Contents

• pg_relation_sql — JOIN BY FOREIGN KEY
  • Install
  • How it works
    • Zero runtime cost
  • Query patterns
  • The generator
    • Naming rules
  • Demo
  • Status
  • Support

pg_relation_sql — JOIN BY FOREIGN KEY

Join relations by foreign keys in plain SQL.

Every foreign key becomes a pair of SQL functions — lookup and list — fully inlined by the PostgreSQL planner: the function is the relation, its argument is the model. Queries navigate declared relations instead of restating them in every ON:

SELECT client.name, profile_detail.phone, delivery_address.city, item.name, quantity
FROM document
, client(document)
, profile_detail(client)
, delivery_address(document)
, document_item_list(document)
, item(document_item_list)
WHERE document.doc_number = 'DOC-1'

The planner turns it into exactly the query you would have written by hand (dbfiddle):

SELECT p.name, pd.phone, a.city, i.name, di.quantity
FROM document d
JOIN profile p ON p.client_id = d.client_id
JOIN profile_detail pd ON pd.profile_id = p.id
JOIN address a ON (a.id, a.profile_id) = (d.delivery_address_id, d.client_id)
JOIN document_item di ON di.document_id = d.id
JOIN item i ON i.id = di.item_id
WHERE d.doc_number = 'DOC-1'

Install

One plain SQL file, no extension:

curl -sf https://raw.githubusercontent.com/asmgit/pg_relation_sql/main/relation_sql.sql | psql postgresql://postgres:postgres@localhost:5432/postgres

Alternatively, open relation_sql.sql and execute it in your SQL client — same result.

Check the state anytime:

SELECT status, command FROM relation_sql();

And the relations that appeared, each with a ready-to-run query:

SELECT tbl || ' → ' || calc_name relation
  , 'SELECT * FROM ' || tbl || ', ' || calc_name || '(' || tbl || ')' example
FROM relation_sql('show'), split_part(calc_args, '.', 2) tbl
ORDER BY 1

        relation        |                   example
------------------------+----------------------------------------------
 address → profile      | SELECT * FROM address, profile(address)
 profile → address_list | SELECT * FROM profile, address_list(profile)

Modes, details and removal: The generator.

How it works

Every foreign key gets a pair of generated functions — the relation query itself, written once: a row in, related rows out.

CREATE TABLE profile (
  id BIGINT PRIMARY KEY
  , name TEXT
);
CREATE TABLE address (
  id BIGINT PRIMARY KEY
  , profile_id BIGINT REFERENCES profile (id)
  , city TEXT
);
-- one sync creates the relation functions the foreign keys call for
SELECT status, command FROM relation_sql('sync')
;
-- or turn autosync on
SELECT status, command FROM relation_sql('install')
;
-- and this is what it created: the lookup and the list
CREATE FUNCTION profile(address) RETURNS SETOF profile LANGUAGE sql STABLE PARALLEL SAFE
AS $$ SELECT * FROM public.profile WHERE (id) = (($1).profile_id) $$
;
CREATE FUNCTION address_list(profile) RETURNS SETOF address LANGUAGE sql STABLE PARALLEL SAFE
AS $$ SELECT * FROM public.address WHERE (profile_id) = (($1).id) $$
;
-- queries navigate
SELECT a.city, p.name FROM address a, profile(a) p
;
SELECT p.name, a.city FROM profile p, address_list(p) a
;

Zero runtime cost

When a relation function is used in FROM, the planner fully inlines it: EXPLAIN is node-for-node identical to the hand-written join — same nested-loop/hash strategies, same indexes, same parallel workers. Verified on 1.35M rows: EXPLAIN.md.

The requirements, all guaranteed by the generator: LANGUAGE sql, STABLE, not STRICT, not SECURITY DEFINER, and PARALLEL SAFE — parallel safety is decided before inlining, so the default PARALLEL UNSAFE would silently block parallel workers for the whole query. The full inlining conditions: Inlining of SQL functions.

Two spellings do not inline — both fine for selective queries, both documented with plans in EXPLAIN.md:

• attribute notation in a select list (ProjectSet, one call per row);
• EXISTS (SELECT FROM f(x)) — sublinks become semi/anti joins before functions inline, so this stays a correlated index probe per row; on full-table anti-joins the classic subquery wins.

Two more practical notes. The generated body is SELECT *, so column-level privileges (GRANT SELECT (id, name)) don’t mix with relation functions — after inlining the query needs the whole row; RLS composes as usual, policies apply to the table after inlining. And queries using relation functions run only where the functions exist — the same portability as queries against views; IDEs autocomplete them as ordinary catalog functions.

Query patterns

-- document lines with item, client and optional manager
SELECT d.doc_number, i.name, di.quantity, p.name, m.name
FROM document_item di, document(di) d, item(di) i, client(d) p
LEFT JOIN manager(d) m ON true
;
-- deliveries to Berlin: the composite FK stays hidden
SELECT d.doc_number, a.street
FROM document d, delivery_address(d) a
WHERE a.city = 'Berlin'
;
-- revenue per client: three hops, zero ON
SELECT p.name, sum(di.quantity * i.price) revenue
FROM profile p, client_document_list(p) d, document_item_list(d) di, item(di) i
GROUP BY p.id
;
-- the whole related row as a field
SELECT doc_number, (document.client).* FROM document
;
-- clients without documents
SELECT name FROM profile
WHERE type = 'client' AND NOT EXISTS (SELECT FROM client_document_list(profile))
;
-- recursive tree walk: the row travels as a value, the relation makes the step
WITH RECURSIVE tree AS (
  SELECT item node FROM item WHERE parent_item_id IS NULL
  UNION ALL
  SELECT c node FROM tree t, item_list(t.node) c
)
SELECT (node).name FROM tree
;

These queries run on the demo schema — an ER diagram where every edge carries its function pair, plus all 19 side-by-side pairs, classic vs relations.

