14 KiB
Decomposition & Reconstruction: Advanced Methodology
Workflow
Copy this checklist for complex decomposition scenarios:
Advanced Decomposition Progress:
- [ ] Step 1: Apply hierarchical decomposition techniques
- [ ] Step 2: Build and analyze dependency graphs
- [ ] Step 3: Perform critical path analysis
- [ ] Step 4: Use advanced property measurement
- [ ] Step 5: Apply optimization algorithms
Step 1: Apply hierarchical decomposition techniques - Multi-level decomposition with consistent abstraction levels. See 1. Hierarchical Decomposition.
Step 2: Build and analyze dependency graphs - Visualize and analyze component relationships. See 2. Dependency Graph Analysis.
Step 3: Perform critical path analysis - Identify bottlenecks using PERT/CPM. See 3. Critical Path Analysis.
Step 4: Use advanced property measurement - Rigorous measurement and statistical analysis. See 4. Advanced Property Measurement.
Step 5: Apply optimization algorithms - Systematic reconstruction approaches. See 5. Optimization Algorithms.
1. Hierarchical Decomposition
Multi-Level Decomposition Strategy
Break into levels: L0 (System) → L1 (3-7 subsystems) → L2 (3-7 components each) → L3+ (only if needed). Stop when component is atomic or further breakdown doesn't help goal.
Abstraction consistency: All components at same level should be at same abstraction type (e.g., all architectural components, not mixing "API Service" with "user login function").
Template:
System → Subsystem A → Component A.1, A.2, A.3
→ Subsystem B → Component B.1, B.2
→ Subsystem C → Component C.1 (atomic)
Document WHY decomposed to this level and WHY stopped.
2. Dependency Graph Analysis
Building Dependency Graphs
Nodes: Components (from decomposition) Edges: Relationships (dependency, data flow, control flow, etc.) Direction: Arrow shows dependency direction (A → B means A depends on B)
Example:
Frontend → API Service → Database
↓
Cache
↓
Message Queue
Graph Properties
Strongly Connected Components (SCCs): Circular dependencies (A → B → C → A). Problematic for isolation. Use Tarjan's algorithm.
Topological Ordering: Linear order where edges point forward (only if acyclic). Reveals safe build/deploy order.
Critical Path: Longest weighted path, determines minimum completion time. Bottleneck for optimization.
Dependency Analysis
Forward: "If I change X, what breaks?" (BFS from X outgoing) Backward: "What must work for X to function?" (BFS from X incoming) Transitive Reduction: Remove redundant edges to simplify visualization.
3. Critical Path Analysis
PERT/CPM (Program Evaluation and Review Technique / Critical Path Method)
Use case: System with sequential stages, need to identify time bottlenecks
Inputs:
- Components with estimated duration
- Dependencies between components
Process:
Step 1: Build dependency graph with durations
A (3h) → B (5h) → D (2h)
A (3h) → C (4h) → D (2h)
Step 2: Calculate earliest start time (EST) for each component
EST(node) = max(EST(predecessor) + duration(predecessor)) for all predecessors
Example:
- EST(A) = 0
- EST(B) = EST(A) + duration(A) = 0 + 3 = 3h
- EST(C) = EST(A) + duration(A) = 0 + 3 = 3h
- EST(D) = max(EST(B) + duration(B), EST(C) + duration(C)) = max(3+5, 3+4) = 8h
Step 3: Calculate latest finish time (LFT) working backwards
LFT(node) = min(LFT(successor) - duration(node)) for all successors
Example (working backwards from D):
- LFT(D) = project deadline (say 10h)
- LFT(B) = LFT(D) - duration(B) = 10 - 5 = 5h
- LFT(C) = LFT(D) - duration(C) = 10 - 4 = 6h
- LFT(A) = min(LFT(B) - duration(A), LFT(C) - duration(A)) = min(5-3, 6-3) = 2h
Step 4: Calculate slack (float)
Slack(node) = LFT(node) - EST(node) - duration(node)
Example:
- Slack(A) = 2 - 0 - 3 = -1h (on critical path, negative slack means delay)
- Slack(B) = 5 - 3 - 5 = -3h (critical)
- Slack(C) = 6 - 3 - 4 = -1h (has some float)
- Slack(D) = 10 - 8 - 2 = 0 (critical)
Step 5: Identify critical path
Components with zero (or minimum) slack form the critical path.
Critical path: A → B → D (total 10h)
Optimization insight: Only optimizing B will reduce total time. Optimizing C (non-critical) won't help.
