An analyst's dashboard slows down whenever the nightly ELT job runs. The cheapest Snowflake fix is to…

Run ELT and BI on separate virtual warehouses over the same data

Make a second copy of the tables

Run ELT and BI on separate virtual warehouses over the same data

Resize the single shared warehouse to 6X-LARGE permanently

Move BI queries to a different cloud region

Snowflake — The Multi-Cluster Shared-Data Architecture — Semantic Web Academy

Theory

One copy of data, many independent warehouses

Snowflake's signature is its multi-cluster, shared-data design. There is one copy of each table in Snowflake-managed object storage, and any number of virtual warehouses can read it at once without contending for the same machines.

Virtual warehouse — A named, resizable compute cluster (X-SMALL → 6X-LARGE). Each size doubles the credits-per-hour and roughly the throughput. Warehouses auto-suspend when idle (you stop paying within seconds) and auto-resume on the next query.
Workload isolation — Give the BI team WH_BI and the ELT job WH_ETL. A giant nightly load on WH_ETL cannot slow a CEO dashboard on WH_BI, because they are different clusters reading the same data. This is the everyday superpower people buy Snowflake for.
Multi-cluster (auto-scale-out) — For a single warehouse under spiky concurrency (e.g. 500 analysts at 9am), Snowflake can add more clusters of the same size and load-balance queries across them, then scale back down. Resize = bigger queries; multi-cluster = more simultaneous queries.

Use Case Example: Month-end, finance runs a monster reconciliation. You bump WH_FINANCE from MEDIUM to X-LARGE for the night (faster single queries), set MIN_CLUSTER=1, MAX_CLUSTER=3 for the 9am dashboard surge (more concurrency), and auto-suspend everything at 60s idle so the rest of the month costs almost nothing.

Analogy

A Snowflake table is a single reference library, and virtual warehouses are separate reading rooms with their own desks and lamps. The law students (BI) and the medics (ELT) read the same books but never fight over the same desks. Need each book read faster? Give a room bigger desks (resize). Too many readers queued at the door? Open identical overflow rooms (multi-cluster). Nobody reading? Switch the lights off and stop paying (auto-suspend) — the books don't move.

Warehouses over shared data

LayoutLabelsClick a node to focus its neighbourhood · drag to pan · scroll to zoom

Shared data, isolated compute

Two warehouses, one storage layer. Resizing changes a warehouse's power; multi-cluster changes its parallelism.

Reflect

Snowflake's elegance is that 90% of its performance and cost story reduces to two dials — size and cluster count — plus auto-suspend. Most teams that overspend never touched auto-suspend.

›If you mapped each team to its own warehouse, what would the monthly bill reveal that you can't see today?
›Where in your stack would 'resize' be the wrong fix because the real problem is concurrency?

Reading in progress · 0 of 2 activities done