Revision policy

remodel_creates_revision

True — model refits open new versions; as_of returns the value published at that date

precedence

observed wins. If both groups move in one republication the row is classed correction, never remodel.

float_precision

6 — float reserialization below this is not a change

digest_algo

sha256/canon-v1 — stable across library versions

One upstream refit writes ~5,979,895 row versions at current scale. Refits are not rare. The alternative (False) keeps the store small but makes every backtest inherit lookahead bias, which is the defect this service exists to fix.

Model outputs could be versioned against the upstream model version rather than per row: one row per refit instead of ~5,979,895, with as_of still resolving correctly. Blocked — nflverse publishes no model version to key on.

src/fbdata/store/revision.py::assess