Handling Uncertainty (PERT Estimates)
When durations are uncertain, use three-point estimates:
- Optimistic (O): Best case
- Most Likely (M): Expected case
- Pessimistic (P): Worst case
Expected duration: E = (O + 4M + P) / 6
Standard deviation: σ = (P - O) / 6
Example:
- Component A: O=2h, M=3h, P=8h
- Expected: E = (2 + 4×3 + 8) / 6 = 3.67h
- Std dev: σ = (8 - 2) / 6 = 1h
Use expected durations for critical path analysis, report confidence intervals
4. Advanced Property Measurement
Quantitative vs Qualitative Properties
Quantitative (measurable):
- Latency (ms), throughput (req/s), cost ($/month), lines of code, error rate (%)
- Measurement: Use APM tools, profilers, logs, benchmarks
- Reporting: Mean, median, p95, p99, min, max, std dev
Qualitative (subjective):
- Code readability, maintainability, user experience, team morale
- Measurement: Use rating scales (1-10), comparative ranking, surveys
- Reporting: Mode, distribution, outliers
Statistical Rigor
For quantitative measurements:
1. Multiple samples: Don't rely on single measurement
- Run benchmark 10+ times, report distribution
- Example: Latency = 250ms ± 50ms (mean ± std dev, n=20)
2. Control for confounds: Isolate what you're measuring
- Example: Measure DB query time with same dataset, same load, same hardware
3. Statistical significance: Determine if difference is real or noise
- Use t-test or ANOVA to compare means
- Report p-value (p < 0.05 typically considered significant)
For qualitative measurements:
1. Multiple raters: Reduce individual bias
- Have 3+ people rate complexity independently, average scores
2. Calibration: Define rating scale clearly
- Example: Complexity 1="< 50 LOC, no dependencies", 10=">1000 LOC, 20+ dependencies"
3. Inter-rater reliability: Check if raters agree
- Calculate Cronbach's alpha or correlation coefficient
Performance Profiling Techniques
CPU Profiling:
- Identify which components consume most CPU time
- Tools: perf, gprof, Chrome DevTools, Xcode Instruments
Memory Profiling:
- Identify which components allocate most memory or leak
- Tools: valgrind, heaptrack, Chrome DevTools, Instruments
I/O Profiling:
- Identify which components perform most disk/network I/O
- Tools: iotop, iostat, Network tab in DevTools
Tracing:
- Track execution flow through distributed systems
- Tools: OpenTelemetry, Jaeger, Zipkin, AWS X-Ray
Result: Component-level resource consumption data for bottleneck analysis
5. Optimization Algorithms
Greedy Optimization
Approach: Optimize components in order of highest impact first
Algorithm:
- Measure impact of optimizing each component (reduction in latency, cost, etc.)
- Sort components by impact (descending)
- Optimize highest-impact component
- Re-measure, repeat until goal achieved or diminishing returns
Example (latency optimization):
- Components: A (100ms), B (500ms), C (50ms)
- Sort by impact: B (500ms), A (100ms), C (50ms)
- Optimize B first → Reduce to 200ms → Total latency improved by 300ms
- Re-measure, continue
Advantage: Fast, often gets 80% of benefit with 20% of effort Limitation: May miss global optimum (e.g., removing B entirely better than optimizing B)
Dynamic Programming Approach
Approach: Find optimal decomposition/reconstruction by exploring combinations
Use case: When multiple components interact, greedy may not find best solution
Example (budget allocation):
- Budget: $1000/month
- Components: A (improves UX, costs $400), B (reduces latency, costs $600), C (adds feature, costs $500)
- Constraint: Total cost ≤ $1000
- Goal: Maximize value
Algorithm:
- Enumerate all feasible combinations: {A}, {B}, {C}, {A+B}, {A+C}, {B+C}
- Calculate value and cost for each
- Select combination with max value under budget constraint
Result: Optimal combination (may not be greedy choice)
Constraint Satisfaction
Approach: Find reconstruction that satisfies all hard constraints
Use case: Multiple constraints (latency < 500ms AND cost < $500/month AND reliability > 99%)
Formulation:
- Variables: Component choices (use component A or B? Parallelize or serialize?)
