Split View: 분산 시스템 완전 정복: CAP theorem부터 분산 ML 학습, Kafka까지

┌─────────────────────────────────────────────┐
│              CAP Triangle                   │
│                                             │
│           Consistency (C)                  │
│               △                             │
│              / \                            │
│             /   \                           │
│            /     \                          │
│           /  CA   \                         │
│          /  (RDBMS)\                        │
│         /───────────\                       │
│        /  CP  │  AP  \                      │
│       / (HBase│(Cassandra)                  │
│      ▽────────┼────────▽                    │
│  Availability  Partition Tolerance          │
│     (A)              (P)                    │
└─────────────────────────────────────────────┘

P → A or C  (during partition)
E → L or C  (during normal operation)

Cassandra: PA/EL  (availability + low latency first)
etcd/Raft: PC/EC  (consistency first)

Linearizability
  └── Strongest. All operations appear instantaneous
  └── Examples: etcd, Zookeeper

Sequential Consistency
  └── All nodes observe operations in the same order
  └── Examples: CPU memory model

Causal Consistency
  └── Causally related operations maintain order
  └── Examples: MongoDB causally consistent sessions

Eventual Consistency
  └── All nodes converge to the same value eventually
  └── Examples: Cassandra, DNS, S3

┌──────────┐  timeout   ┌──────────────┐  majority   ┌────────┐
│          │──────────▶│   Candidate  │────────────▶│        │
│ Follower │            └──────────────┘             │ Leader │
│          │◀─────────────────────────────────────── │        │
└──────────┘          receive heartbeat              └────────┘
     ▲                                                    │
     └────────────────── heartbeat ──────────────────────┘

# Distributed locking with etcd
import etcd3
import time

def distributed_lock(etcd_client, lock_name, ttl=10):
    """Distributed lock using Raft-backed etcd"""
    lease = etcd_client.lease(ttl)
    lock_key = f"/locks/{lock_name}"

    # atomic Compare-And-Swap
    success, _ = etcd_client.transaction(
        compare=[
            etcd3.transactions.create(lock_key, '==', 0)
        ],
        success=[
            etcd3.transactions.put(lock_key, 'locked', lease=lease)
        ],
        failure=[]
    )

    if success:
        print(f"Lock acquired: {lock_name}")
        return lease
    else:
        print(f"Lock already held: {lock_name}")
        return None

def release_lock(etcd_client, lease):
    """Release the distributed lock"""
    if lease:
        lease.revoke()
        print("Lock released")

# Usage
client = etcd3.client(host='localhost', port=2379)

lease = distributed_lock(client, 'my-resource')
if lease:
    try:
        # Perform work in critical section
        time.sleep(5)
    finally:
        release_lock(client, lease)

┌─────────────────────────────────────────────────────┐
│                  etcd Cluster (3 nodes)             │
│                                                     │
│  ┌──────────┐    ┌──────────┐    ┌──────────┐      │
│  │  Leader  │───▶│Follower 1│    │Follower 2│      │
│  │  node-1  │───▶│  node-2  │    │  node-3  │      │
│  └──────────┘    └──────────┘    └──────────┘      │
│       ▲               │               │             │
│       └───────────────┴───────────────┘             │
│                  Raft Consensus                     │
└─────────────────────────────────────────────────────┘

Phase 1 (Prepare):
  Coordinator ──▶ Participant A: "Are you ready?"
  Coordinator ──▶ Participant B: "Are you ready?"
  Participant A ──▶ Coordinator: "YES"
  Participant B ──▶ Coordinator: "YES"

Phase 2 (Commit):
  Coordinator ──▶ Participant A: "Commit!"
  Coordinator ──▶ Participant B: "Commit!"

Order Saga:
  1. Create Order      → compensate: Cancel Order
  2. Deduct Inventory  → compensate: Restore Inventory
  3. Process Payment   → compensate: Refund Payment
  4. Start Shipping    → compensate: Cancel Shipment

from typing import Callable, List
from dataclasses import dataclass
import logging

logger = logging.getLogger(__name__)

