fix(tradein/db): widen offer_price_history.diff_percent (5,2)→(8,2) + clamp в триггере (краш full_load) #1827

Merged
lekss361 merged 1 commit from fix/tradein-diff-percent-overflow into main 2026-06-20 08:46:59 +00:00

View file

@ -0,0 +1,103 @@
-- Migration 131: fix NumericValueOutOfRange crash in record_listing_price_change()
--
-- Incident: prod run_id=260, 2026-06-20
-- A listing scrubbed by avito_full_load had an OLD.price_rub that was
-- garbage-small (data entry error or placeholder value). The computed
-- diff_percent = round((NEW-OLD)/OLD*100, 2) exceeded ±999.99, which
-- overflows numeric(5,2) (precision=5, scale=2, max value=±999.99).
-- PostgreSQL raised DataError: "numeric field overflow" inside the
-- trigger, which caused the INSERT into listings to abort, which caused
-- the entire avito_full_load batch to crash.
-- The trigger is shared across all non-cian sources → same crash would
-- occur for yandex/domclick full_load batches on similar data.
--
-- Fix (two parts):
-- 1. Widen diff_percent column to numeric(8,2) (max ±999999.99%).
-- Widening precision is safe and lossless for existing data.
-- 2. Replace the trigger function to clamp the computed value into
-- [999999.99, +999999.99] so even future extreme deltas cannot
-- overflow the column. All other logic is preserved exactly as in
-- migration 116.
--
-- Dependencies: offer_price_history table (023), listings table,
-- offer_price_history_listing_change_uq unique constraint,
-- trigger listings_price_change_trg (116).
--
-- Idempotent: ALTER TYPE widening is a no-op if already numeric(8,2);
-- CREATE OR REPLACE FUNCTION is always safe to re-run.
-- Trigger is NOT recreated — the existing trigger already points to
-- record_listing_price_change(); replacing the function body is enough.
--
-- Deploy order: standalone; auto-applied by deploy pipeline via
-- _schema_migrations. Do NOT apply manually on prod — causes drift.
BEGIN;
-- Part 1: widen the column.
-- numeric(5,2) → numeric(8,2): raises max from ±999.99 to ±999999.99.
-- PostgreSQL widens precision without rewriting rows (catalogue-only change
-- for heap tables when scale does not decrease and precision does not decrease).
ALTER TABLE offer_price_history
ALTER COLUMN diff_percent TYPE numeric(8,2);
-- Part 2: replace trigger function with clamped diff_percent computation.
-- All guards from migration 116 are preserved verbatim:
-- • same-value skip (IS NOT DISTINCT FROM)
-- • NULL new-price skip
-- • cian source skip (backfill exclusion — see 116 for rationale)
-- • OLD.price_rub > 0 gate (avoid division by zero / nonsensical pct)
-- • ON CONFLICT DO NOTHING (idempotent re-scrape)
CREATE OR REPLACE FUNCTION record_listing_price_change()
RETURNS trigger
LANGUAGE plpgsql
AS $$
BEGIN
-- Skip if price did not actually change (same-value upsert write).
IF NEW.price_rub IS NOT DISTINCT FROM OLD.price_rub THEN
RETURN NEW;
END IF;
-- Skip if the new price is NULL (defensive; price_rub is NOT NULL in practice).
IF NEW.price_rub IS NULL THEN
RETURN NEW;
END IF;
-- Skip cian: its history is source-provided via cian_price_history.py backfill.
-- That service selects listings WHERE oph.listing_id IS NULL (zero rows in
-- offer_price_history). Inserting a trigger row here would permanently
-- exclude those listings from the backfill, losing accurate historical timestamps.
IF NEW.source = 'cian' THEN
RETURN NEW;
END IF;
INSERT INTO offer_price_history (
listing_id,
change_time,
price_rub,
source,
diff_percent,
recorded_at
) VALUES (
NEW.id,
now(),
NEW.price_rub,
NEW.source,
CASE
WHEN OLD.price_rub IS NOT NULL AND OLD.price_rub > 0
THEN LEAST(GREATEST(
round(
(NEW.price_rub::numeric - OLD.price_rub::numeric)
/ OLD.price_rub::numeric * 100,
2
),
-999999.99), 999999.99)
ELSE NULL
END,
now()
)
ON CONFLICT ON CONSTRAINT offer_price_history_listing_change_uq
DO NOTHING;
RETURN NEW;
END;
$$;
COMMIT;