Snowflake "warehouse overloaded" what to do

TL;DR

Check whether queries are queued (too much concurrency) or just slow (too little compute), then scale out with multi-cluster for queuing or scale up a size for slow single queries. An overloaded warehouse is a capacity problem with two different shapes, and the fix depends on which one you have.

Snowflake "warehouse overloaded" what to do

Use this when

• Queries sit queued instead of executing
• You get warehouse overload or queuing messages
• You need to choose between a bigger warehouse and more clusters

Not for this skill when

• Queries fail with SQL syntax or permission errors
• The question is purely about credit costs
• One specific query is slow but nothing is queuing

Steps

1. Confirm queuing in the query history. Queued means overloaded; slow-but-running means undersized:

SELECT query_text, warehouse_name, execution_status,
       queued_overload_time, total_elapsed_time
FROM snowflake.account_usage.query_history
WHERE start_time > DATEADD(hour, -2, CURRENT_TIMESTAMP())
  AND queued_overload_time > 0
ORDER BY queued_overload_time DESC
LIMIT 20;

Expected output: the queries that waited, and how long each waited. If this returns rows, you have a concurrency problem.

1. Check how loaded the warehouse actually was:

SELECT warehouse_name,
       AVG(avg_running) AS avg_running,
       AVG(avg_queued_load) AS avg_queued
FROM snowflake.account_usage.warehouse_load_history
WHERE start_time > DATEADD(hour, -2, CURRENT_TIMESTAMP())
GROUP BY warehouse_name;

Expected output: average running vs queued load per warehouse. Sustained queuing with running at the max means you need more capacity, not better queries.

1. For queuing under concurrency, enable multi-cluster (scale out):

ALTER WAREHOUSE bi_warehouse SET warehouse_size = 'MEDIUM'
  MIN_CLUSTER_COUNT = 1 MAX_CLUSTER_COUNT = 3
  SCALING_POLICY = 'STANDARD';

Expected output: Snowflake spins up extra clusters automatically when queries queue. Scale-out adds parallel capacity, which is what concurrency needs.

1. For a single heavy query that is slow on its own, scale up one size and test:

ALTER WAREHOUSE etl_warehouse SET warehouse_size = 'LARGE';
-- rerun the heavy query, compare elapsed time, then consider scaling back

Expected output: the query runs faster on more compute per node. Scale-up helps one big query; it does almost nothing for fifty small queued ones.

1. Separate the workloads so they stop fighting. ETL and BI on one warehouse is the usual overload story:

CREATE WAREHOUSE etl_wh WITH warehouse_size = 'LARGE';
CREATE WAREHOUSE bi_wh WITH warehouse_size = 'MEDIUM'
  MIN_CLUSTER_COUNT = 1 MAX_CLUSTER_COUNT = 2;
-- point ETL jobs at etl_wh, dashboards at bi_wh

Expected output: the nightly ETL spike no longer queues the morning dashboard rush. Isolation beats sizing when the contention is between workload types.

Variant phrasings

snowflake queries queued warehouse busy

Step 1 confirms it, step 3 fixes it. Queued overload time is the metric that proves overload vs slowness.

snowflake scale up vs scale out

Up (bigger size) for single-query speed, out (multi-cluster) for concurrent-query throughput. Steps 3-4 are the decision in action.

warehouse size XSMALL overloaded

An XSMALL has very little compute; even modest concurrency queues on it. Medium is the usual floor for shared BI workloads.

Why it happens

A Snowflake warehouse is a fixed pool of compute: one cluster runs a limited number of queries in parallel, and the rest queue. Overload arrives two ways, too many queries at once (concurrency) or queries too big for the cluster size (compute). The engine queues rather than fails, so overload shows up as latency first and errors only when queue timeouts hit.

Edge cases

• Auto-suspend plus a flood of queries causes a cold-start queue spike; a minimum cluster count of 1 avoids the worst of it for latency-sensitive dashboards.
• One runaway query can hog a whole cluster; set STATEMENTQUEUEDTIMEOUTINSECONDS so bad queries fail instead of blocking everyone.
• Multi-cluster in ECONOMY mode only scales for queued queries, STANDARD also scales for anticipated load; pick based on how bursty the traffic is.
• Credits: scaling up or out costs more while it runs. Auto-suspend aggressively on the new clusters so idle capacity doesnt burn credits.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstdt5wZP8PTUgC4eZhV5lZg