@dataclass
class SagaStep:
    name: str
    action: Callable
    compensating_action: Callable

class SagaOrchestrator:
    """Saga Orchestrator pattern implementation"""

    def __init__(self, steps: List[SagaStep]):
        self.steps = steps
        self.completed_steps = []

    def execute(self, context: dict) -> bool:
        for step in self.steps:
            try:
                logger.info(f"Executing step: {step.name}")
                step.action(context)
                self.completed_steps.append(step)
                logger.info(f"Step completed: {step.name}")
            except Exception as e:
                logger.error(f"Step failed: {step.name}, error: {e}")
                self._compensate(context)
                return False
        return True

    def _compensate(self, context: dict):
        """Execute compensating transactions in reverse order"""
        logger.info("Starting compensation...")
        for step in reversed(self.completed_steps):
            try:
                logger.info(f"Compensating: {step.name}")
                step.compensating_action(context)
            except Exception as e:
                logger.error(f"Compensation failed: {step.name}, error: {e}")
                raise RuntimeError(f"Saga compensation failed at {step.name}")

# Define the Order Saga
def create_order(ctx):
    ctx['order_id'] = 'ORD-001'
    print(f"Order created: {ctx['order_id']}")

def cancel_order(ctx):
    print(f"Order cancelled: {ctx['order_id']}")

def deduct_inventory(ctx):
    ctx['inventory_reserved'] = True
    print("Inventory deducted")

def restore_inventory(ctx):
    print("Inventory restored")

def process_payment(ctx):
    raise Exception("Payment gateway timeout")

def refund_payment(ctx):
    print("Payment refunded")

order_saga = SagaOrchestrator([
    SagaStep("create_order", create_order, cancel_order),
    SagaStep("deduct_inventory", deduct_inventory, restore_inventory),
    SagaStep("process_payment", process_payment, refund_payment),
])

context = {"user_id": "user-123", "amount": 50000}
result = order_saga.execute(context)
print(f"Saga result: {'success' if result else 'failed with compensation'}")

CQRS Architecture:

  Command Side (Writes)        Query Side (Reads)
  ─────────────────            ──────────────────
  POST /orders          ──▶    GET /orders/123
  PUT  /inventory       ──▶    GET /inventory/stats
       │                              ▲
       ▼                              │
  Event Store ──── Projection ────────┘
  (append-only)    (updates read models)

┌─────────────────────────────────────────────────────────────┐
│                     Kafka Cluster                           │
│                                                             │
│  Topic: orders (4 partitions, replication factor: 3)        │
│                                                             │
│  Broker 1      Broker 2      Broker 3                       │
│  ┌──────┐      ┌──────┐      ┌──────┐                       │
│  │ P0 L │      │ P0 F │      │ P0 F │  L=Leader, F=Follower │
│  │ P1 F │      │ P1 L │      │ P1 F │                       │
│  │ P2 F │      │ P2 F │      │ P2 L │                       │
│  │ P3 L │      │ P3 F │      │ P3 F │                       │
│  └──────┘      └──────┘      └──────┘                       │
│                                                             │
│  Consumer Group A          Consumer Group B                 │
│  Consumer 1: P0, P1        Consumer X: P0, P1, P2, P3      │
│  Consumer 2: P2, P3                                         │
└─────────────────────────────────────────────────────────────┘

from confluent_kafka import Producer, Consumer, KafkaError
import json
import time

# Producer configuration
producer_config = {
    'bootstrap.servers': 'localhost:9092',
    'acks': 'all',           # Wait for all ISR replicas to acknowledge
    'retries': 3,
    'enable.idempotence': True,   # Exactly-once semantics
    'compression.type': 'snappy',
}

def delivery_report(err, msg):
    if err is not None:
        print(f'Message delivery failed: {err}')
    else:
        print(f'Message delivered to {msg.topic()} [{msg.partition()}] @ {msg.offset()}')

producer = Producer(producer_config)

# Publish order events
for i in range(10):
    order_event = {
        'order_id': f'ORD-{i:04d}',
        'user_id': f'user-{i % 5}',
        'amount': 10000 * (i + 1),
        'timestamp': time.time()
    }
    producer.produce(
        topic='orders',
        key=order_event['user_id'],   # Same user → same partition
        value=json.dumps(order_event),
        callback=delivery_report
    )
    producer.poll(0)

producer.flush()
print("All messages sent")

# Consumer configuration
consumer_config = {
    'bootstrap.servers': 'localhost:9092',
    'group.id': 'order-processor-group',
    'auto.offset.reset': 'earliest',
    'enable.auto.commit': False,   # Manual commit for at-least-once
}

consumer = Consumer(consumer_config)
consumer.subscribe(['orders'])

try:
    while True:
        msg = consumer.poll(timeout=1.0)
        if msg is None:
            continue
        if msg.error():
            if msg.error().code() == KafkaError._PARTITION_EOF:
                continue
            else:
                raise KafkaError(msg.error())

        event = json.loads(msg.value().decode('utf-8'))
        print(f"Processing order: {event['order_id']}, amount: {event['amount']}")

