multi-agent-performance-profiling（多 Agent 性能分析）

Multi-Agent Performance Profiling

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

Overview

Prescriptive workflow for spawning parallel profiling agents to comprehensively identify performance bottlenecks across multiple system layers. Successfully discovered that QuestDB ingests at 1.1M rows/sec (11x faster than target), proving database was NOT the bottleneck - CloudFront download was 90% of pipeline time.

When to Use This Skill

Use this skill when:

• Performance below SLO (e.g., 47K vs 100K rows/sec target)
• Multi-stage pipeline optimization (download → extract → parse → ingest)
• Database performance investigation
• Bottleneck identification in complex workflows
• Pre-optimization analysis (before making changes)

Key outcomes:

• Identify true bottleneck (vs assumed bottleneck)
• Quantify each stage's contribution to total time
• Prioritize optimizations by impact (P0/P1/P2)
• Avoid premature optimization of non-bottlenecks

Core Methodology

1. Multi-Layer Profiling Model (5-Agent Pattern)

Agent 1: Profiling (Instrumentation)

• Empirical timing of each pipeline stage
• Phase-boundary instrumentation with time.perf_counter()
• Memory profiling (peak usage, allocations)
• Bottleneck identification (% of total time)

Agent 2: Database Configuration Analysis

• Server settings review (WAL, heap, commit intervals)
• Production vs development config comparison
• Expected impact quantification (<5%, 10%, 50%)

Agent 3: Client Library Analysis

• API usage patterns (dataframe vs row-by-row)
• Buffer size tuning opportunities
• Auto-flush behavior analysis

Agent 4: Batch Size Analysis

• Current batch size validation
• Optimal batch range determination
• Memory overhead vs throughput tradeoff

Agent 5: Integration & Synthesis

• Consensus-building across agents
• Prioritization (P0/P1/P2) with impact quantification
• Implementation roadmap creation

2. Agent Orchestration Pattern

Parallel Execution (all 5 agents run simultaneously):

Agent 1 (Profiling)          → [PARALLEL]
Agent 2 (DB Config)          → [PARALLEL]
Agent 3 (Client Library)     → [PARALLEL]
Agent 4 (Batch Size)         → [PARALLEL]
Agent 5 (Integration)        → [PARALLEL - reads tmp/ outputs from others]

Key Principle: No dependencies between investigation agents (1-4). Integration agent synthesizes findings.

Dynamic Todo Management:

• Start with investigation plan (5 agents)
• Spawn agents in parallel using single message with multiple Task tool calls
• Update todos as each agent completes
• Integration agent waits for all findings before synthesizing

3. Profiling Script Structure

Each agent produces:

1. Investigation Script (e.g., profile_pipeline.py)

- time.perf_counter() instrumentation at phase boundaries - Memory profiling with tracemalloc - Structured output (phase, duration, % of total)

1. Report (markdown with findings, recommendations, impact quantification)
2. Evidence (benchmark results, config dumps, API traces)

Example Profiling Code:

import time

# Profile multi-stage pipeline
def profile_pipeline():
    results = {}

    # Phase 1: Download
    start = time.perf_counter()
    data = download_from_cdn(url)
    results["download"] = time.perf_counter() - start

    # Phase 2: Extract
    start = time.perf_counter()
    csv_data = extract_zip(data)
    results["extract"] = time.perf_counter() - start

    # Phase 3: Parse
    start = time.perf_counter()
    df = parse_csv(csv_data)
    results["parse"] = time.perf_counter() - start

    # Phase 4: Ingest
    start = time.perf_counter()
    ingest_to_db(df)
    results["ingest"] = time.perf_counter() - start

    # Analysis
    total = sum(results.values())
    for phase, duration in results.items():
        pct = (duration / total) * 100
        print(f"{phase}: {duration:.3f}s ({pct:.1f}%)")

    return results

4. Impact Quantification Framework

Priority Levels:

• P0 (Critical): >5x improvement, addresses primary bottleneck
• P1 (High): 2-5x improvement, secondary optimizations
• P2 (Medium): 1.2-2x improvement, quick wins
• P3 (Low): <1.2x improvement, minor tuning

Impact Reporting Format:

### Recommendation: [Optimization Name] (P0/P1/P2) - [IMPACT LEVEL]

**Impact**: 🔴/🟠/🟡 **Nx improvement**
**Effort**: High/Medium/Low (N days)
**Expected Improvement**: CurrentK → TargetK rows/sec

**Rationale**:

- [Why this matters]
- [Supporting evidence from profiling]
- [Comparison to alternatives]

**Implementation**:
[Code snippet or architecture description]

5. Consensus-Building Pattern

Integration Agent Responsibilities:

1. Read all investigation reports (Agents 1-4)
2. Identify consensus recommendations (all agents agree)
3. Flag contradictions (agents disagree)
4. Synthesize master integration report
5. Create implementation roadmap (P0 → P1 → P2)

Consensus Criteria:

• ≥3/4 agents recommend same optimization → Consensus
• 2/4 agents recommend, 2/4 neutral → Investigate further
• Agents contradict (one says "optimize X", another says "X is not bottleneck") → Run tie-breaker experiment

Workflow: Step-by-Step

Step 1: Define Performance Problem

Input: Performance metric below SLO Output: Problem statement with baseline metrics

Example Problem Statement:

Performance Issue: BTCUSDT 1m ingestion at 47K rows/sec
Target SLO: >100K rows/sec
Gap: 53% below target
Pipeline: CloudFront download → ZIP extract → CSV parse → QuestDB ILP ingest

Step 2: Create Investigation Plan

Directory Structure:

tmp/perf-optimization/
  profiling/              # Agent 1
    profile_pipeline.py
    PROFILING_REPORT.md
  questdb-config/         # Agent 2
    CONFIG_ANALYSIS.md
  python-client/          # Agent 3
    CLIENT_ANALYSIS.md
  batch-size/             # Agent 4
    BATCH_ANALYSIS.md
  MASTER_INTEGRATION_REPORT.md  # Agent 5

Agent Assignment:

• Agent 1: Empirical profiling (instrumentation)
• Agent 2: Database configuration analysis
• Agent 3: Client library usage analysis
• Agent 4: Batch size optimization analysis
• Agent 5: Synthesis and integration

Step 3: Spawn Agents in Parallel

IMPORTANT: Use single message with multiple Task tool calls for true parallelism

Example:

I'm going to spawn 5 parallel investigation agents:

[Uses Task tool 5 times in a single message]
- Agent 1: Profiling
- Agent 2: QuestDB Config
- Agent 3: Python Client
- Agent 4: Batch Size
- Agent 5: Integration (depends on others completing)

Execution:

# All agents run simultaneously (user observes 5 parallel tool calls)
# Each agent writes to its own tmp/ subdirectory
# Integration agent polls for completed reports

Step 4: Wait for All Agents to Complete

Progress Tracking:

• Update todo list as each agent completes
• Integration agent polls tmp/ directory for report files
• Once 4/4 investigation reports exist → Integration agent synthesizes

Completion Criteria:

• All 4 investigation reports written
• Integration report synthesizes findings
• Master recommendations list created

Step 5: Review Master Integration Report

Report Structure:

# Master Performance Optimization Integration Report

## Executive Summary

- Critical discovery (what is/isn't the bottleneck)
- Key findings from each agent (1-sentence summary)

## Top 3 Recommendations (Consensus)

1. [P0 Optimization] - HIGHEST IMPACT
2. [P1 Optimization] - HIGH IMPACT
3. [P2 Optimization] - QUICK WIN

