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.

What it’ll do

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.

Built for

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.

faq

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.