Understanding and Optimizing Snowflake Cloud Services Billing

Introduction

When reviewing Snowflake usage statements, you’ll notice credit consumption recorded under “Cloud Services” separately from warehouse compute. This article covers what Cloud Services are, under what conditions they are billed, and how to track and reduce them.

Official documentation: Understanding compute cost / Optimizing cloud services for cost

What Is the Cloud Services Layer?

Snowflake’s architecture is broadly divided into three layers.

┌────────────────────────────────────────────┐
│        Cloud Services Layer                │  ← Focus of this article
│  Authentication / Metadata management /    │
│  Query compilation / Access control /      │
│  Transaction management                    │
├────────────────────────────────────────────┤
│        Compute Layer (Virtual Warehouses)  │  ← Query & DML execution
├────────────────────────────────────────────┤
│        Storage Layer                       │  ← Data persistence
└────────────────────────────────────────────┘

The Cloud Services layer is a component that Snowflake keeps running internally at all times, even when users have not explicitly started a warehouse. Its main responsibilities are as follows.

The 10% Rule (Cloud Services Adjustment)

How Billing Works

Cloud Services credit consumption always occurs, but it is only billed (charged) when it exceeds a certain threshold.

Cloud Services consumption is only billed when it exceeds 10% of the virtual warehouse consumption for the same day (UTC). Only the excess amount is charged.

This mechanism where only the excess is billed is called the “Cloud Services Adjustment.” The CREDITS_ADJUSTMENT_CLOUD_SERVICES column in the METERING_DAILY_HISTORY view records a negative value, representing the adjusted (deducted) amount.

Calculation Examples

Here are specific examples from the official documentation.

• Nov 1: CS=20 > VW×10%=10 → 10 excess credits are billed. Adjustment is −10
• Nov 2: CS=10 < VW×10%=12 → CS is under 10% → Adjustment is −10 (full CS deduction)
• Nov 3: CS=5 < VW×10%=8 → Full adjustment. Adjustment is −5 (full CS deduction)
• Nov 4: CS=13 > VW×10%=10 → 3 excess credits are billed. Adjustment is −10

Use Cases That Generate Cloud Services Costs

The following patterns all consume Cloud Services resources without using a virtual warehouse.

How to Track Cloud Services Costs

1. Daily Billing Summary (METERING_DAILY_HISTORY)

The SNOWFLAKE.ACCOUNT_USAGE.METERING_DAILY_HISTORY view shows daily CS usage and the actual billed amount after adjustments.

-- CS billing summary for the past month (ordered by highest billed amount)
SELECT
    usage_date,
    credits_used_cloud_services,
    credits_adjustment_cloud_services,
    credits_used_cloud_services + credits_adjustment_cloud_services AS billed_cloud_services
FROM snowflake.account_usage.metering_daily_history
WHERE usage_date >= DATEADD(month, -1, CURRENT_TIMESTAMP())
    AND credits_used_cloud_services > 0
ORDER BY 4 DESC;

Example output:

Rows where billed_cloud_services is 0 indicate that the day’s CS consumption stayed within 10% of VW usage and was not billed.

Key column definitions:

2. Per-Warehouse Analysis (WAREHOUSE_METERING_HISTORY)

Identifying warehouses with a high CS ratio helps narrow down problematic workloads.

-- Identify warehouses with high CS ratios
SELECT
    warehouse_name,
    SUM(credits_used)                AS credits_used,
    SUM(credits_used_cloud_services) AS credits_used_cloud_services,
    ROUND(
        SUM(credits_used_cloud_services) / NULLIF(SUM(credits_used), 0) * 100,
        2
    )                                AS pct_cloud_services
FROM snowflake.account_usage.warehouse_metering_history
WHERE TO_DATE(start_time) >= DATEADD(month, -1, CURRENT_TIMESTAMP())
    AND credits_used_cloud_services > 0
GROUP BY 1
ORDER BY 4 DESC;

Example output:

Warehouses where PCT_CLOUD_SERVICES significantly exceeds 10% likely have workloads matching the patterns described in the previous section.

3. Per-Query-Type Analysis (QUERY_HISTORY)

Check which types of operations are consuming the most CS credits.

-- CS consumption by query type over the past 7 days
SELECT
    query_type,
    ROUND(SUM(credits_used_cloud_services), 6) AS cs_credits,
    COUNT(1)                                   AS num_queries
FROM snowflake.account_usage.query_history
WHERE start_time >= TIMESTAMPADD(day, -7, CURRENT_TIMESTAMP)
GROUP BY 1
ORDER BY 2 DESC
LIMIT 10;

Example output:

If SHOW queries have a high count and significant CS consumption, it’s likely that third-party tools are frequently fetching metadata.

4. Checking in Snowsight

You can visually review costs from Admin > Cost Management in Snowsight. It supports filtering by service type and time period, making it useful for understanding trends without writing queries.

Additional Notes

Differences Between INFORMATION_SCHEMA and ACCOUNT_USAGE

When writing tracking queries, be aware that the choice of schema affects behavior.

For monitoring purposes, using ACCOUNT_USAGE is often more cost-effective.

Notes on METERING_DAILY_HISTORY

• View latency is up to 180 minutes (3 hours)
• The CREDITS_USED_CLOUD_SERVICES column has up to 6 hours of latency (on the WAREHOUSE_METERING_HISTORY side)
• When reconciling with monthly billing, set the session timezone to UTC before querying

ALTER SESSION SET TIMEZONE = UTC;

Cost Reduction Priority

Listed in order of typical effectiveness:

1. Switch from INFORMATION_SCHEMA to ACCOUNT_USAGE — Can be addressed with code changes alone
2. Reduce frequent SHOW commands — Can be addressed by reviewing tool configurations
3. Convert single-row INSERTs to bulk loading — Requires architectural changes but yields significant results
4. Restructure S3 buckets for COPY commands — Can be addressed incrementally

Summary

• The Cloud Services layer handles authentication, metadata management, and query compilation, consuming credits at all times
• Billing follows the rule that “only the excess beyond 10% of daily VW consumption is charged,” and in many cases the adjustment covers the full amount
• Three views — METERING_DAILY_HISTORY, WAREHOUSE_METERING_HISTORY, and QUERY_HISTORY — allow you to progressively identify the root cause
• When CS costs are high, first investigate “frequent SHOW/INFORMATION_SCHEMA queries” and “single-row INSERTs”

References

Understanding compute cost | Snowflake Documentation

docs.snowflake.com

Optimizing cloud services for cost | Snowflake Documentation

docs.snowflake.com

Exploring compute cost | Snowflake Documentation

docs.snowflake.com

METERING_DAILY_HISTORY view | Snowflake Documentation

docs.snowflake.com