A column compared with a function of itself no longer collapses onto one Painless parameter (#370)

[fupelaqu]

The defect

PainlessParam identity was the doc-value string (Identifier.param = doc['d'].value), while a folding function's method is attached to the parameter object (toPainless → ctx.find(base).addPainlessMethod). The second use of a column in one script resolved, through PainlessContext.get, to the first object — which by then carried the first chain. Measured on ES 8.18 before the fix:

shape	emitted	outcome
WHERE DATE_TRUNC(d, MONTH) < d	left1.isBefore(param1), truncation folded INTO param1	always false, HTTP 200
WHERE DATE_ADD(d, INTERVAL 1 DAY) = d	left1.isEqual(param1)	always true
WHERE YEAR(d) > MONTH(d)	one parameter with .get(YEAR).get(MONTH_OF_YEAR)	int.get → shard failure
CASE … WHEN CAST(lastUpdated AS date) - INTERVAL 3 DAY THEN lastUpdated …	THEN returned lastUpdated − 3 days	wrong branch value — and the bridge fixture had pinned it

The WHERE translator emits each AND branch as its own script filter, so YEAR(d) = 2025 AND MONTH(d) = 1 was never wrong live; the collapse needs one expression, or one script (CASE, COALESCE, Criteria.painless).

The fix — seven rules, each forced by a measurement

1. Identity = column + the folded leading run of its chain (Identifier.contextKey), keyed on the folded functions' renderings so two spellings of one fold share (YEAR(d) ≡ EXTRACT(YEAR FROM d)). TransformFunction.foldsOntoOperand is the same startsWith(".") test toPainless folds on, taken context-free; the run stops at the first non-folding function (YEAR(DATE_PARSE(name, …)) keys on the raw name, which the DATE_PARSE argument shares). DATE_TRUNC … QUARTER binds instead of folding and overrides the test at the site that knows. equals/hashCode untouched — GenericIdentifier inherits them and the AST must not move.
2. UTC normalisation first (prependPainlessMethod): a .toLocalDate() narrowing is appended at coercion time; on one shared object the fold masked the order, on two objects …toLocalDate().toInstant() fails.
3. DateTrunc reads its receiver (rendersLocalDate) off the registered parameter — two instances of one identifier can share a parameter, and reading expr appended both renderings. Context-free the receiver is the type only, else the identity key depends on predicate order.
4. The interval family folds onto its operand's parameter (ctx.find(base)), not onto "the last parameter created" (ctx.last). Found by review: CASE WHEN DATE_ADD(d, INTERVAL 1 DAY) > ts THEN DATE_ADD(d, INTERVAL 1 DAY) ELSE d END added the day to ts and returned it; with THEN/ELSE swapped it folded onto the boolean condition.
5. A no-schema CAST declares the fold its toPainless really does, so CAST(x AS DATE) - INTERVAL 3 DAY keys apart from a bare x.
6. A CAST is the identity on a receiver already narrowed to LocalDate. Rule 4 keeps a coercion ctx.last used to discard — the only reason CAST(d AS DATE) - INTERVAL 3 DAY = <date literal> worked before; and plain CAST(d AS DATE) = <date literal> failed the shard on main itself. One rule fixes both.
7. Same for LocalTime (CAST(ts AS TIME) + INTERVAL 1 HOUR > …).

Rules 6 and 7 repair two regressions rule 4 introduced; both were found by the independent reviewer executing ~130 renderings against a control clone, not by any suite.

Verification

Every emission that moved was executed on real Elasticsearch, before and after: the four defect shapes; the CASE/COALESCE interval shapes over a fixture covering every branch (before: d → 2024-06-01 (ELSE), after: d → 2025-01-03; e → lastUpdated = 2025-01-04, where the old pin answered 2025-01-01); the CAST-vs-literal shapes ([b, c]); the QUARTER, CAST, YEAR(DATE_PARSE(…)), LAST_DAY, UPPER must-not-move shapes (byte-identical).

• ParameterIdentitySpec (19): the mechanism (distinct parameters iff folded renderings differ; DATE_FORMAT's outer/inner share), the defect pins, the must-not-move pins, one pin per rule — each rule seen red by mutation.
• PredicateTransformSurvivalSpec's 4,209-predicate sweep now renders the operand for its predicate (it rendered it in isolation, which rule 3 made a wrong reference), and its DATE_TRUNC = CAST(…) pin — which had certified an emission ES rejects (LocalDate.toInstant) — pins the executed one.
• PredicateFunctionResultSpec on the five clients / ES 6.8 (rest, jest), 7.17, 8.18, 9.0: the A column compared with a function of itself collapses onto one Painless parameter — silent wrong answer #370 row flipped from the recorded-wrong Set() to the truth; YEAR(d) > MONTH(d), DATE_ADD(d, INTERVAL 1 DAY) = d, CAST DATE/TIME rows added; AND forms and QUARTER kept as controls.
• Both bridge copies' CASE pins moved and were executed; caseWhenExpr had pinned the wrong branch.
• sql 1405 · core 1129 · five bridge copies 208 each · 2.12.20 cross-compile · CI lint line — green on the committed bytes.
• Parse cost ×5 interleaved against a control clone at 650345fa: control median 3404 µs (3200–3467), branch 3408 µs (3247–3646), ranges overlap — no signal. Measured at rules 1–3; the later rules are emission-time only.

Independent review, three passes on its own clone, before commit: the first pass found rule 4's population (ctx.last) and the pinned wrong CASE branch; the second found the two regressions rule 4 introduced; the third confirmed the repairs. The reviewer's own "order A executes correctly" claim did not reproduce — both predicate orders of that shape are loud (residual 1 below) and the pin says so.

Release notes

• DATE_TRUNC/DATE_ADD/DATE_SUB/extractor chains compared with, or combined with another chain of, their own column now answer correctly; CASE/COALESCE branches over the same column with different intervals return the right value. Queries that silently returned the wrong rows will change.
• CAST(<date column> AS DATE) = <date literal> (and >, BETWEEN, = CURRENT_DATE) now works; it failed the shard before.
• Three mixed LocalDate/ZonedDateTime comparisons that were silently false now fail loudly (residual 1).
• Downstream fixtures pinning these Painless emissions need updating; PainlessParam gains contextKey, prependPainlessMethod, rendersLocalDate/rendersLocalTime; TransformFunction gains foldsOntoOperand (additive, sealed/trait-default).

Residuals — pre-existing, deliberately out of scope, all in one follow-up

See #373 — nine items, each executed on real Elasticsearch before and after. Per the lead's decision they ship as one PR.

Closes #370

🤖 Generated with Claude Code