skill-107（技能 107）

Skill 107: Distributed System Patterns & Design

Quality Grade: 94-95/100 Author: OpenClaw Assistant Last Updated: March 2026 Difficulty: Advanced (requires architectural thinking, trade-off analysis)

Overview

Distributed System Patterns are proven solutions to recurring problems in systems that span multiple machines, networks, and datacenters. As systems scale beyond single machines, coordination, fault tolerance, and consistency become non-negotiable.

This skill covers:

• Replication and consistency models
• Partitioning strategies and data distribution
• Consensus algorithms and leader election
• Failure recovery and resilience patterns
• Message passing and event ordering
• Coordination across services

Part 1: Replication Patterns

Master-Slave Replication

How it works:

• All writes go to master
• Master propagates to slaves asynchronously
• Reads can come from slaves (eventual consistency)

Trade-offs:

• ✓ Scalable reads
• ✗ Write bottleneck at master
• ✗ Stale reads from slaves
• ✗ Slave lag under high load

When to use: Read-heavy workloads, geographic distribution, backup resilience

Peer-to-Peer Replication

How it works:

• All nodes accept reads and writes
• Changes propagate peer-to-peer (gossip protocol)
• Eventual consistency with conflict resolution

Trade-offs:

• ✓ Scalable both reads and writes
• ✓ High availability (no single master)
• ✗ Conflict resolution complexity
• ✗ Higher network overhead

When to use: High availability needs, offline-first systems, global distribution

Chain Replication

How it works:

• Writes go to head, flow through chain to tail
• Tail is readable, provides strong consistency
• Head can be rebalanced independently

Trade-offs:

• ✓ Strong consistency
• ✓ Read tail scalability
• ✗ Slower writes (latency of chain length)
• ✗ Head failure needs rebalance

When to use: Consistent reads critical, moderate write frequency

Part 2: Partitioning Strategies

Range-Based Partitioning

Partition 0: UserIDs [0, 1000000)
Partition 1: UserIDs [1000000, 2000000)
Partition 2: UserIDs [2000000, ∞)

Pros: Simple, range queries efficient Cons: Uneven distribution (hotspots), rebalancing expensive

Hash-Based Partitioning

Partition = hash(key) % num_partitions

Pros: Even distribution, fast lookup Cons: Range queries require full scan, rebalancing complex

Consistent Hashing

Nodes arranged in ring, key maps to first node clockwise
Adding/removing node affects only adjacent partitions (~1/N data moves)

Pros: Minimal rebalancing, scalable additions Cons: Uneven distribution without virtual nodes, algorithm complexity

Part 3: Consensus & Coordination

Two-Phase Commit (2PC)

Flow:

1. Coordinator asks all participants: "Can you commit?"
2. Participants respond Yes/No (reserve resources)
3. If all Yes, coordinator tells all: "Commit"
4. If any No, coordinator tells all: "Abort"

Guarantees: Atomic across all participants Problems: Blocking, not partition-tolerant, slow

Use case: Database transactions across shards

Raft Consensus

Leader election + log replication:

1. Nodes elect a leader via voting
2. Leader accepts all writes
3. Leader replicates log entries to followers
4. Majority replication = safe to commit

Guarantees: Safety (never lose committed data), liveness (will elect leader) Performance: Lower throughput than 2PC, but more resilient

Use case: Distributed consensus (etcd, Consul), metadata stores

CRDT (Conflict-free Replicated Data Types)

Approach: Assign unique IDs, track causal history Guarantees: Automatic conflict resolution, commutative operations

Example: Vector clocks + last-write-wins for distributed counters

Use case: Collaborative editing, offline-first applications

Part 4: Failure Recovery

Idempotency

Make operations repeatable—if a request is retried, result is same:

def transfer_funds(from_id, to_id, amount, idempotency_key):
    # Check: did we already process this key?
    if idempotency_cache.get(idempotency_key):
        return idempotency_cache[idempotency_key]
    
    result = _do_transfer(from_id, to_id, amount)
    idempotency_cache[idempotency_key] = result
    return result

Key: Idempotency key must be client-chosen and immutable

Retries with Backoff

Attempt 1: immediate
Attempt 2: wait 1s
Attempt 3: wait 2s
Attempt 4: wait 4s
Attempt 5: wait 8s (give up if still failing)

Jitter: add random delay to avoid thundering herd
backoff_time = min(max_backoff, base * (2 ^ attempt)) + random(0, jitter)

Circuit Breaker

State: CLOSED (normal) → OPEN (failing) → HALF_OPEN (testing)

CLOSED → OPEN: When error rate > threshold for duration
OPEN → HALF_OPEN: After cooldown period
HALF_OPEN → CLOSED: If test request succeeds
HALF_OPEN → OPEN: If test request fails

Part 5: Message Passing & Ordering

FIFO Ordering

Messages between two nodes arrive in send order. Implementation: Sequence numbers, TCP guarantees

Causal Ordering

If event A causally precedes B, A's message arrives before B's. Implementation: Vector clocks or version vectors

Total Ordering

All nodes receive all messages in same order. Implementation: Consensus-based broadcast, sequencer node

Trade-offs: Ordering strength vs. latency cost

Conclusion

Distributed system patterns are essential vocabulary for building scalable, reliable systems. Understanding replication, partitioning, consensus, and failure recovery lets you design systems that survive failures, scale horizontally, and provide guarantees users can depend on.

Key Takeaway: Choose patterns based on your actual requirements (CAP theorem), not ideals. Consistency, availability, and partition tolerance—pick two.