pg-dump v17 — Wire Compatibility Spec

Normative spec for what pg_dump --version 17 expects from a Postgres-wire server. Used as the contract that infra/data/kdb (kdbpgwire / kdbgateway) must satisfy to support pg_dump roundtrips, and as the sourceof-truth for the per-table coverage work tracked in infra/data/kdb#341.

1. Provenance

The catalogue of queries below was captured live from PostgreSQL 17.9 (Debian 17.9-0+deb13u1) using infra/data/kdb/scripts/pgdump-v17-capture.sh.

The script spins a throw-away PG 17 cluster, applies a varied schema (5 tables, 4 indexes including partial, 2 explicit sequences, FK + CHECK constraints, view, materialized view, enum type), and runs pg_dump under log_statement = all so the server records the exact stream of statements pg_dump issued.

Captures are checked in under meta/docs/stack/specs/postgres-compat/captures/pg-dump-v17-queries-<flag>.sql for the following flag combinations:

Recapture by running the script — output is bytestable across runs of the same PG 17 minor (pg_dump does not emit non-deterministic queries).

2. Session prelude (every invocation)

The first ~13 statements in every pg_dump run are session-shape GUCs. Every server claiming pg_dump-compat MUST accept these without error (returning SET / SELECT is sufficient; the values can be ignored if the server has no equivalent knob):

SELECT pg_catalog.set_config('search_path', '', false);
SELECT pg_catalog.pg_is_in_recovery();
SELECT pg_catalog.set_config('search_path', '', false);
SET DATESTYLE = ISO;
SET INTERVALSTYLE = POSTGRES;
SET extra_float_digits TO 3;
SET synchronize_seqscans TO off;
SET statement_timeout = 0;
SET lock_timeout = 0;
SET idle_in_transaction_session_timeout = 0;
SET transaction_timeout = 0;
SET row_security = off;
SELECT set_config(name, 'view, foreign-table', false)
  FROM pg_settings
  WHERE name = 'restrict_nonsystem_relation_kind';
BEGIN;
SET TRANSACTION ISOLATION LEVEL REPEATABLE READ, READ ONLY;

The final two open a readonly repeatableread transaction that wraps the entire dump; servers without MVCC snapshots can accept these as no-ops provided they expose a stable read for the duration of the dump.

3. Catalogue queries — required surface

pg_dump v17 issues queries against the following pg_catalog relations. Servers MUST respond with the row shape PostgreSQL 17 produces; missing columns cause pg_dump to bail with a parser error in its PQgetvalue(... , <col>) calls. Functions in *old*are outofband helpers (not relations) — they MUST exist and return the documented type.

3.1 Relations (must be queryable, even if empty)

3.2 Relations newly required by PG 17 (gaps vs PG 15/16)

These either did not exist or had different shape in earlier PG and are likely to be missing in a server claiming PG 13/14 compatibility:

3.3 Functions

3.4 PG 17 row-shape additions

The big v15→v17 movement in pg_attribute matters:

• attstorage is "char" in v17 (was oid historically; some legacy

  shims still return oid — pg_dump v17 will reject those rows).

• attcompression exists from v14+ ("char", default 'p' for pglz).
• attmissingval is anyarray (often emitted as text).
• attstattarget is int2 (became nullable in v17 — NULL allowed).
• attgenerated is "char" ('s' for STORED, '' for none).
• attidentity is "char" ('a' for ALWAYS, 'd' for BY DEFAULT, '' for none).

pg_class.relrowsecurity and relforcerowsecurity are bool (RLS-related). pg_class.relispartition is bool (PG 10+); MUST exist even on non-partitioned tables (default false).

4. Object-fetch loop (the heart of pg_dump)

After the session prelude, pgdump runs *ne large `pgclass` query*to materialize the entire tableofobjects:

