Data Analysis

Answer business questions by querying the data warehouse. The kernel starts automatically on first use.

Prerequisites

uv must be installed:

curl -LsSf https://astral.sh/uv/install.sh | sh

Scripts are located relative to this skill file.

MANDATORY FIRST STEP

Before any other action, check for cached patterns:

uv run scripts/cli.py pattern lookup "<user's question>"

This is NON-NEGOTIABLE. Patterns contain proven strategies that save time and avoid failed queries.

Workflow

Analysis Progress:
- [ ] Step 1: pattern lookup (check for cached strategy)
- [ ] Step 2: concept lookup (check for known tables)
- [ ] Step 3: Search codebase for table definitions (Grep)
- [ ] Step 4: Read SQL file to get table/column names
- [ ] Step 5: Execute query via kernel (run_sql)
- [ ] Step 6: learn_concept (ALWAYS before presenting results)
- [ ] Step 7: learn_pattern (ALWAYS if discovery required)
- [ ] Step 8: record_pattern_outcome (if you used a pattern in Step 1)
- [ ] Step 9: Present findings to user

CLI Commands

Kernel Management

uv run scripts/cli.py start           # Start kernel with Snowflake
uv run scripts/cli.py exec "..."      # Execute Python code
uv run scripts/cli.py status          # Check kernel status
uv run scripts/cli.py restart         # Restart kernel
uv run scripts/cli.py stop            # Stop kernel
uv run scripts/cli.py install plotly  # Install additional packages

Concept Cache (concept -> table mappings)

# Look up a concept
uv run scripts/cli.py concept lookup customers

# Learn a new concept
uv run scripts/cli.py concept learn customers HQ.MART_CUST.CURRENT_ASTRO_CUSTS -k ACCT_ID

# List all concepts
uv run scripts/cli.py concept list

# Import concepts from warehouse.md
uv run scripts/cli.py concept import -p /path/to/warehouse.md

Pattern Cache (query strategies)

# Look up patterns for a question
uv run scripts/cli.py pattern lookup "who uses operator X"

# Learn a new pattern
uv run scripts/cli.py pattern learn operator_usage \
    -q "who uses X operator" \
    -q "which customers use X" \
    -s "1. Query TASK_RUNS for operator_class" \
    -s "2. Join with ORGS on org_id" \
    -t "HQ.MODEL_ASTRO.TASK_RUNS" \
    -t "HQ.MODEL_ASTRO.ORGANIZATIONS" \
    -g "TASK_RUNS is huge - always filter by date"

# Record pattern outcome
uv run scripts/cli.py pattern record operator_usage --success

# List all patterns
uv run scripts/cli.py pattern list

# Delete a pattern
uv run scripts/cli.py pattern delete operator_usage

Table Schema Cache

# Look up cached table schema
uv run scripts/cli.py table lookup HQ.MART_CUST.CURRENT_ASTRO_CUSTS

# Cache a table schema
uv run scripts/cli.py table cache DB.SCHEMA.TABLE -c '[{"name":"id","type":"INT"}]'

# List all cached tables
uv run scripts/cli.py table list

# Delete from cache
uv run scripts/cli.py table delete DB.SCHEMA.TABLE

Cache Management

# View cache statistics
uv run scripts/cli.py cache status

# Clear all caches
uv run scripts/cli.py cache clear

# Clear only stale entries (older than 90 days)
uv run scripts/cli.py cache clear --stale-only

Quick Start Example

# 1. Check for existing patterns
uv run scripts/cli.py pattern lookup "how many customers"

# 2. Check for known concepts
uv run scripts/cli.py concept lookup customers

# 3. Execute query
uv run scripts/cli.py exec "df = run_sql('SELECT COUNT(*) FROM HQ.MART_CUST.CURRENT_ASTRO_CUSTS')"
uv run scripts/cli.py exec "print(df)"

# 4. Cache what we learned
uv run scripts/cli.py concept learn customers HQ.MART_CUST.CURRENT_ASTRO_CUSTS -k ACCT_ID

Available Functions in Kernel

Once kernel starts, these are available:

run_sql(query, limit=100)

run_sql_pandas(query, limit=100)

pl

pd

Table Discovery via Codebase

If concept/pattern cache miss, search the codebase:

Grep pattern="<concept>" glob="**/*.sql"

dags/declarative/04_metric/

06_reporting/

05_mart/

models/marts/

models/staging/

Known Tables Quick Reference

Large tables (always filter by date): TASK_RUNS (6B rows), DAG_RUNS (500M rows)

Query Tips

• Use LIMIT during exploration
• Filter early with WHERE clauses
• Prefer pre-aggregated tables (

  METRICS_*

  ,

  MART_*

  ,

  AGG_*

  )
• For 100M+ row tables: no JOINs or GROUP BY on first query

Reference

• reference/discovery-warehouse.md - Large table handling, warehouse discovery