        # Commit offset after successful processing
        consumer.commit(asynchronous=False)

except KeyboardInterrupt:
    pass
finally:
    consumer.close()

# Kafka Topic configuration (Strimzi Operator)
apiVersion: kafka.strimzi.io/v1beta2
kind: KafkaTopic
metadata:
  name: orders
  labels:
    strimzi.io/cluster: my-cluster
spec:
  partitions: 12 # Max parallelism = partition count
  replicas: 3 # Replication factor
  config:
    retention.ms: '604800000' # Retain messages for 7 days
    min.insync.replicas: '2' # Require 2 ISR replicas for ack
    compression.type: snappy
    cleanup.policy: delete

import hashlib
from sortedcontainers import SortedDict

class ConsistentHashRing:
    """Consistent hashing ring implementation"""

    def __init__(self, virtual_nodes=150):
        self.virtual_nodes = virtual_nodes
        self.ring = SortedDict()
        self.nodes = set()

    def _hash(self, key: str) -> int:
        return int(hashlib.md5(key.encode()).hexdigest(), 16)

    def add_node(self, node: str):
        """Add a node to the ring with virtual nodes"""
        self.nodes.add(node)
        for i in range(self.virtual_nodes):
            virtual_key = f"{node}#{i}"
            hash_val = self._hash(virtual_key)
            self.ring[hash_val] = node
        print(f"Added node: {node} with {self.virtual_nodes} virtual nodes")

    def remove_node(self, node: str):
        """Remove a node from the ring"""
        self.nodes.discard(node)
        keys_to_remove = [k for k, v in self.ring.items() if v == node]
        for k in keys_to_remove:
            del self.ring[k]
        print(f"Removed node: {node}")

    def get_node(self, key: str) -> str:
        """Return the node responsible for a key"""
        if not self.ring:
            raise Exception("Ring is empty")

        hash_val = self._hash(key)
        idx = self.ring.bisect_left(hash_val)
        if idx == len(self.ring):
            idx = 0  # Wrap around the ring

        return self.ring.values()[idx]

    def get_nodes_for_replication(self, key: str, n: int) -> list:
        """Return n distinct nodes for replication"""
        if len(self.nodes) < n:
            raise Exception(f"Not enough nodes: {len(self.nodes)} < {n}")

        hash_val = self._hash(key)
        idx = self.ring.bisect_left(hash_val)

        result_nodes = []
        seen_nodes = set()
        total = len(self.ring)

        for i in range(total):
            curr_idx = (idx + i) % total
            node = self.ring.values()[curr_idx]
            if node not in seen_nodes:
                seen_nodes.add(node)
                result_nodes.append(node)
                if len(result_nodes) == n:
                    break

        return result_nodes

# Example usage
ring = ConsistentHashRing(virtual_nodes=150)
for i in range(5):
    ring.add_node(f"cassandra-node-{i}")

for key in ["user:1001", "order:5555", "product:abc", "session:xyz"]:
    node = ring.get_node(key)
    replicas = ring.get_nodes_for_replication(key, 3)
    print(f"Key '{key}' -> Primary: {node}, Replicas: {replicas}")

┌──────────────────────────────────────────────────────────────────┐
│                   Cassandra Ring (6 nodes)                       │
│                                                                  │
│                        Node 1                                    │
│                       /      \                                   │
│                   Node 6      Node 2                             │
│                     |            |                               │
│                   Node 5      Node 3                             │
│                       \      /                                   │
│                        Node 4                                    │
│                                                                  │
│  Write Path:                                                     │
│  Client → Coordinator → Hash partition key → Responsible nodes   │
│                       → CommitLog (WAL) → Memtable → SSTable    │
│                                                                  │
│  Read Path:                                                      │
│  Client → Coordinator → Responsible nodes → Row Cache / Bloom    │
│                       → Merge SSTables → Return latest data      │
└──────────────────────────────────────────────────────────────────┘

Parameter Server approach:
  ───────────────────────
  Worker 1 ─┐
  Worker 2 ─┤── PS (Parameter Server) ──▶ Aggregate + broadcast gradients
  Worker 3 ─┘

  Problem: PS is a bottleneck. Bandwidth requirement is O(n).