The generator

relation_sql.sql is a single function, relation_sql(mode); the event trigger and its support function are created and dropped by install / uninstall. The dashboard right after install:

                                 status                                 |                        command
------------------------------------------------------------------------+-------------------------------------------------------
 pg_relation_sql 0.2.2 — relation functions generated from foreign keys | SELECT status, command FROM relation_sql()
 event trigger: installed                                               | SELECT status, command FROM relation_sql('uninstall')
 relation functions: 16 ok, 0 to sync, 0 foreign, 0 duplicate           | SELECT status, command FROM relation_sql('drop')
 details                                                                | SELECT * FROM relation_sql('show')

The function always returns the same wide row — status and command plus every introspection column; project what you need:

• relation_sql() — read-only dashboard: shows the state and suggests the next command; only status and command are filled.
• relation_sql('show') — the heart: one row per FK per direction; status is the diff status, command the ready-to-run sync SQL, the rest of the columns hold the computed name, arguments, return type and body plus the actual state in the database.
• relation_sql('sync') — apply the diff once; no special rights. The team default: as a migration step, so the functions ship through the same pipeline and review as the rest of your DDL.
• relation_sql('install') / relation_sql('uninstall') — a single ddl_command_end event trigger that re-syncs automatically on CREATE / ALTER / DROP TABLE, CREATE / DROP SCHEMA. Needs superuser; without the rights the trigger is skipped with a WARNING while the sync still runs. Safe by construction: SECURITY DEFINER with a pinned search_path, and a blocked sync step logs a WARNING instead of failing your DDL.
• relation_sql('drop') — remove everything generated; together with 'uninstall' and DROP FUNCTION relation_sql(text) the module leaves without a trace.

Statuses of relation_sql('show'):

status	meaning	on sync
OK	function matches the computed definition	—
NEED_CREATE_OR_REPLACE	absent, or body/parallel flag differs	CREATE OR REPLACE
NEED_DROP_CREATE	return type differs (FK repointed)	DROP + CREATE
FOREIGN_FUNCTION	name taken by a function without the marker	never touched
ORPHANED	marked function whose FK is gone	dropped
DUPLICATE_NAME	two FKs computed the same name	nothing, human resolves

Generated functions carry a COMMENT marker — that is how foreign functions are protected and orphans are found, and how both IDE tooling and schema-diff tools (migra, atlas) can filter the generated set out; with sync as a migration step they are ordinary managed DDL anyway. Renames cascade both ways. Dropping one of two same-target FKs makes the collision disappear — one sync drops the role-prefixed orphans and creates the short names. Adding a second FK to a target renames the existing short list function to a role-prefixed one — dependent queries fail loudly at compile time, and 'show' displays the new names.

Dropping a table with relation functions requires DROP TABLE ... CASCADE — the functions depend on its row type; the event trigger then removes the partner-side leftovers. pg_dump / restore round-trips cleanly: generated functions travel as ordinary objects and the event trigger is restored after the schema, so it does not fire mid-restore.

Naming rules

• Lookup role = FK column minus the referenced column’s name: (client_id) → (id) gives client(document); identical names (client_id → client_id) subtract the target’s PK name instead. Subtraction respects _/CamelCase boundaries — paid and clientid are never mangled.
• Explicit role: name the FK constraint <role>_rel and the role is taken verbatim — delivery_address_rel gives delivery_address(document); renaming the constraint renames the relation. ORM-generated names (_fkey, fk_*) never match the suffix.
• Composite FK: the semantic column is the one referencing the target’s single-column PK — (org_id, customer_id) → (org_id, id) gives customer. Both keys composite — the shared context columns are discarded: from_warehouse / to_warehouse.
• No semantic column left (a column named just id) — the referenced table name: profile(profile_extra).
• List: <referencing_table>_list — document_item_list(document); the same name over different argument types is plain overloading.
• Several FKs to one target — lists get role prefixes: client_document_list, manager_document_list.
• 1:1 (FK columns unique) — singular, no suffix: profile_detail(profile).
• Name taken by a column — _ref suffix: document.client stays the column, document.client_ref navigates.
• Names over 63 bytes are pre-truncated to the identifier limit; any collision, truncation or not, is DUPLICATE_NAME — nothing created, a human resolves.

Functions live in the schema of their argument table; cross-schema FKs work; partition clones and temp tables are skipped.

Demo

No install at all — a db<>fiddle sandbox with the demo schema, pre-generated relation functions, six navigation queries and the two plans side by side.

Full environment:

git clone https://github.com/asmgit/pg_relation_sql.git
cd pg_relation_sql/test && docker compose up -d

Fresh PostgreSQL in Docker with the example schema, bulk data (300k documents, 1.35M lines) and the module fully installed on first start, in about a minute — details in the test environment guide.

Status

Verified live, end to end:

• DDL automation — CREATE TABLE with an FK produces its pair mid-command, dropping the FK cleans it up, renames cascade on collisions.
• Plans — node-for-node identical to hand-written joins on 1.35M rows, parallel plans included.
• Edge matrix — mixed-case and reserved-word identifiers, cross-schema FKs, composite keys with positional roles, 63-byte truncation, duplicate names, foreign functions, orphans, partitioned and temp tables, dump/restore, DDL under arbitrary search_path.

Support

If relation functions save you keystrokes, a star helps others find the project. Sponsoring keeps it maintained — via the Sponsor button above or a coffee at ko-fi.com/asmgit — thank you.