Warehouse Routing
When a query has to hit Snowflake, send it to a warehouse that’s already warm. Melt watches the fleet in real time and lands each statement on the cheapest warm cluster that can run it — no Enterprise-tier multi-cluster, no cold-start tax, no per-tenant pinning spreadsheet.
The capabilities, at a glance.
Real-time warm-warehouse detection
The proxy already terminates every Snowflake session and observes every passthrough query end-to-end.
Concurrency-aware placement, not just "warmest"
The naive policy — always route to the most-recently-used warehouse — collapses a fleet into one warehouse that queues.
Per-statement override via SQL hint
Routing is a heuristic; we’d rather give operators an out than ship a black box.
Where this fits.
Data platform team on Snowflake Standard with bursty ad-hoc traffic
5–20 person team, $20K–$200K/month Snowflake bill, a handful of warehouses sized M to XL, AUTO_SUSPEND at 60–300 seconds because every credit hurts.
Multi-tenant SaaS with a warehouse-per-tenant pattern
Per-tenant isolation is contractual, but most tenants are quiet most of the time; the cold-start bill on the long tail is several times the bill on the active tenants.
Anyone running an agent fleet against Snowflake
Coding agents, research agents, autonomous pipelines — workloads that fire small queries at machine cadence with no human-driven amortisation window.
Frequently asked questions.
What is warehouse routing?
When a query has to hit Snowflake, warehouse routing lands it on a warehouse that is already warm. Melt observes the warehouse fleet in real time off the wire (no INFORMATION_SCHEMA polling, no synthetic probes) and scores each candidate by warmth recency, in-flight count, concurrency headroom, queue depth, and size. The goal is to stop paying the 60-second cold-start minimum Snowflake bills every time a warehouse resumes.
Why not just use Snowflake multi-cluster warehouses?
Multi-cluster warehouses solve cross-cluster fan-out natively, but the feature is gated to Snowflake’s Enterprise tier, which carries roughly a 50% per-credit premium. Warehouse routing gives Standard-tier accounts the same shape — a pool of warehouses with traffic distributed across them — without paying the Enterprise premium. You still own the warehouse list; Melt just picks the right one per statement.
Does warehouse routing change which warehouse my driver connects to?
The driver still connects to the warehouse it is configured for. Warehouse routing rewrites the warehouse at the session-swap layer inside the proxy for a given statement; the driver does not see it. A leading /*+ MELT_WAREHOUSE('TRANSFORM_WH') */ SQL hint pins a query to a named warehouse and bypasses the router when you need explicit control — for example, on a tenant warehouse with isolation requirements.
Is warehouse routing production-ready?
Warehouse routing is in alpha. Design and prototyping are active, with a flag-gated implementation planned in a near-term release and a public design RFC ahead of GA. The warmth-ledger plumbing piggybacks on the existing per-statement instrumentation, and the session-swap path is an extension of the existing passthrough rewrite, so the surface area is incremental rather than a new service.
Want warehouse routing in your stack early?
We’re shipping this with a small group of design partners. Tell us about your workload and we’ll set you up.