Ring-AllReduce approach:
  ────────────────────────
  Worker 1 ──▶ Worker 2 ──▶ Worker 3
     ▲                          │
     └──────────────────────────┘

  Each worker communicates only with neighbors.
  Total data transferred: 2 * (n-1)/n * data_size (independent of n)

import os
import torch
import torch.nn as nn
import torch.optim as optim
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data import DataLoader, DistributedSampler

def setup(rank, world_size):
    """Initialize distributed training"""
    os.environ['MASTER_ADDR'] = 'localhost'
    os.environ['MASTER_PORT'] = '12355'

    # NCCL backend: high-speed GPU communication (NVLink, InfiniBand)
    dist.init_process_group(
        backend='nccl',
        rank=rank,
        world_size=world_size
    )
    torch.cuda.set_device(rank)

def cleanup():
    dist.destroy_process_group()

class SimpleModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.layers = nn.Sequential(
            nn.Linear(1024, 2048),
            nn.ReLU(),
            nn.Linear(2048, 1024),
            nn.ReLU(),
            nn.Linear(1024, 10)
        )

    def forward(self, x):
        return self.layers(x)

def train(rank, world_size, epochs=5):
    setup(rank, world_size)

    # Place model on GPU, wrap with DDP
    model = SimpleModel().to(rank)
    ddp_model = DDP(model, device_ids=[rank])

    # DistributedSampler ensures each process gets a different data shard
    dataset = torch.utils.data.TensorDataset(
        torch.randn(1000, 1024),
        torch.randint(0, 10, (1000,))
    )
    sampler = DistributedSampler(dataset, num_replicas=world_size, rank=rank)
    dataloader = DataLoader(dataset, batch_size=32, sampler=sampler)

    optimizer = optim.Adam(ddp_model.parameters(), lr=1e-3)
    criterion = nn.CrossEntropyLoss()

    for epoch in range(epochs):
        sampler.set_epoch(epoch)  # Reseed shuffle each epoch
        total_loss = 0.0

        for batch_x, batch_y in dataloader:
            batch_x = batch_x.to(rank)
            batch_y = batch_y.to(rank)

            optimizer.zero_grad()
            output = ddp_model(batch_x)
            loss = criterion(output, batch_y)
            loss.backward()  # Gradient AllReduce happens automatically
            optimizer.step()
            total_loss += loss.item()

        if rank == 0:  # Only rank 0 logs
            print(f"Epoch {epoch}: Loss = {total_loss/len(dataloader):.4f}")

    cleanup()

# Launch: torchrun --nproc_per_node=4 train.py
if __name__ == '__main__':
    world_size = torch.cuda.device_count()
    torch.multiprocessing.spawn(
        train,
        args=(world_size,),
        nprocs=world_size,
        join=True
    )

NCCL (NVIDIA Collective Communications Library):

  Single node (NVLink):
  GPU0 ←──── NVLink ────→ GPU1
    │                       │
   600 GB/s bidirectional bandwidth

  Multi-node (InfiniBand + RDMA):
  Node 1 ←── IB (200 Gb/s) ──→ Node 2
    │                              │
  RDMA: direct GPU-to-GPU without CPU involvement

Backend selection:
  - nccl:  GPU distributed training (recommended)
  - gloo:  CPU training or debugging
  - mpi:   HPC environments

Request flow (traced by Jaeger/Zipkin):

  Browser → API GW → Order Svc → Inventory Svc → DB
    │           │          │            │          │
    │  TraceID: abc123     │            │          │
    │  SpanID: 001    SpanID: 002  SpanID: 003     │
    │                      │            │          │
    └──────────────────────┴────────────┴──────────┘
             Total latency: 120ms
             Order Svc: 80ms (bottleneck!)

OpenTelemetry layers:
  - API: write instrumentation code
  - SDK: process and export telemetry data
  - Exporter: send to Jaeger / Zipkin / OTLP endpoint

┌─────────────────────────────────────────────────────┐
│                   ELK Stack                         │
│                                                     │
│  App Logs ──▶ Filebeat ──▶ Logstash ──▶ Elasticsearch│
│  K8s Logs ──▶ Fluentd ──/             ──▶ Kibana    │
│                                         (dashboards)│
│                                                     │
│  Modern alternative:                               │
│  App Logs ──▶ Promtail ──▶ Loki ──▶ Grafana         │
└─────────────────────────────────────────────────────┘

# Prometheus scrape configuration
scrape_configs:
  - job_name: 'kafka'
    static_configs:
      - targets: ['kafka-broker-1:9090', 'kafka-broker-2:9090']
    metrics_path: '/metrics'
    scrape_interval: 15s

  - job_name: 'pytorch-training'
    static_configs:
      - targets: ['trainer-0:8080', 'trainer-1:8080']