The scheduled rescore job rewrote its whole book on every pass, whether or not a record’s score had moved. On a 4,000-record score that was 4,000 row rewrites a day for nothing, and the database was warning that its disk IO budget was being spent.
The score cache upsert now carries a condition on its state columns: the score, the converted flag, the delivery method and the feature columns. A row whose state is unchanged is not written, so there is no new row version, no log entry and nothing for vacuum. Score history, on Enterprise, is logged only for records whose cached score changed, with the real previous score and delta; before, every pass logged every record as a fresh baseline. Opportunity snapshots skip an open opportunity whose stage, amount and tracked fields equal its latest row.
Each run row now says how many cache rows changed and how many unchanged rows
were skipped. One reading changes with this: a cached score’s scored_at is
now the time of its last real change, not the last sweep.