Scaling Warehouses: Up vs Out
Snowflake gives you two scaling levers, and they solve different problems. Scaling up means resizing: same single cluster, more compute, so each individual query gets more muscle. Scaling out means multi-cluster: same size, more identical clusters, so more queries run at once. Confusing them is the most reliably planted trap in Domain 4.
SCALE UP · RESIZE
BEFOREMediumone cluster
resizeAFTERLargeone cluster2× CPU and memory
ALTER WAREHOUSE wh SET
WAREHOUSE_SIZE = 'LARGE';
- Fixes spilling, slow heavy queries
- Still one queue at peak
Credits double with each size step
SCALE OUT · MULTI-CLUSTER
BEFOREMediumone cluster
add clustersALTER WAREHOUSE wh SET
MAX_CLUSTER_COUNT = 3;
- Fixes queuing at peak
- A single query runs no faster
Enterprise edition or higher
- Each query gets more CPU, memory, and SSD
- Fixes spilling and heavy single queries
- One cluster still serves one queue
- Cost doubles per size step, busy or not
- More identical clusters, started only when needed
- Fixes queuing when many users hit at once
- Doesn't make any single query faster
- Requires Enterprise edition
The one-line rule: one slow query, scale up; many queued queries, scale out. The Query Profile tells you which world you're in before you spend anything: spilled bytes point up, queued time points out.
A multi-cluster warehouse is one warehouse name backed by several clusters of identical size, up to its MAX_CLUSTER_COUNT (see the limits below). Snowflake adds clusters when queries queue and retires them when load drops. It exists to solve concurrency, and it requires Enterprise edition or higher.
The problem
Monday 09:00, three hundred analysts open dashboards against BI_WH. A single Medium cluster runs perhaps eight queries at once; the rest queue, and dashboards spin. Resizing to 4X-Large wouldn't help much: each query is small, so the extra power per query is wasted while the queue logic stays the same. What you need is more lanes, not a faster car.
Two modes
Set MIN_CLUSTER_COUNT and MAX_CLUSTER_COUNT. Equal values (both above 1) give Maximized mode: all clusters run whenever the warehouse runs, for predictable full capacity. Different values give Auto-scale mode: Snowflake starts at the minimum and adds clusters up to the maximum as queries queue.
Scaling policies (auto-scale only)
| Policy | Starts a new cluster | Shuts a cluster down | Optimizes for |
|---|---|---|---|
| Standard (default) | Immediately when a query queues | After a sustained period of low load, the least-loaded cluster is shut down once its running queries finish | User experience: no queuing |
| Economy | Only if the new cluster would stay busy for at least 6 minutes of queued work | When Snowflake estimates less than 6 minutes of work remains for the least-loaded cluster, after its queries finish | Credits: tolerates some queuing |
A Monday morning in auto-scale, MIN 1 / MAX 3
A few early queries, no queue. The minimum cluster handles it alone.
Dashboards pile in, queries queue. Standard policy starts cluster 2 the moment the first query waits.
Peak. All three lanes open; nobody queues. This is the ceiling, MAX_CLUSTER_COUNT, and further demand would queue again.
Load fades. After consecutive checks confirm fewer clusters suffice, 3 and then 2 retire. Alice's late query lands on cluster 1 like nothing happened.
Billing follows the badges: you paid for one cluster until 09:02, two for the eight minutes after that, three for the half hour from 09:10 to 09:40, and one again afterward. Figure 2.5 adds it up: about 9.5 credits for the morning, against 15 if all three clusters had run from 08:45 to 10:00. That per-cluster, per-second arithmetic is what makes auto-scale cheaper than a permanently bigger warehouse for spiky concurrency.
A cluster starts as soon as a query queues or Snowflake estimates the running clusters lack capacity. Queues drain fast; credits are spent sooner.
A cluster starts only if Snowflake estimates enough queued load to keep it busy for at least 6 minutes. Short bursts queue instead, saving credits.
Billing is simply the sum of running clusters: a Medium (4 credits/hr) running 3 clusters burns 12 credits per hour while all three are up. Each cluster suspends independently under auto-scale.
- Maximum clusters: the long-standing answer is 10, and Snowsight still caps the setting there. Since February 2025, SQL accepts more on smaller sizes: up to 300 for X-Small to Medium, 160 for Large, 80 for X-Large, 40 for 2X-Large, 20 for 3X-Large, and 10 for 4X-Large and above. Check where the value is set and the warehouse size before you answer 10.
- Multi-cluster answers concurrency scenarios (many users, queuing). If the scenario is one giant slow query, the answer is resizing, and multi-cluster is the distractor.
- All clusters in one multi-cluster warehouse are the same size. You can't mix a Small and a Large.
- Minimum edition: Enterprise. A Standard-edition scenario asking for multi-cluster is really an edition question.
"Size for weight, clusters for crowd." A heavy query needs a bigger warehouse. A crowd of users needs more clusters. And the policy names mean what they say: Standard is the standard user experience (no waiting), Economy is the economy fare (you might wait at the gate).
- Clusters are all the same size, Enterprise edition minimum. The long-standing cap of 10 clusters still applies in Snowsight and for 4XL and larger; since February 2025, SQL allows more on smaller sizes (up to 300 for XS to Medium). min = max is Maximized, min < max is Auto-scale.
- Standard starts a cluster the moment anything queues; Economy starts one only for at least 6 minutes of estimated work and retires clusters when less than 6 minutes remains.
- Billing is per running cluster: a Medium on 3 clusters burns 12 credits per hour, and each cluster suspends independently in auto-scale.

