< blog
12 min read

Snowflake Multi-Cluster Warehouses: Scaling Policies, Queuing and Cost (2026)

Illustration of a queue of spheres flowing into a row of identical server blocks, each extra block topped with a gold coin, the largest block in coral

A Snowflake multi-cluster warehouse runs extra clusters of the same size when many queries arrive at once, so they do not have to queue. Each running cluster bills at the full rate for its size, so a multi-cluster warehouse can cost two, three or ten times as much as a single cluster while the extra clusters run. Standard, the default scaling policy, starts an extra cluster as soon as a query queues; Economy starts one only when Snowflake estimates it will have at least six minutes of work, which costs less but lets some queries wait. Queued queries carry no charge of their own, but every running cluster bills for every second it is up. Each time the warehouse starts or resumes, Snowflake bills at least 60 seconds. This guide explains how multi-cluster warehouses scale, what the Standard and Economy policies do, how to tell whether queuing calls for more clusters or a bigger warehouse, and how to keep the cost in check. Every fact and figure in this guide was checked against Snowflake’s documentation.

This article is part of our guide to Snowflake warehouse optimization.

What is a multi-cluster warehouse?

A multi-cluster warehouse is a warehouse that can run more than one cluster of compute at the same time. All clusters are the same size. Snowflake starts extra clusters when queries start to queue and stops them when the load drops. Multi-cluster warehouses need Enterprise Edition or higher.

Two settings decide how many clusters can run:

  • MIN_CLUSTER_COUNT: the clusters that run whenever the warehouse runs.
  • MAX_CLUSTER_COUNT: the most clusters the warehouse can run at once.

Both default to 1, which is a normal single-cluster warehouse. When the two are equal and above 1, the warehouse runs in Maximized mode: all clusters run whenever it runs. When they differ, it runs in Auto-scale mode: Snowflake adds and removes clusters between the two as load changes. Snowflake suggests starting small, in Auto-scale mode with a minimum of 1 and a maximum of 2 or 3.

The highest allowed maximum depends on size: 300 clusters 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. Snowsight lets you choose up to 10; higher values need SQL.

By default, Snowflake aims for about 8 concurrent statements per cluster (MAX_CONCURRENCY_LEVEL); the real number depends on how heavy the queries are. When that level is reached, a single-cluster or Maximized warehouse queues the extra queries, while an Auto-scale warehouse starts another cluster.

How much does a multi-cluster warehouse cost?

Each running cluster bills at the full rate for its size, for every second it runs. A Large warehouse uses 8 credits per hour on Gen1, so:

Clusters running Credits per hour (Large, Gen1) Per hour at $3.00 per credit
1 8 $24
2 16 $48
3 24 $72
4 32 $96

$3.00 is the Enterprise Edition on-demand price in AWS US East in Snowflake’s Service Consumption Table; Enterprise is the lowest edition with multi-cluster warehouses.

Snowflake’s own example: a 3X-Large warehouse (64 credits per hour) that runs one cluster for an hour and two clusters for the next hour uses 64 + 128 = 192 credits.

Because the extra clusters only run while there is load, the real cost depends on how long they stay up, and that is mostly decided by the scaling policy.

Standard vs Economy: which scaling policy should you use?

The scaling policy decides how eagerly Snowflake starts extra clusters and how quickly it stops them.

Standard (default) Economy
Starts a cluster As soon as a query queues, or Snowflake estimates the running clusters cannot take more queries Only if Snowflake estimates there is enough load to keep the new cluster busy for at least 6 minutes
Stops a cluster After a sustained period of low load, the least-loaded clusters shut down once their queries finish When Snowflake estimates the least-loaded cluster has less than 6 minutes of work left
Favors Little or no queuing Lower cost

With a maximum of 10 clusters or fewer, clusters start and stop one at a time.

Use Standard where queuing hurts: user-facing dashboards and interactive work. Use Economy where some queuing is acceptable and cost matters more: batch jobs and background workloads.

My queries are queuing: should I resize or add clusters?

It depends on why they queue.