## Agent Investigation Summary

### Agent 1: Profiling

### Agent 2: Database Config

### Agent 3: Client Library

### Agent 4: Batch Size

## Implementation Roadmap

### Phase 1: P0 Optimizations (Week 1)

### Phase 2: P1 Optimizations (Week 2)

### Phase 3: P2 Quick Wins (As time permits)

Step 6: Implement Optimizations (P0 First)

For each recommendation:

1. Implement highest-priority optimization (P0)
2. Re-run profiling script
3. Verify expected improvement achieved
4. Update report with actual results
5. Move to next priority (P1, P2, P3)

Example Implementation:

# Before optimization
uv run python tmp/perf-optimization/profiling/profile_pipeline.py
# Output: 47K rows/sec, download=857ms (90%)

# Implement P0 recommendation (concurrent downloads)
# [Make code changes]

# After optimization
uv run python tmp/perf-optimization/profiling/profile_pipeline.py
# Output: 450K rows/sec, download=90ms per symbol * 10 concurrent (90%)

Real-World Example: QuestDB Refactor Performance Investigation

Context: Pipeline achieving 47K rows/sec, target 100K rows/sec (53% below SLO)

Assumptions Before Investigation:

• QuestDB ILP ingestion is the bottleneck (4% of time)
• Need to tune database configuration
• Need to optimize Sender API usage

Findings After 5-Agent Investigation:

1. Profiling Agent: CloudFront download is 90% of time (857ms), ILP ingest only 4% (40ms)
2. QuestDB Config Agent: Database already optimal, tuning provides <5% improvement
3. Python Client Agent: Sender API already optimal (using dataframe() bulk ingestion)
4. Batch Size Agent: 44K batch size is within optimal range
5. Integration Agent: Consensus recommendation - optimize download, NOT database

Top 3 Recommendations:

1. 🔴 P0: Concurrent multi-symbol downloads (10-20x improvement)
2. 🟠 P1: Multi-month pipeline parallelism (2x improvement)
3. 🟡 P2: Streaming ZIP extraction (1.3x improvement)

Impact: Discovered database ingests at 1.1M rows/sec (11x faster than target) - proving database was never the bottleneck

Outcome: Avoided wasting 2-3 weeks optimizing database when download was the real bottleneck

Common Pitfalls

1. Profiling Only One Layer

❌ Bad: Profile database only, assume it's the bottleneck ✅ Good: Profile entire pipeline (download → extract → parse → ingest)

2. Serial Agent Execution

❌ Bad: Run Agent 1, wait, then run Agent 2, wait, etc. ✅ Good: Spawn all 5 agents in parallel using single message with multiple Task calls

3. Optimizing Without Profiling

❌ Bad: "Let's optimize the database config first" (assumption-driven) ✅ Good: Profile first, discover database is only 4% of time, optimize download instead

4. Ignoring Low-Hanging Fruit

❌ Bad: Only implement P0 (highest impact, highest effort) ✅ Good: Implement P2 quick wins (1.3x for 4-8 hours effort) while planning P0

5. Not Re-Profiling After Changes

❌ Bad: Implement optimization, assume it worked ✅ Good: Re-run profiling script, verify expected improvement achieved

Resources

scripts/

Not applicable - profiling scripts are project-specific (stored in tmp/perf-optimization/)

references/

• profiling_template.py - Template for phase-boundary instrumentation
• integration_report_template.md - Template for master integration report
• impact_quantification_guide.md - How to assess P0/P1/P2 priorities

assets/

Not applicable - profiling artifacts are project-specific

Troubleshooting

Post-Execution Reflection

After this skill completes, reflect before closing the task:

1. Locate yourself. — Find this SKILL.md's canonical path before editing.
2. What failed? — Fix the instruction that caused it.
3. What worked better than expected? — Promote to recommended practice.
4. What drifted? — Fix any script, reference, or dependency that no longer matches reality.
5. Log it. — Evolution-log entry with trigger, fix, and evidence.

Do NOT defer. The next invocation inherits whatever you leave behind.