When a model is updated, existing scores are deprecated. New scores should be backfilled using the updated model.
I can't imagine how to update scores in-place, so I think we have to purge all scores for the old model. Somehow don't break any research queries running on that table, or have some kind of social protocol for downtime.
| Status | Subtype | Assigned | Task | ||
|---|---|---|---|---|---|
| Open | None | T209611 [Epic] Make ORES scores for wikidata available as a dump | |||
| Open | None | T209742 Purge ORES scores from Hadoop and begin backfill when model version changes |
Indeed in hadoop there is no such thing as 'in place'. The way to go could be to use model version as a partition-key. You'd backfill a new version, keeping the old ones (and therefore not breaking queries on old ones). Then the decision is on how many versions to keep, and whether to keep them update or not.