  • Many queries at once, each reasonably fast: add clusters. More clusters let more queries run at the same time. Snowflake’s guidance is that resizing is not intended for handling concurrency; use a multi-cluster warehouse or more warehouses instead.
  • A few big queries that run slowly or spill to disk: resize. A bigger warehouse gives each query more compute and memory; more clusters would not make a single query faster.

To tell them apart, look at how much time queries spend waiting and why:

-- Queuing and run time per warehouse, last 7 days
SELECT warehouse_name,
       COUNT(*)                            AS queries,
       COUNT_IF(queued_overload_time > 0)  AS queued_queries,
       AVG(queued_overload_time) / 1000    AS avg_queued_seconds,
       AVG(execution_time) / 1000          AS avg_execution_seconds
FROM snowflake.account_usage.query_history
WHERE start_time >= DATEADD(day, -7, CURRENT_TIMESTAMP())
  AND warehouse_name IS NOT NULL
GROUP BY 1
ORDER BY avg_queued_seconds DESC;

QUEUED_OVERLOAD_TIME is time spent waiting because the warehouse was busy, which more clusters fix. QUEUED_PROVISIONING_TIME is time spent waiting for the warehouse to start or resize. Both are in milliseconds. For a warehouse-level view, WAREHOUSE_LOAD_HISTORY shows average running and queued load (AVG_RUNNING, AVG_QUEUED_LOAD) in 5-minute intervals. Both are load ratios, the time queries spent in that state divided by the length of the interval, not counts of queries or clusters.

Read the two numbers together. High avg_queued_seconds with low avg_execution_seconds is a concurrency problem: the queries are fine, there are simply too many at once, so add clusters. High avg_execution_seconds with little queuing is a size problem, so resize. Queuing that is almost all QUEUED_PROVISIONING_TIME is neither, and usually means the warehouse is suspending and resuming too often.

Snowflake does not charge per queued query. A queued query adds no charge of its own, but the warehouse keeps billing while it runs, whether or not it is doing work. If queries should not wait forever, STATEMENT_QUEUED_TIMEOUT_IN_SECONDS cancels statements that queue too long; it is 0, meaning no timeout, by default.

How to control multi-cluster costs: six warehouse actions

  • Set MAX_CLUSTER_COUNT from real demand. Look at how many clusters actually run at peak, and cap the warehouse there rather than at a round number.
  • Keep MIN_CLUSTER_COUNT at 1 unless you need guaranteed capacity. Every cluster in the minimum runs whenever the warehouse runs.
  • Use Economy for workloads that can wait.
  • Split workloads that do not belong together. A dashboard warehouse and a batch warehouse have different needs; one multi-cluster warehouse serving both is usually sized for the worse case.
  • Check Query Acceleration. New multi-cluster warehouses get it turned on by default, and it bills whenever it is used.
  • Watch the tail of each burst. Clusters that stay up after the load has gone are the main waste in multi-cluster warehouses.

None of that is visible by default. Snowflake bills by warehouse, not by cluster, so there is no per-cluster line anywhere in your usage: the tail of a burst shows up only as credits that bought very little work. Compare credits against the load that earned them:

-- Credits per hour against the load that earned them, last 7 days
SELECT m.warehouse_name,
       m.start_time                      AS hour,
       ROUND(m.credits_used, 2)          AS credits,
       ROUND(AVG(l.avg_running), 2)      AS avg_running_load,
       ROUND(AVG(l.avg_queued_load), 2)  AS avg_queued_load
FROM snowflake.account_usage.warehouse_metering_history m
JOIN snowflake.account_usage.warehouse_load_history l
  ON l.warehouse_name = m.warehouse_name
 AND l.start_time >= m.start_time
 AND l.start_time < m.end_time
WHERE m.start_time >= DATEADD(day, -7, CURRENT_TIMESTAMP())
GROUP BY 1, 2, 3
ORDER BY credits DESC;

Hours with high credits and a low running load are the burst tails.

