Split View: 시스템 디자인 면접 완전 가이드 2025: 대규모 시스템 설계, 확장성, 가용성 패턴

Interviewer: "Design a URL shortening service."

Good questions:
Q: "What's the expected DAU?"
A: "100 million."

Q: "Which is more important, URL shortening or redirection?"
A: "Redirection is more important. Reads are much more frequent."

Q: "Is there a length limit for shortened URLs?"
A: "As short as possible."

Q: "Should we support custom URLs?"
A: "Yes, optionally."

+-----------+     +---------------+     +---------------+
|  Client   |---->| Load Balancer |---->|  Web Server   |
+-----------+     +---------------+     +------+--------+
                                               |
                    +--------------------------+
                    v                          v
             +-----------+            +---------------+
             |   Cache   |            |   Database    |
             +-----------+            +---------------+

Before:                 After:
+---------------+      +---------------+
|  4 CPU        |      |  32 CPU       |
|  16GB RAM     | ---> |  256GB RAM    |
|  500GB SSD    |      |  4TB SSD      |
+---------------+      +---------------+

                    +---------------+
              +---->|  Server 1     |
              |     +---------------+
+-----------+ |     +---------------+
|   Load    |-+---->|  Server 2     |
| Balancer  | |     +---------------+
+-----------+ |     +---------------+
              +---->|  Server 3     |
              |     +---------------+
              |     +---------------+
              +---->|  Server N     |
                    +---------------+

Bad (Stateful):
+-----------+  session  +-----------+
|  Client   |---------->| Server A  |  (Session lost if Server A dies)
+-----------+           +-----------+

Good (Stateless):
+-----------+     +-----------+     +-----------+
|  Client   |---->|   Any     |---->|   Redis   |
+-----------+     |  Server   |     | (Session) |
                  +-----------+     +-----------+

# AWS Auto Scaling configuration example
AutoScalingGroup:
  MinSize: 2
  MaxSize: 20
  DesiredCapacity: 4
  TargetTrackingScaling:
    TargetValue: 70.0          # Maintain 70% CPU
    PredefinedMetricType: ASGAverageCPUUtilization

  # Schedule-based scaling (predictable traffic)
  ScheduledActions:
    - Schedule: "0 8 * * MON-FRI"   # Weekdays 8 AM
      MinSize: 6
    - Schedule: "0 22 * * *"        # Daily 10 PM
      MinSize: 2

1. Round Robin (Sequential Distribution)
   Request 1 -> Server A
   Request 2 -> Server B
   Request 3 -> Server C
   Request 4 -> Server A  (back to start)

2. Weighted Round Robin
   Server A (weight=5): handles 5 requests
   Server B (weight=3): handles 3 requests
   Server C (weight=2): handles 2 requests

3. Least Connections
   Server A: 10 connections  <-- new request goes here
   Server B: 25 connections
   Server C: 18 connections

4. IP Hash
   hash(client_ip) % server_count = target_server
   -> Same client always goes to same server

L4 (Transport Layer):
+----------+    TCP/UDP     +-----------+
| Client   |--------------->| L4 LB     |-- routes by IP:Port
+----------+                +-----------+
  Fast, simple, no SSL offload

L7 (Application Layer):
+----------+    HTTP/HTTPS  +-----------+
| Client   |--------------->| L7 LB     |-- routes by URL, Header, Cookie
+----------+                +-----------+
  Content-based routing, SSL termination, more flexible

Regular Hashing:
hash(key) % 3 = server number
-> Adding a server: hash(key) % 4 = most keys move!

Consistent Hashing (Hash Ring):

        Server A
        /      \
       /   0    \
      /          \
  Server D     Server B
      \          /
       \   key  /
        \  |  /
        Server C

-> Adding a server: only adjacent keys move (~1/N of total)

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

        for node in nodes:
            for i in range(virtual_nodes):
                key = self._hash(f"node-vn-i")
                self.ring[key] = node

    def get_node(self, data_key):
        if not self.ring:
            return None
        hash_val = self._hash(data_key)
        # Find nearest node clockwise on the hash ring
        idx = self.ring.bisect_right(hash_val)
        if idx == len(self.ring):
            idx = 0
        return self.ring.peekitem(idx)[1]

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

1. Cache-Aside (Lazy Loading):
   +--------+  1.read   +---------+
   |  App   |---------->|  Cache  |
   +---+----+  miss     +---------+
       | 2.read DB         ^
       v             3.store in cache
   +--------+
   |   DB   |
   +--------+

2. Write-Through:
   +--------+  1.write  +---------+  2.sync write  +--------+
   |  App   |---------->|  Cache  |--------------->|   DB   |
   +--------+           +---------+                +--------+
   (Always consistent, higher write latency)

3. Write-Back (Write-Behind):
   +--------+  1.write  +---------+  2.async batch write  +--------+
   |  App   |---------->|  Cache  |======================>|   DB   |
   +--------+           +---------+                       +--------+
   (Great write performance, data loss risk)

4. Write-Around:
   +--------+  1.write  +--------+
   |  App   |---------->|   DB   |
   +---+----+           +--------+
       | 2.load to cache on read
       v
   +---------+
   |  Cache  |
   +---------+

# Redis - supports various data structures
import redis

r = redis.Redis(host='localhost', port=6379)

# String
r.set('user:1001:name', 'Alice', ex=3600)  # 1-hour TTL

# Hash
r.hset('user:1001', mapping={
    'name': 'Alice',
    'email': 'alice@example.com',
    'login_count': 42
})

# Sorted Set (leaderboard)
r.zadd('leaderboard', {'player_a': 1500, 'player_b': 1800})
top_10 = r.zrevrange('leaderboard', 0, 9, withscores=True)

