pg_get_indexdef() output not idempotent for partial indexes with ALL(ARRAY[…])::text[]
*PostgreSQL version*: 18.3 (Homebrew) on aarch64-apple-darwin24.6.0
*pg_get_indexdef()* produces SQL that, when executed, yields a different
pg_get_indexdef() output. This means a pg_dump → pg_restore cycle silently
changes the deparsed form of partial index WHERE clauses that use NOT IN
(...) on a varchar column, causing cosmetic drift in tools that compare
index definitions (e.g. ORM schema dumps, annotation generators).
*Reproduction:*
-- Setup
CREATE TABLE test_idempotent (
id bigint PRIMARY KEY,
state varchar NOT NULL,
space_id bigint
);-- Step 1: Create index using NOT IN
CREATE UNIQUE INDEX idx_original ON test_idempotent (space_id)
WHERE state NOT IN ('completed', 'failed', 'cancelled')
AND space_id IS NOT NULL;-- Step 2: Capture pg_get_indexdef output
SELECT pg_get_indexdef(indexrelid) AS pass_1
FROM pg_stat_user_indexes
WHERE indexrelname = 'idx_original';-- pass_1 result:
-- CREATE UNIQUE INDEX idx_original ON public.test_idempotent USING btree
(space_id)
-- WHERE (((state)::text <> ALL ((ARRAY['completed'::character varying,
-- 'failed'::character varying, 'cancelled'::character
varying])::text[]))
-- AND (space_id IS NOT NULL))-- Step 3: Feed pass_1 output back as a new index
CREATE UNIQUE INDEX idx_round_trip ON public.test_idempotent USING btree
(space_id)
WHERE (((state)::text <> ALL ((ARRAY['completed'::character varying,
'failed'::character varying, 'cancelled'::character varying])::text[]))
AND (space_id IS NOT NULL));-- Step 4: Compare both
SELECT indexrelname, pg_get_indexdef(indexrelid)
FROM pg_stat_user_indexes
WHERE indexrelname IN ('idx_original', 'idx_round_trip')
ORDER BY indexrelname;-- Cleanup
DROP TABLE test_idempotent;
*Observed result (step 4):*
idx_original | ... WHERE (((state)::text <> ALL
((ARRAY['completed'::character varying, 'failed'::character varying,
'cancelled'::character varying])::text[])) AND (space_id IS NOT NULL))
idx_round_trip | ... WHERE (((state)::text <> ALL
(ARRAY[('completed'::character varying)::text, ('failed'::character
varying)::text, ('cancelled'::character varying)::text])) AND (space_id IS
NOT NULL))
The array-level cast (ARRAY[...])::text[] is redistributed to per-element
casts ARRAY[(...::text), (...::text), (...::text)] after a single
parse→deparse cycle.
*Expected result:*
Both indexes should produce identical pg_get_indexdef() output since
idx_round_trip was created using the exact SQL that pg_get_indexdef()
produced for idx_original.
Marcelo Lauxen <marcelolauxen16@gmail.com> writes:
*PostgreSQL version*: 18.3 (Homebrew) on aarch64-apple-darwin24.6.0
*pg_get_indexdef()* produces SQL that, when executed, yields a different
pg_get_indexdef() output. This means a pg_dump → pg_restore cycle silently
changes the deparsed form of partial index WHERE clauses that use NOT IN
(...) on a varchar column, causing cosmetic drift in tools that compare
index definitions (e.g. ORM schema dumps, annotation generators).
You are assuming a property that we've never guaranteed and don't plan
to start guaranteeing, ie that the output of expression decompilation
matches the input even in semantically-insignificant details.
My own advice about how to fix this particular example is not to use
varchar --- especially not unconstrained varchar, which doesn't even
have the thin excuse of being spec-compliant. Postgres' native string
type is text.
regards, tom lane
Gotcha, it wasn't clear to me that this was never guaranteed. I will change
the type of this column to text to resolve this.
Appreciate the quick response!
Regards,
Marcelo
On Wed, May 13, 2026 at 10:51 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Show quoted text
Marcelo Lauxen <marcelolauxen16@gmail.com> writes:
*PostgreSQL version*: 18.3 (Homebrew) on aarch64-apple-darwin24.6.0
*pg_get_indexdef()* produces SQL that, when executed, yields a different
pg_get_indexdef() output. This means a pg_dump → pg_restore cyclesilently
changes the deparsed form of partial index WHERE clauses that use NOT IN
(...) on a varchar column, causing cosmetic drift in tools that compare
index definitions (e.g. ORM schema dumps, annotation generators).You are assuming a property that we've never guaranteed and don't plan
to start guaranteeing, ie that the output of expression decompilation
matches the input even in semantically-insignificant details.My own advice about how to fix this particular example is not to use
varchar --- especially not unconstrained varchar, which doesn't even
have the thin excuse of being spec-compliant. Postgres' native string
type is text.regards, tom lane