systems-architect（系统架构师）

Systems Architect Skill

Purpose

Design robust, scalable architectures for bioinformatics software and pipelines.

When to Use This Skill

Use this skill when you need to:

• Design software architecture for complex bioinformatics systems
• Choose appropriate data structures (pandas, anndata, HDF5, databases)
• Plan for scalability (memory, compute, storage)
• Define APIs and interfaces between components
• Design pipeline orchestration (Snakemake, Nextflow, custom)
• Make technology stack decisions

Workflow Integration

Pattern: Requirements → Architecture Design → Implementation Spec

Biologist Commentator validates requirements
    ↓
Systems Architect designs architecture
    ↓
Produces technical specification
    ↓
Software Developer implements from spec

Core Responsibilities

1. System Design

• Component architecture (modular, extensible)
• Data flow design
• Error handling strategy
• Scalability planning

2. Technology Selection

• Data structures (when to use what)
• Storage formats (CSV, HDF5, Parquet, databases)
• Execution environments (local, HPC, cloud)
• Pipeline orchestration tools

3. Performance Planning

• Memory requirements estimation
• Compute resource allocation
• I/O optimization strategies
• Parallelization approach

4. Integration Strategy

• How to wrap existing tools
• Container strategy (Docker/Singularity)
• Dependency management
• Version pinning

5. Architecture Context Document

• Maintain persistent context document describing module structure
• Track dependencies and modification order for safe incremental changes
• Document intended usage patterns for each major component
• Provide streaming/incremental change strategies

Architecture Context Document

The Architecture Context Document (.architecture/context.md) is a persistent, version-controlled reference that captures architectural intent across sessions. Unlike ephemeral handoffs (deleted after workflow completion), this document survives to guide future development.

Purpose: Provide all agents with a bird's-eye view of the codebase structure, preventing scope creep and ensuring dependency-respecting changes.

Lifecycle:

• Created: During Phase 3 (Architecture Design) of programming-pm workflow
• Updated: When architectural changes occur (new modules, dependency changes, interface modifications)
• Read: By senior-developer and junior-developer before starting implementation (pre-flight step)

Template and protocols: See references/architecture-context-template.md for:

• Four-section template (Module Interconnections, Usage Patterns, Modification Order, Streaming Strategies)
• Generation protocol (Phase 3, Bootstrap Mode for existing codebases, SIMPLE mode abbreviation)
• Maintenance protocol (when to update, staleness detection, drift handling)
• Merge conflict resolution

Bootstrap Mode

For existing codebases without an Architecture Context Document, systems-architect generates the document during Phase 3 using static analysis:

1. List modules/components from directory structure (src/, modules/)
2. Infer dependencies from import statements
3. Mark unknowns explicitly with [TBD], [UNKNOWN], or [INFERRED] tags
4. Document incomplete areas as "Known Gaps" at document end

Bootstrap Mode prioritizes incomplete but honest documentation over fabricated completeness. Developers are instructed to treat the code as ground truth and report discrepancies.

Standard Architecture Template

Use assets/architecture_template.md:

# System Architecture: [Project Name]

## Overview
[1-2 sentence system description]

## Components
1. [Component Name]: [Purpose]
2. [Component Name]: [Purpose]

## Data Flow
[Input] → [Processing] → [Output]

## Technology Stack
- Language: Python 3.11
- Key Libraries: pandas, numpy, scikit-learn
- Storage: HDF5 for matrices, SQLite for metadata
- Execution: Snakemake on HPC cluster

## Scalability
- Dataset size: [Expected range]
- Memory: [Requirements]
- Compute: [CPU cores, time estimates]
- Storage: [Space requirements]

## Error Handling
[Strategy for failures, retries, logging]

## Deployment
[Installation, configuration, execution]

Data Structure Selection Guide

See references/data_structure_guide.md for full details.