-- A multi-cluster warehouse with a cost-conscious setup
CREATE WAREHOUSE batch_wh
  WAREHOUSE_SIZE = MEDIUM
  MIN_CLUSTER_COUNT = 1
  MAX_CLUSTER_COUNT = 3
  SCALING_POLICY = 'ECONOMY'
  AUTO_SUSPEND = 60
  AUTO_RESUME = TRUE;

-- Change the policy or the cap on an existing warehouse
ALTER WAREHOUSE batch_wh SET SCALING_POLICY = 'STANDARD' MAX_CLUSTER_COUNT = 4;

Multi-cluster warehouses are not Automatic Clustering

Two Snowflake features share the word cluster and neither bill touches the other. A multi-cluster warehouse runs extra copies of compute so queries do not queue, and bills as warehouse compute at the size you chose. Automatic Clustering reorganizes the micro-partitions of a table so queries can skip more of it, and bills on its own: Clustering Classic bills serverless compute, while Optima Clustering bills on the volume of uncompressed data ingested. Capping clusters on a warehouse does nothing to an Automatic Clustering bill, and dropping a clustering key does nothing to a warehouse bill.

To see what reclustering costs, read AUTOMATIC_CLUSTERING_HISTORY. VERSION tells you whether the work came from Clustering Classic or Optima Clustering, and the view can lag by up to 3 hours.

-- What Automatic Clustering costs by table, last 30 days
SELECT table_name,
       version,
       ROUND(SUM(credits_used), 2) AS credits,
       SUM(num_bytes_reclustered)  AS bytes_reclustered
FROM snowflake.account_usage.automatic_clustering_history
WHERE start_time >= DATEADD(day, -30, CURRENT_TIMESTAMP())
GROUP BY 1, 2
ORDER BY credits DESC;

Suspending reclustering only pauses the charges: on Optima Clustering, resuming later adds a catch-up cost for the data ingested in the meantime. To stop paying for a table altogether, drop its clustering key. Our guide to Snowflake cost optimization goes further into when a clustering key earns its cost.

How Seemore keeps extra clusters from idling

Snowflake’s scaling policies are built to protect performance, not cost, so extra clusters can keep running after the load that justified them is gone. Seemore’s Auto-Scaler watches multi-cluster warehouses in real time and scales cluster count down the moment extra clusters stop contributing meaningful performance. It pauses scale-in while any queries are queued, considers a cluster for scale-in only after it has been up for at least a minute, and never goes below the minimum or above the maximum you set. SmartPulse builds an hourly schedule of sizes and generations and applies it once you approve it, so concurrency and size are handled together.

Snowflake’s own answer to the same problem is the adaptive warehouse, generally available on Enterprise Edition and higher in selected regions. An adaptive warehouse has no size, no cluster count, no Query Acceleration setting and no suspend timer: Snowflake allocates compute per query and bills per query. That removes the settings rather than tuning them, and it does not yet cover X5LARGE and X6LARGE warehouses, or Snowpark-optimized and interactive ones. Auto-Scaler applies to standard multi-cluster warehouses, which keep their size, cluster count and scaling policy.

Frequently asked questions

What is a multi-cluster warehouse in Snowflake?

A warehouse that can run several clusters of the same size at once to handle more concurrent queries. It needs Enterprise Edition or higher.

How are Snowflake multi-cluster warehouses billed?

Each running cluster bills at the full rate for the warehouse’s size. Three Large clusters running for an hour use 3 x 8 = 24 credits on Gen1.

What is the difference between Standard and Economy scaling in Snowflake?

Standard starts a cluster as soon as queries queue, favoring speed. Economy starts one only when there is enough work to keep it busy for at least 6 minutes and shuts clusters down sooner, so it queues more and costs less.

Should I resize a Snowflake warehouse or add clusters?

Add clusters when many queries queue at the same time; resize when individual queries are slow or spill to disk.

What is the default cluster count for Snowflake multi-cluster warehouses?

One. MIN_CLUSTER_COUNT and MAX_CLUSTER_COUNT both default to 1, which is a normal single-cluster warehouse.

Do queued queries cost money in Snowflake?

Not on their own. Snowflake bills the warehouse while it runs, not each query, so queued queries add no separate charge, but the warehouse keeps billing while they wait.