spaa（spaa 搜索）

Analyzing SPAA Files

SPAA (Stack Profile for Agentic Analysis) is an NDJSON file format for representing sampled performance stack traces. Each line is a self-contained JSON object with a type field.

Before analyzing an SPAA file, read the full specification at references/SPEC.md.

Quick Start

1. Examine the header

The first line is always the header. It tells you what profiler generated the data and how to interpret metrics:

head -1 profile.spaa | jq .

Key header fields:

• source_tool: The profiler that generated this data (e.g., "perf", "dtrace")
• frame_order: Either "leaf_to_root" or "root_to_leaf" - determines how to read stack frames
• events[].sampling.primary_metric: The authoritative metric for weighting (e.g., "period" for perf, "samples" for DTrace)

2. Understand the record types

# Count records by type
grep -o '"type":"[^"]*"' profile.spaa | sort | uniq -c | sort -rn

Common types:

• header - File metadata (exactly one, always first)
• dso - Shared libraries and binaries
• frame - Individual stack frame definitions
• thread - Thread/process info
• stack - Aggregated call stacks with weights (the main data)
• sample - Individual sample events (optional, for temporal analysis)

3. Find performance hotspots

Extract the heaviest stacks by the primary metric:

# For perf data (uses "period" metric)
grep '"type":"stack"' profile.spaa | \
  jq -s 'sort_by(-.weights[] | select(.metric=="period") | .value) | .[0:10]'

# For DTrace data (uses "samples" metric)
grep '"type":"stack"' profile.spaa | \
  jq -s 'sort_by(-.weights[] | select(.metric=="samples") | .value) | .[0:10]'

4. Find hot functions (exclusive time)

Exclusive time shows where the CPU actually spent time, not just functions on the call path:

grep '"type":"stack"' profile.spaa | \
  jq -s '[.[] | select(.exclusive) | {frame: .exclusive.frame, value: (.exclusive.weights[] | select(.metric=="period") | .value)}] | group_by(.frame) | map({frame: .[0].frame, total: (map(.value) | add)}) | sort_by(-.total) | .[0:20]'

Then look up the frame IDs to get function names:

# Get frame details for a specific ID
grep '"type":"frame"' profile.spaa | jq 'select(.id == 101)'

Analyzing Memory Profiles

SPAA also supports heap/allocation profilers. Memory events use different metrics:

# Find top allocation sites by bytes allocated
grep '"type":"stack"' profile.spaa | \
  jq -s '[.[] | select(.weights[] | .metric == "alloc_bytes")] | sort_by(-.weights[] | select(.metric=="alloc_bytes") | .value) | .[0:10]'

# Find potential memory leaks (high live_bytes)
grep '"type":"stack"' profile.spaa | \
  jq -s '[.[] | select(.weights[] | .metric == "live_bytes")] | sort_by(-.weights[] | select(.metric=="live_bytes") | .value) | .[0:10]'

Key memory metrics:

• alloc_bytes / alloc_count - Total allocations
• live_bytes / live_count - Currently unreleased memory (potential leaks)
• peak_bytes - High-water mark

Reconstructing Call Stacks

Stack records contain frame IDs. To see the actual function names:

# Extract a stack and resolve its frames
STACK_FRAMES=$(grep '"type":"stack"' profile.spaa | head -1 | jq -r '.frames | @csv')

# Build a frame lookup table, then query it
grep '"type":"frame"' profile.spaa | jq -s 'INDEX(.id)' > /tmp/frames.json
echo $STACK_FRAMES | tr ',' '\n' | while read fid; do
  jq --arg id "$fid" '.[$id] | "\(.func) (\(.srcline // "unknown"))"' /tmp/frames.json
done

Common Analysis Patterns

Filter by thread/process

grep '"type":"stack"' profile.spaa | jq 'select(.context.tid == 4511)'

Filter by event type

grep '"type":"stack"' profile.spaa | jq 'select(.context.event == "cycles")'

Find kernel vs userspace time

# Kernel stacks
grep '"type":"stack"' profile.spaa | jq 'select(.stack_type == "kernel")'

# Or check frame kinds
grep '"type":"frame"' profile.spaa | jq 'select(.kind == "kernel")' | head -20

Temporal analysis (if sample records exist)

# Check if raw samples are included
grep -c '"type":"sample"' profile.spaa

# Plot sample distribution over time
grep '"type":"sample"' profile.spaa | jq -s 'group_by(.timestamp | floor) | map({time: .[0].timestamp | floor, count: length})'

Tips for Performance Analysis

1. Start with the header - Understand the profiler, sampling mode, and time range
2. Check the primary metric - Use period for perf, samples for DTrace
3. Look at exclusive time first - This shows actual hotspots, not just callers
4. Cross-reference frame IDs - Build a lookup table for readable output
5. Filter by context - Narrow down by thread, CPU, or event type
6. For memory issues - Focus on live_bytes to find leaks, alloc_bytes for churn