Analyze Workload Performance with CoCo

The workload-performance-analysis bundled skill in Cortex Code analyzes Snowflake SQL query execution using ACCOUNT_USAGE views. You ask questions in natural language; the skill queries historical performance data, summarizes what it finds, and can suggest next steps.

The skill is built into Cortex Code. You don’t install or configure it separately.

Overview

Use this skill when you need to understand how queries ran and what to tune next: spilling, partition pruning, cache efficiency, clustering, and eligibility for Search Optimization Service (SOS) or Query Acceleration Service (QAS).

The skill is a single entry point for performance analysis. Depending on what you ask, it focuses on a specific query, warehouse, table, query pattern, stored procedure run, a set of queries, or a broad account health scan.

When the skill applies

Cortex Code selects workload-performance-analysis when your request matches performance analysis topics such as:

  • Spilling, memory pressure, or spill to local or remote storage
  • Partition pruning, scan volume, or tables that prune poorly
  • Cache hit rates or warehouse cache efficiency
  • Clustering keys, CLUSTER BY, or predicate columns for clustering or SOS
  • Search optimization or SOS candidates
  • Query Acceleration Service (QAS) or acceleration eligibility
  • Slow SQL diagnosis, bottlenecks, root cause, or concurrency issues
  • Per-warehouse spill, prune, or cache metrics

What this skill is not for

Use other Cortex Code skills for the following topics:

TopicUse instead
Cost, credits, or spend analysiscost-intelligence
Access audit or data governancedata-governance
Writing, editing, or debugging your SQLGeneral Cortex Code assistance outside this skill

What the skill can do

The skill analyzes SQL query performance using ACCOUNT_USAGE views. It can help you:

  • Disk spilling: Find queries with excessive local or remote spilling and suggest fixes.
  • Partition pruning: Review pruning ratios and scan efficiency on large tables.
  • Cache hit rates: Evaluate local disk cache and warehouse cache efficiency.
  • Clustering keys: Recommend clustering keys when pruning is poor.
  • Search Optimization Service (SOS): Identify workloads that may benefit from SOS.
  • Query Acceleration Service (QAS): Find queries that may be eligible for QAS.
  • Predicate analysis: Relate filter columns to clustering or SOS opportunities.
  • Stored procedure runs: Analyze a parent CALL and its child queries as one run.
  • Query sets and patterns: Compare multiple queries, recurring patterns, or filtered workloads.

Results are summarized with top findings first, not raw dumps of every row.

Questions the skill can help answer

  • Which queries are spilling to disk, and should I upsize the warehouse or change the query?
  • Is warehouse ANALYTICS_WH showing spill, prune, or cache problems across recent runs?
  • Are scans on DB.SCHEMA.ORDERS pruning enough partitions?
  • Which recurring query patterns are slowest or most expensive to execute?
  • What happened inside this stored procedure run, including child queries?
  • Are there SOS or QAS candidates in my recent workload?
  • Give me a quick account-level performance health check (spilling, pruning, cache, QAS).

Be specific when you can: a query ID, warehouse name, fully qualified table name, time range, or symptom (for example “remote spilling on ANALYTICS_WH last 7 days”) helps the skill scope the analysis.

How deep the analysis goes

You control depth with how you phrase the request. The following table summarizes typical outcomes:

You ask forWhat you get
Summary, overview, quick look, health checkHigh-level findings; you can ask to go deeper afterward
What’s wrong, root cause, bottlenecks, why is X slowSummary plus issue detection
Recommend, how to fix, optimize, action itemsSummary, detection, and concrete recommendations

If you don’t specify depth, the skill starts with a summary and offers deeper analysis or recommendations.

Where you can use the skill

The skill is available in four places. In Snowsight, select Analyze with CoCo to pass page context to the skill. In Cortex Code CLI, Cortex Code Desktop, or the Cortex Code panel in Snowsight, describe what to analyze or invoke /workload-performance-analysis.

Query Details

From a single query’s detail view in Snowsight, select Analyze with CoCo. Cortex Code receives that query’s context and analyzes that execution (including stored procedure runs when the query is a CALL). For more information, see Review details and profile of a specific query.

Query Details page with Analyze with CoCo selected and Cortex Code summarizing query profile metrics such as partitions scanned and cache hit rate

Query History

From Query History, select Analyze with CoCo when you have one or more queries in view. The skill uses the list context from the page to summarize or compare those queries. For more information, see Query history in Snowsight.

Query History page with Analyze with CoCo selected and Cortex Code summarizing query status and key findings for the visible queries

Performance Explorer

From the Performance Explorer dashboard, select Analyze with CoCo to interpret workload trends, hotspots, or changes over time in the context of the charts and filters you already applied. For more information, see Analyzing query workloads with Performance Explorer.

Performance Explorer dashboard with Analyze with CoCo selected and Cortex Code presenting a health metrics summary and key takeaways

Cortex Code (CLI, Desktop, Snowsight panel)

Open Cortex Code directly and describe what to analyze, or type /workload-performance-analysis to invoke the skill. This path also applies when you open the Cortex Code panel in Snowsight without using a page-specific Analyze action. For more information, see Cortex Code CLI bundled skills.

Cortex Code panel showing the workload-performance-analysis skill description and analysis type options

Snowsight entry points pass page context so you don’t need to copy query IDs or re-enter filters. In CLI, Desktop, or the general Snowsight panel, include the scope in your prompt (warehouse name, query ID, table name, time range, and so on).

Example ways to use the skill

These outcomes match common workloads the skill is designed for:

  • Identify queries with excessive disk spilling and get concrete fix recommendations.
  • Analyze partition pruning ratios and suggest better clustering keys for large tables.
  • Find queries eligible for Search Optimization Service or Query Acceleration Service.

Example prompts:

Summarize spill and cache metrics for warehouse ANALYTICS_WH over the last 7 days.

Why is query 01b24bb0-0007-9627-0000-0001234abcde slow? Root cause and recommendations.

Which tables have the worst partition pruning? Recommend clustering keys for the top offenders.

Are any of my recurring query patterns good QAS candidates?

Health check: spilling, pruning, cache, and QAS eligibility account-wide for the last week.

In Cortex Code CLI, you can also run /workload-performance-analysis and then describe your question.

Prerequisites and access

You need:

  • Access to Cortex Code in your account. See Cortex Code in Snowsight and Cortex Code CLI.

  • Privileges to read the ACCOUNT_USAGE views the skill queries. Many analyses require the SNOWFLAKE.USAGE_VIEWER database role (or related usage or governance roles for specific views). An administrator can grant access, for example:

    GRANT DATABASE ROLE SNOWFLAKE.USAGE_VIEWER TO ROLE <your_role>;

What you can see matches the query and account activity your roles can access in Snowsight and ACCOUNT_USAGE.

Limitations

Keep these constraints in mind:

  • Historical data only: The skill analyzes past runs in ACCOUNT_USAGE; it does not predict future performance.
  • Latency: ACCOUNT_USAGE data can lag behind real time (for example, up to about 45 minutes for query history and longer for some pruning views).
  • Hybrid tables: Visibility in these views can be limited for hybrid tables.
  • Benefit estimates: The skill can flag SOS or clustering candidates but can’t guarantee exact savings from enabling them.

When recommending warehouse sizing or queuing tradeoffs, the skill may ask whether your workload prioritizes speed (minimize wait time and runtime) or cost (accept some queuing or spilling to save credits). That framing helps tailor guidance; for credit and spend analysis, use the cost-intelligence skill instead.