILTER aggregate clauses, LATERAL, and ON CONFLICT are not universally available. Confirm support before generating them.

Distributed deployments

• On distributed GaussDB, large tables carry DISTRIBUTE BY HASH (col, ...); reference and dimension tables may use DISTRIBUTE BY REPLICATION. Centralized deployments (and openGauss) take no DISTRIBUTE BY clause.
• Distribution key columns cannot be UPDATEd; never generate UPDATE statements that modify a table's distribution key.
• Prefer joins and GROUP BY on distribution keys to avoid cross-node data shuffling.

Writes and transactions

• Use INSERT, UPDATE, and DELETE with PostgreSQL syntax. For upserts, prefer INSERT ... ON DUPLICATE KEY UPDATE or MERGE INTO — ON CONFLICT support is version-dependent. ON DUPLICATE KEY UPDATE is a GaussDB extension accepted in every compatibility mode and is the one intentional exception to the "no MySQL syntax" rule above.
• In A mode, inserting an empty string stores NULL; source data containing empty strings will not round-trip.
• Do not mix DDL and DML in one multi-statement transaction; some DDL is restricted inside explicit transactions, especially on distributed deployments.

Avoid common dialect leaks

Before returning SQL, reject MySQL backticks and table options, Oracle-only functions and (+) joins, Snowflake three-part names, unguarded ON CONFLICT, and = '' emptiness checks that break under A mode.