Power BI Diagnostics Skill

Troubleshoot performance, trace queries, and verify the pbi-cli environment.

Prerequisites

pipx install pbi-cli-tool
pbi-cli skills install
pbi connect

Environment Check

# Verify pythonnet and .NET DLLs are installed
pbi setup

# Show detailed environment info (version, DLL paths, pythonnet status)
pbi setup --info
pbi --json setup --info

# Check CLI version
pbi --version

Quick Troubleshooting

If pbi-cli isn't working, run these checks in order:

# 1. Is pbi-cli installed correctly?
pbi --version
pbi setup --info

# 2. Is Power BI Desktop running with a model open?
pbi connect

# 3. Is the connection still alive?
pbi connections last

# 4. Can you query the model?
pbi dax execute "EVALUATE ROW(\"test\", 1)"

Model Health Check

# Quick model overview
pbi --json model get

# Object counts (tables, columns, measures, relationships, partitions)
pbi --json model stats

# List all tables with column/measure counts
pbi --json table list

Query Tracing

Capture diagnostic events during DAX query execution:

# Start a trace
pbi trace start

# Execute the query you want to profile
pbi dax execute "EVALUATE SUMMARIZECOLUMNS(Products[Category], \"Total\", SUM(Sales[Amount]))"

# Stop the trace
pbi trace stop

# Fetch captured trace events
pbi --json trace fetch

# Export trace events to a file
pbi trace export ./trace-output.json

Cache Management

# Clear the formula engine cache (do this before benchmarking)
pbi dax clear-cache

Connection Diagnostics

# List all saved connections
pbi connections list
pbi --json connections list

# Show the last-used connection
pbi connections last

# Reconnect to a specific data source
pbi connect -d localhost:54321

# Disconnect
pbi disconnect

Workflow: Profile a Slow Query

# 1. Clear cache for a clean benchmark
pbi dax clear-cache

# 2. Start tracing
pbi trace start

# 3. Run the slow query
pbi dax execute "EVALUATE SUMMARIZECOLUMNS(Products[Category], \"Total\", SUM(Sales[Amount]))" --timeout 300

# 4. Stop tracing
pbi trace stop

# 5. Export trace for analysis
pbi trace export ./slow-query-trace.json

# 6. Review trace events
pbi --json trace fetch

Workflow: Model Health Audit

# 1. Model overview
pbi --json model get
pbi --json model stats

# 2. Check table sizes and structure
pbi --json table list

# 3. Review relationships
pbi --json relationship list

# 4. Check security roles
pbi --json security-role list

# 5. Export full model for offline review
pbi database export-tmdl ./audit-export/

Best Practices

• Clear cache before benchmarking: pbi dax clear-cache
• Use --timeout for long-running queries to avoid premature cancellation
• Export traces to files for sharing with teammates
• Run pbi setup --info first when troubleshooting environment issues
• Use --json output for automated monitoring scripts
• Use pbi repl for interactive debugging sessions with persistent connection

Gotchas

• pbi connect succeeds even against a stale port: Desktop reassigns its local Analysis Services port on every restart. Connection from the previous session may bind to a port now occupied by a different model — queries return data from the wrong file. After Desktop restart, always pbi connect again rather than reusing the saved connection.
• pbi setup --info reports pythonnet status, not runtime health: A green pythonnet: ok line means the DLL loaded — it does NOT verify the AdomdClient can talk to your model. Confirm with a smoke query like EVALUATE ROW("test", 1) before assuming the environment is healthy.
• Trace events arrive asynchronously: pbi trace fetch immediately after trace stop may return an empty list. Wait 1-2 seconds or call fetch in a short retry loop — events buffer in the AS server before flushing.
• Clearing formula engine cache leaves storage engine cache intact: Repeat-running a benchmark query after clear-cache still hits cached VertiPaq segments. For a true cold benchmark, restart Desktop or reconnect.
• model stats counts hidden objects too: Calculation groups, auto-date hierarchies, and translation tables inflate the table count vs. what a user sees in the field list. Cross-check with table list filtered on visibility.
• --timeout is in seconds, not milliseconds: Easy slip — --timeout 300 is 5 minutes, not 0.3 seconds. Default is 60s, which silently kills any moderately-complex query on a cold cache.