ARA-C01 Question 238
Select 2CostsA retail company runs a Snowflake-based analytics platform with three major workloads: 1) hourly ELT pipelines, 2) business intelligence dashboards used during business hours, and 3) ad hoc data science queries. Finance has reported that Snowflake compute costs increased sharply over the last two months, even though data volume and user counts remained relatively stable. An architect reviews usage and finds that a single large warehouse is shared by all workloads, auto-suspend is set to 30 minutes, and analysts frequently complain about inconsistent dashboard performance during ELT windows. The company wants to reduce compute cost without materially affecting user experience. Which TWO actions should the architect recommend?
- A
Split the shared warehouse into separate warehouses for ELT, BI, and ad hoc workloads, and size each warehouse based on its workload characteristics
- B
Reduce the auto-suspend setting from 30 minutes to a much shorter interval appropriate for the workload patterns
- C
Enable a multi-cluster warehouse for all workloads on the existing shared warehouse so queries finish faster and overall credits decrease automatically
- D
Increase the size of the existing shared warehouse so queued queries complete sooner, which will consistently reduce total compute spend
- E
Move all workloads to Snowpark-optimized warehouses because they are designed to lower cost for mixed analytics workloads
Show answer and explanation
Correct answers: A, B
Explanation
The best answer is to isolate workloads and reduce idle warehouse runtime. In Snowflake, compute cost is driven largely by warehouse size, runtime duration, and scaling behavior. When multiple workload types share one warehouse, teams often oversize it to satisfy peak demand, which increases credits and introduces contention. Separating ELT, BI, and ad hoc workloads allows each warehouse to be sized, scheduled, and governed independently. In addition, shortening auto-suspend reduces credits spent on idle warehouses, which is a common source of waste.
Snowflake best practices emphasize using separate warehouses for separate workloads, enabling right-sizing and avoiding interference. They also recommend aggressive auto-suspend settings where feasible because warehouses accrue charges while running, even when not actively processing queries. Multi-cluster warehouses can be valuable for concurrency-sensitive workloads, but they should be used selectively and with an understanding that scaling out can increase cost. Similarly, larger warehouses do not inherently reduce spend, and specialized warehouse types such as Snowpark-optimized warehouses should be chosen only for appropriate workload patterns.
Relevant Snowflake documentation and guidance include virtual warehouse considerations, auto-suspend/auto-resume behavior, warehouse sizing, workload isolation, and cost optimization best practices in Snowflake's performance and cost management documentation.
- A. Correct.
Correct. Isolating workloads onto separate virtual warehouses is a core Snowflake cost and performance best practice. Different workloads have different concurrency, latency, and runtime profiles. ELT often benefits from dedicated sizing and scheduling, BI requires predictable response times, and ad hoc data science workloads are bursty and less predictable. Separating them prevents resource contention, improves performance consistency, and enables independent right-sizing, scheduling, and suspension behavior. This often reduces unnecessary overprovisioning caused by trying to satisfy all workloads with one oversized shared warehouse.
- B. Correct.
Correct. Auto-suspend at 30 minutes often leads to substantial idle-credit consumption, especially for intermittent BI or ad hoc workloads. Snowflake charges warehouse compute while the warehouse is running, not only while queries are executing. Lowering auto-suspend to a shorter interval, such as a few minutes or less where operationally appropriate, can significantly reduce idle time while preserving acceptable user experience. This is one of the most common and effective levers for lowering compute costs.
- C. Incorrect.
Incorrect. Multi-cluster warehouses are primarily used to handle concurrency for a single workload pattern, especially to reduce queuing during spikes. They do not automatically lower cost; in fact, additional clusters can increase compute consumption. Using multi-cluster on a single shared warehouse may improve concurrency, but it does not address the root issue of mixed workloads and long idle periods. It is a performance-scaling feature, not a universal cost-optimization mechanism.
- D. Incorrect.
Incorrect. Increasing warehouse size may reduce elapsed runtime for some queries, but it does not consistently reduce overall spend. In this scenario, the warehouse is already shared across dissimilar workloads and remains running for long idle windows due to the 30-minute auto-suspend. A larger warehouse can increase credit consumption per unit of time and may worsen cost if the root causes are poor workload isolation and idle runtime. Right-sizing should be workload-specific, not a blanket increase.
- E. Incorrect.
Incorrect. Snowpark-optimized warehouses are intended for memory-intensive Snowpark workloads, such as certain machine learning or complex DataFrame processing use cases. They are not a general cost-reduction option for mixed BI, ELT, and ad hoc SQL workloads. Moving all workloads to Snowpark-optimized warehouses would likely increase cost and misalign warehouse type to the actual workload characteristics.