GCP Deployment

Planned — Phase 7

Contents

Cloud Run + Cloud Tasks + durable graph store.

The load-bearing constraint, stated precisely because the headline figure is the wrong one. Cloud Run services allow a request up to 60 minutes (5 minutes by default), but a request dispatched through Cloud Tasks is capped at 30 minutes, defaulting to 10. Since expansions are dispatched, 30 is the number the design actually has to live under — and Google advises tolerating client reconnection beyond 15.

Each graph expansion is therefore one short request, and a long run is many short requests over a durable graph. Work that is blocked rather than slow does not occupy a request at all: see waiting, which is what keeps these ceilings from bounding how long an agent may wait.

Open: AlloyDB is technically right for the memory tier (Postgres + pgvector + ScaNN in one store) but is cluster-based and does not scale to zero. Verify current pricing before committing; Cloud SQL with pgvector may serve the public demo while AlloyDB stays the documented production recommendation.