Hi all, Following up on this thread -- I've written the proposal up as a FLIP draft (view-only Google Doc, FLIP-XXX placeholder per the non-committer process):
FLIP-XXX: Model Routing for ML_PREDICT https://docs.google.com/document/d/1Xgf4hqHr6NhKoVVbyqOzLilLTPhVlyFf-uehkaYaj6A/edit The FLIP proposes a routing model: an ordinary CREATE MODEL with 'provider' = 'routing' that selects one candidate model per row inside the existing ML_PREDICT path -- e.g. a cheap model for easy rows and a premium model for hard ones, with declarative fallback when a candidate fails. No SQL grammar changes or new runtime operators; the one framework change is a small extension so candidate models referenced by name are resolved from the catalog and persisted into compiled plans. A proof-of-concept implementation exists and I'll share the branch link in this thread. Looking forward to feedback -- in particular whether the catalog-resolution and plan-persistence extension to ModelProviderFactory.Context / ModelSpec looks acceptable (details in the doc). Best, Purshotam Shah