- Domains: Possible values for each choice
- Constraints: Rules that must be satisfied
Algorithm: Backtracking search, constraint propagation Tools: CSP solvers (Z3, MiniZinc)
Sensitivity Analysis
Goal: Understand how sensitive reconstruction is to property estimates
Process:
- Build reconstruction based on measured/estimated properties
- Vary each property by ±X% (e.g., ±20%)
- Re-run reconstruction
- Identify which properties most affect outcome
Example:
- Baseline: Component A latency = 100ms → Optimize B
- Sensitivity: If A latency = 150ms → Optimize A instead
- Conclusion: Decision is sensitive to A's latency estimate, need better measurement
6. Advanced Reconstruction Patterns
Caching & Memoization
Pattern: Add caching layer for frequently accessed components
When: Component is slow, accessed repeatedly, output deterministic
Example: Database query repeated 1000x/sec → Add Redis cache → 95% cache hit rate → 20× latency reduction
Trade-offs: Memory cost, cache invalidation complexity, eventual consistency
Batch Processing
Pattern: Process items in batches instead of one-at-a-time
When: Per-item overhead is high, latency not critical
Example: Send 1000 individual emails (1s each, total 1000s) → Batch into groups of 100 → Send via batch API (10s per batch, total 100s)
Trade-offs: Increased latency for individual items, complexity in failure handling
Asynchronous Processing
Pattern: Decouple components using message queues
When: Component is slow but result not needed immediately
Example: User uploads video → Process synchronously (60s wait) → User unhappy Reconstruction: User uploads → Queue processing → User sees "processing" → Email when done
Trade-offs: Complexity (need queue infrastructure), eventual consistency, harder to debug
Load Balancing & Sharding
Pattern: Distribute load across multiple instances of a component
When: Component is bottleneck, can be parallelized, load is high
Example: Single DB handles 10K req/s, saturated → Shard by user ID → 10 DBs each handle 1K req/s
Trade-offs: Operational complexity, cross-shard queries expensive, rebalancing cost
Circuit Breaker
Pattern: Fail fast when dependent component is down
When: Component depends on unreliable external service
Example: API calls external service → Service is down → API waits 30s per request → API becomes slow Reconstruction: Add circuit breaker → Detect failures → Stop calling for 60s → Fail fast (< 1ms)
Trade-offs: Reduced functionality during outage, tuning thresholds (false positives vs negatives)
7. Failure Mode & Effects Analysis (FMEA)
FMEA Process
Goal: Identify weaknesses and single points of failure in decomposed system
Process:
Step 1: List all components
Step 2: For each component, identify failure modes
- How can this component fail? (crash, slow, wrong output, security breach)
Step 3: For each failure mode, assess:
- Severity (S): Impact if failure occurs (1-10, 10 = catastrophic)
- Occurrence (O): Likelihood of failure (1-10, 10 = very likely)
- Detection (D): Ability to detect before impact (1-10, 10 = undetectable)
Step 4: Calculate Risk Priority Number (RPN) RPN = S × O × D
Step 5: Prioritize failures by RPN, design mitigations
Example
| Component | Failure Mode | S | O | D | RPN | Mitigation |
|---|---|---|---|---|---|---|
| Database | Crashes | 9 | 2 | 1 | 18 | Add replica, automatic failover |
| Cache | Stale data | 5 | 6 | 8 | 240 | Reduce TTL, add invalidation |
| API | DDoS attack | 8 | 4 | 3 | 96 | Add rate limiting, WAF |
Highest RPN = 240 (Cache stale data) → Address this first
Mitigation Strategies
Redundancy: Multiple instances, failover Monitoring: Early detection, alerting Graceful degradation: Degrade functionality instead of total failure Rate limiting: Prevent overload Input validation: Prevent bad data cascading Circuit breakers: Fail fast when dependencies down
8. Case Study Approach
Comparative Analysis
Compare reconstruction alternatives in table format (Latency, Cost, Time, Risk, Maintainability). Make recommendation with rationale based on trade-offs.
Iterative Refinement
If initial decomposition doesn't reveal insights, refine: go deeper in critical areas, switch decomposition strategy, add missing relationships. Re-run analysis. Stop when further refinement doesn't change recommendations.
9. Tool-Assisted Decomposition
Static analysis: CLOC, SonarQube (dependency graphs, complexity metrics) Dynamic analysis: Flame graphs, perf, Chrome DevTools (CPU/memory/I/O), Jaeger/Zipkin (distributed tracing)
Workflow: Static analysis → Dynamic measurement → Manual validation → Combine quantitative + qualitative
Caution: Tools miss runtime dependencies, overestimate coupling, produce overwhelming detail. Use as guide, not truth.
10. Communication & Visualization
Diagrams: Hierarchy trees, dependency graphs (color-code critical path), property heatmaps, before/after comparisons
Stakeholder views:
- Executives: 1-page summary, key findings, business impact
- Engineers: Detailed breakdown, technical rationale, implementation
- Product/Business: UX impact, cost-benefit, timeline
Adapt depth to audience expertise.