instruments-analyzer（仪器分析仪）

Instruments Analyzer

This tool gives you programmatic access to Apple Instruments trace data. Instruments is normally a GUI tool — this tool bridges the gap by exporting .trace files into DuckDB, where you can query them with SQL.

When to use this tool

• You have (or want to record) an Instruments .trace file
• You want to analyze performance data: CPU profiling, hitches, hangs, signposts, Core Animation, SwiftUI updates, RunLoop activity, etc.
• You want root-cause analysis backed by frame-level or event-level evidence

For scroll & animation jank diagnosis: See scroll_and_animation.md — a frame-first workflow for isolating interaction windows, ranking dropped frames, cascade analysis, per-frame attribution, and producing a prioritized fix plan.

Workflow overview

1. Record a trace (or use an existing one)
2. Export the trace to DuckDB
3. Explore the exported tables
4. Prepare derived views (optional, for frame-level analysis)
5. Analyze using SQL queries against the DuckDB database

Step 1: Record a trace

Use xctrace to record from the command line:

# Attach to a running app
xcrun xctrace record --template 'SwiftUI' --time-limit 20s \
  --output ./traces/recording.trace \
  --attach AppName --no-prompt

# Or launch the app
xcrun xctrace record --template 'SwiftUI' --time-limit 20s \
  --output ./traces/recording.trace \
  --launch /path/to/App.app --no-prompt

You can also use the included PerfDebugging.tracetemplate in Instruments.

Choosing a template

• SwiftUI: SwiftUI view updates, hitches, Core Animation, signposts
• Time Profiler: CPU sampling with backtraces
• Animation Hitches: Frame lifetimes, hitch detection
• Custom: Combine instruments as needed

Use xcrun xctrace list templates to see available templates.

Step 2: Export to DuckDB

The export script converts an Instruments .trace file into a DuckDB database with Parquet backing:

./scripts/export_to_duckdb.py traces/recording.trace traces/recording/analysis.duckdb

The script:

• Uses uv run via shebang — run it directly (not with python3)
• Requires uv
• Creates any missing parent directories for the output path
• Exports each Instruments table as a compressed Parquet file
• Creates a DuckDB database with views referencing the Parquet files

If key tables are empty after export, recommend a different Instruments template or a longer recording.

Step 3: Explore the exported tables

After export, connect to the DuckDB database and explore what's available:

-- List all tables/views
SHOW TABLES;

-- Check row counts
SELECT 'updates' AS tbl, COUNT(*) AS rows FROM updates
UNION ALL SELECT 'hitches', COUNT(*) FROM hitches
UNION ALL SELECT 'time_profile', COUNT(*) FROM time_profile
UNION ALL SELECT 'os_signpost_intervals', COUNT(*) FROM os_signpost_intervals
UNION ALL SELECT 'runloop_intervals', COUNT(*) FROM runloop_intervals
UNION ALL SELECT 'potential_hangs', COUNT(*) FROM potential_hangs;

Key tables

Full schema reference: SCHEMAS.md

Common exploration queries

-- Signpost overview (what instrumentation exists)
SELECT
  name, category, subsystem,
  COUNT(*) AS n,
  MAX(CAST(duration_ns AS BIGINT))/1e6 AS max_ms
FROM os_signpost_intervals
WHERE name IS NOT NULL
GROUP BY 1,2,3
ORDER BY max_ms DESC
LIMIT 50;

-- Worst hitches
SELECT start_ns/1e9 AS time_s, duration_ns/1e6 AS ms, narrative_description
FROM hitches
ORDER BY duration_ns DESC
LIMIT 20;

-- Heaviest CPU backtraces
SELECT
  COUNT(*) AS samples,
  SUM(weight_ns)/1e6 AS approx_ms,
  backtrace_json
FROM time_profile
WHERE backtrace_json IS NOT NULL
GROUP BY backtrace_json
ORDER BY approx_ms DESC
LIMIT 10;

-- Hang summary
SELECT hang_type, COUNT(*) AS count, MAX(duration_ns)/1e6 AS max_ms
FROM potential_hangs
GROUP BY hang_type
ORDER BY max_ms DESC;

Step 4: Prepare derived views (for frame analysis)

The prepare_analysis.py script creates analysis-ready views on top of the raw exported data:

./scripts/prepare_analysis.py traces/recording/analysis.duckdb

This creates a frames view from hitches_frame_lifetimes with:

• Computed missed_frames count
• Severity buckets (Low / Medium / High / Extreme)
• Inferred frame budget from coreanimation_lifetime_intervals.acceptable_latency_ns
• Falls back to 60fps (16.67ms) if unavailable

Override for 120fps displays:

./scripts/prepare_analysis.py traces/recording/analysis.duckdb --budget-ms 8.33

Frame-specific cascade analysis

Analyze a specific frame with surrounding context:

./scripts/prepare_analysis.py traces/recording/analysis.duckdb --swap-id 166871
./scripts/prepare_analysis.py traces/recording/analysis.duckdb --swap-id 166871 --context-frames 10

This outputs:

• Target frame details
• Preceding frames with budget status
• Root cause identification (first over-budget frame in a cascade)
• Signposts and logs during the root cause frame

Time units

• All timestamps and durations are in nanoseconds
• start_ns is relative to trace start (not wall clock)
• Convert: duration_ns / 1e6 for milliseconds, start_ns / 1e9 for seconds
• os_signpost_* timestamps are strings — cast to BIGINT when comparing

Use-case companion resources

• scroll_and_animation.md — SwiftUI scroll and animation jank diagnosis