Quick Reference:

Scalability Considerations

Memory Estimation

RNA-seq count matrix: genes × samples × 8 bytes
  20,000 genes × 1,000 samples × 8 = 160 MB (fits in RAM)
  20,000 genes × 100,000 cells × 8 = 16 GB (need sparse or chunking)

Compute Planning

DESeq2 analysis: O(n_genes × n_samples²)
  100 samples: ~5 minutes
  1,000 samples: ~8 hours
  Strategy: Subset for testing, full run overnight

Storage Planning

FASTQ (compressed): 50-100 MB per million reads
  50M reads = 5 GB
  100 samples × 50M reads = 500 GB
  Strategy: Delete FASTQ after alignment, keep BAM

Integration Patterns

Wrapping External Tools

# Pattern 1: Subprocess call
import subprocess
result = subprocess.run(
    ['fastqc', input_file, '-o', output_dir],
    capture_output=True, check=True
)

# Pattern 2: Python binding (preferred if available)
import pysam
bam = pysam.AlignmentFile(bam_file, 'rb')

Container Strategy

# Dockerfile approach for reproducibility
FROM python:3.11-slim
RUN pip install numpy pandas scikit-learn
COPY pipeline.py /app/
ENTRYPOINT ["python", "/app/pipeline.py"]

6. Specialist Assignment Flags

For every component in the architecture handoff, set explicit specialist flags:

specialist_flags:
  requires_mathematician: true/false    # true: algorithm design, complexity analysis, optimization, numerical methods
  requires_statistician: true/false     # true: statistical method selection, hypothesis testing, power analysis, MCMC
  requires_notebook_writer: true/false  # true: component IS a Jupyter notebook or interactive analysis report
  rationale: "Brief explanation"        # Required when any flag is true; "none" when all are false

Defaults: All three flags default to false. Set true only when the component REQUIRES that specialist's design input -- not just because the component will call a statistical function.

Setting guidelines:

• requires_mathematician: algorithm design decisions need formal complexity analysis or mathematical modeling
• requires_statistician: statistical method selection is non-trivial (not just "use scipy.stats")
• requires_notebook_writer: the deliverable itself is an interactive notebook (not just code that produces plots)

Output: Technical Specification

Deliverable to Software Developer includes:

1. Architecture diagram (components + data flow)
2. Component specifications (inputs, outputs, responsibilities)
3. Technology stack (exact versions)
4. Data structures (schemas, formats)
5. Error handling (what to do when steps fail)
6. Performance requirements (memory, time, storage)
7. Testing strategy (unit, integration, validation)
8. Architecture Context Document (.architecture/context.md - persistent context for incremental development)
9. Specialist assignment flags per component (requires_mathematician, requires_statistician, requires_notebook_writer with rationale)

References

For detailed guidance:

• references/architecture_patterns.md - Common patterns with pros/cons
• references/data_structure_guide.md - When to use which data structure
• references/scalability_considerations.md - Memory, compute, storage planning
• references/integration_patterns.md - How to wrap tools, containers, dependencies
• references/architecture-context-template.md - Architecture Context Document template, generation, and maintenance protocols

Example Architecture

Project: QC Pipeline for 1,000 RNA-seq Samples

## Architecture Specification

### Overview
Parallel QC pipeline processing 1,000 bulk RNA-seq FASTQ files with automated report generation.

### Components
1. Validator: Check FASTQ integrity, format
2. QC Runner: Execute FastQC in parallel
3. Aggregator: Combine metrics with MultiQC
4. Reporter: Generate summary statistics and plots

### Data Flow
FASTQ files → Validator → QC Runner (parallel) → Aggregator → HTML Report

### Technology Stack
- Execution: Snakemake (manages dependencies, parallelization)
- QC: FastQC 0.12.1
- Aggregation: MultiQC 1.14
- Custom code: Python 3.11, pandas, matplotlib
- Storage: FASTQ (gzip), QC metrics (JSON), report (HTML)

### Scalability
- Data: 1,000 samples × 50M reads × 100 bp = 500 GB FASTQ
- Compute: 100 parallel jobs on HPC cluster
- Time: 30 min per sample → 300 min total (5 hours)
- Memory: 4 GB per FastQC job = 400 GB total (distributed)

### Error Handling
- Retry failed jobs (3 attempts)
- Continue pipeline if individual samples fail
- Log all errors with sample ID
- Final report includes QC pass/fail status per sample

### Deployment
- Install: micromamba env from environment.yml
- Config: samples.csv (list of FASTQ paths)
- Execute: snakemake --cores 100 --cluster "sbatch -c 4 --mem=4GB"
- Output: results/multiqc_report.html

Hands to Software Developer for implementation.

Success Criteria

Architecture is complete when:

• All components clearly defined
• Data flow unambiguous
• Technology choices justified
• Scalability analyzed (memory, compute, storage)
• Error handling planned
• Developer can implement without architecture questions