SELECT c.tableoid, c.oid, c.relname, c.relnamespace, c.relkind, c.reltype,
       c.relowner, c.relchecks, c.relhasindex, c.relhasrules, c.relpages,
       c.relhastriggers, c.relpersistence, c.reloftype, c.relacl,
       acldefault(CASE WHEN c.relkind = 'S' THEN 's'::"char" ELSE 'r'::"char" END,
                  c.relowner) AS acldefault,
       CASE WHEN c.relkind = 'f'
            THEN (SELECT ftserver FROM pg_catalog.pg_foreign_table WHERE ftrelid = c.oid)
            ELSE 0 END AS foreignserver,
       c.relfrozenxid, tc.relfrozenxid AS tfrozenxid, tc.oid AS toid,
       tc.relpages AS toastpages, tc.reloptions AS toast_reloptions,
       d.refobjid AS owning_tab, d.refobjsubid AS owning_col,
       tsp.spcname AS reltablespace, false AS relhasoids,
       c.relispopulated, c.relreplident, c.relrowsecurity,
       c.relforcerowsecurity, c.relminmxid, tc.relminmxid AS tminmxid,
       array_remove(array_remove(c.reloptions, 'check_option=local'),
                                  'check_option=cascaded') AS reloptions,
       CASE WHEN 'check_option=local' = ANY (c.reloptions) THEN 'LOCAL'::text
            WHEN 'check_option=cascaded' = ANY (c.reloptions) THEN 'CASCADED'::text
            ELSE NULL END AS checkoption,
       am.amname, (d.deptype = 'i') IS TRUE AS is_identity_sequence,
       c.relispartition AS ispartition
FROM pg_class c …

(Truncated — see captures/pg-dump-v17-queries-schema-only.sql for the full text.) The relkind values returned drive the rest of the dump:

The LOCK TABLE … IN ACCESS SHARE MODE is non-negotiable: servers MUST either implement table locks or accept the syntax as a no-op (returning LOCK TABLE). Without it, pg_dump v17 aborts the dump.

5. Data path (--data-only)

For each non-system table whose relkind = 'r', pg_dump issues:

COPY public.users (id, handle, email, score, created_at,
                   tags, metadata, avatar, status) TO STDOUT;

The column list is materialized from the pg_attribute query above. The server MUST emit the response on the wire as CopyOutResponse → CopyData* → CopyDone → CommandComplete("COPY n"). Text format is the default; binary is not used by pg_dump v17.

Implementation in kdbpgwire: ticket [infradatakdb#340](......infradatakdbbacklogdone340pgwirecopyto-stdout.kmd).

6. Outofscope for v17 compat

• *oreignkey reconstruction across schemas with `-schema=…`*—

  pg_dump v17 chases FKs to outside-schema parents; the audit script pins to public so the captures do not exercise this corner.

• *erverside COMMENT ON LARGE OBJECT*— largeobject dump uses a

  separate code path (--blobs / --no-blobs) not exercised by the audit. Tracked separately if/when lobs become a goal.

• *arallel pg_dump (-j N)*— parallel restore uses extra COPY

  sessions but the per-session query stream is the same.

• *pg_dump -nserts** — uses INSERT INTO … VALUES (…)` instead of

  COPY. Audit captures focus on the default COPY path.

7. Versioning

Captures are pinned to PG 17.9. Re-run the capture script when:

1. PostgreSQL 18 ships and pg_dump v18 diverges (verify with a fresh

   capture against PG 18 in /usr/lib/postgresql/18/bin/).

2. A user reports a pg_dump failure on kdb-gateway that the diff report

   does not cover.

See kdb-gateway-v0.0.13-vs-pg17-pgdump.md for the perquery coverage delta against the current kdbpgwire surface.

8. Conformance status (20260511)

The 8 originallyidentified RCRITICAL gaps have been closed across a multiday series of fixes culminating in the FK roundtrip milestone. Latest harness state (scripts/pg-dump-roundtrip.sh):

Closed work-items

• #341 umbrella — pg_catalog coverage for the 8 R-CRITICAL gaps
• #342 — round-trip test harness
• #343 — extended-query set_config emission
• #344 (16 waves) — universal tableoid + per-table column gaps
• #345 — planner IS TRUE / IS FALSE
• #346 — subquery catalog inheritance
• #347 — IS DISTINCT FROM
• #348 — text [N] subscript
• #349 — ARRAY (SELECT …) constructor
• #350 — Text → Array<T> cast
• #351 — pg_catalog. prefix-strip for aggregates
• #353 — FK round-trip umbrella (PK ALTER, FK ALTER, conindid linkage)
• #357 — regclass cast → oid
• #360 — FK capture under tenant config
• #361 — synthetic pgindex + pgclass for PK/UNIQUE

Open polish items (post-milestone)

• #352 — schema-diff cosmetic cleanup (35 lines remaining; \restrict

  random tokens unavoidable; public SCHEMA placeholder block deferred)

• #359 — public SCHEMA block suppression (requires pg_dump v17 source

  vendoring to map the suppression algorithm)

Regression coverage

infra/data/kdb/crates/kdb-pgwire/tests/pgwire_fk_roundtrip.rs guards the milestone's three load-bearing wirings (see registry entries 518-520).