# Pub/Sub
r.publish('notifications', 'New message!')

Without CDN:
User(Seoul) ---- 5000km ----> Origin Server(US)
RTT: ~200ms

With CDN:
User(Seoul) ---- 50km ----> CDN Edge(Seoul) -- cache hit
RTT: ~5ms

On Cache Miss:
User(Seoul) -> CDN Edge(Seoul) -> Origin(US)
                | store in cache
         Next request served from Edge

1. TTL-based: Cache-Control: max-age=86400
2. Versioning: /static/app.v2.3.1.js
3. Manual invalidation: CloudFront Invalidation
4. Tag-based: Surrogate-Key header for selective purging

1. Cache Stampede (Thundering Herd):
   TTL expires -> thousands of requests hit DB simultaneously
   Solution: Mutex lock + pre-refresh

2. Cache Penetration:
   Repeated queries for non-existent keys -> every request hits DB
   Solution: Bloom Filter + cache empty results

3. Hot Key Problem:
   Traffic concentrated on specific key -> single cache node overloaded
   Solution: Local cache + key replication (key_1, key_2, ...)

SQL Selection Criteria:
- Complex data relationships with frequent joins
- ACID transactions essential (finance, payments)
- Data integrity is top priority
- Stable schema

NoSQL Selection Criteria:
- Data structure changes frequently
- Massive data volume (multiple TB+)
- Very low latency required
- Horizontal scaling essential
- Unstructured data (logs, social media)

Master-Slave Replication:
+-----------+     sync/async replication     +-----------+
|  Master   | =============================> |  Slave 1  |  <- Read-only
|  (Write)  | =============================> |  Slave 2  |  <- Read-only
+-----------+                                +-----------+

Multi-Master Replication:
+-----------+ <=============================> +-----------+
| Master 1  |    bidirectional replication     | Master 2  |
| (R/W)     |                                 | (R/W)     |
+-----------+                                 +-----------+
(Conflict resolution required)

Before Sharding:
+-----------------------+
|   Single Database     |
|   10TB, latency grows |
+-----------------------+

After Sharding (Hash-based):
hash(user_id) % 4 = shard_number

+-----------+ +-----------+ +-----------+ +-----------+
|  Shard 0  | |  Shard 1  | |  Shard 2  | |  Shard 3  |
|  2.5TB    | |  2.5TB    | |  2.5TB    | |  2.5TB    |
+-----------+ +-----------+ +-----------+ +-----------+

1. Join queries: Cross-shard joins impossible -> denormalization or application-level joins
2. Rebalancing: Shard redistribution on growth -> use Consistent Hashing
3. Hotspots: Traffic concentrated on specific shard -> redesign shard key
4. Transactions: Distributed transactions -> Saga pattern or 2PC
5. Unique IDs: Need globally unique IDs -> Snowflake ID, UUID

-- Normalized design (joins needed):
SELECT u.name, o.total, p.name AS product_name
FROM users u
JOIN orders o ON u.id = o.user_id
JOIN order_items oi ON o.id = oi.order_id
JOIN products p ON oi.product_id = p.id;

-- Denormalized design (single table):
-- orders table includes user_name, product_name
SELECT user_name, total, product_name
FROM denormalized_orders
WHERE user_id = 1001;

       Consistency
         /          \
        /            \
       /     CA       \
      /  (single DC)   \
     /------------------\
    |                    |
 CP |  Distributed Sys   | AP
    |  P is unavoidable  |
    |--------------------|
         Availability

During network partition:
CP system -> maintains consistency, rejects some requests
AP system -> maintains availability, allows temporary inconsistency

if (Partition) then
    Choose between Availability and Consistency
else
    Choose between Latency and Consistency

Examples:
- DynamoDB:   PA/EL (availability + low latency)
- MongoDB:    PC/EC (consistency-first)
- Cassandra:  PA/EL (availability + low latency)
- PostgreSQL(single): PC/EC (consistency + consistency)

Synchronous Processing:
Client -> API -> Send Email(3s) -> Resize Image(2s) -> Log(1s) -> Response
Total: 6 seconds waiting

Asynchronous Processing:
Client -> API -> Message Queue -> Response (immediately)
                    |
             +------+------+
             v             v
       Email Worker   Image Worker
         (3s)           (2s)
Total: 0.1 seconds waiting

Apache Kafka:
+-----------+     +---------------------------+
| Producer  |---->|  Topic: orders            |
+-----------+     |  +-------+-------+------+ |
                  |  |  P0   |  P1   |  P2  | |
                  |  +---+---+---+---+--+---+ |
                  +------+-------+------+-----+
                         v       v      v
                  +------------------------+
                  |  Consumer Group A       |
                  |  (one consumer per      |
                  |   partition)            |
                  +------------------------+

1. At-Most-Once:
   Producer -> Broker (no ACK wait)
   Fast but messages can be lost
   Use: Logs, metrics (loss acceptable)

2. At-Least-Once:
   Producer -> Broker -> ACK
   Retry on ACK failure -> duplicates possible
   Use: Most systems (idempotent processing needed)

3. Exactly-Once:
   Transaction-based or idempotency keys
   Performance overhead
   Use: Finance, payments (accuracy critical)

import time
import threading

class TokenBucket:
    def __init__(self, capacity, refill_rate):
        self.capacity = capacity        # Max tokens in bucket
        self.tokens = capacity           # Current tokens
        self.refill_rate = refill_rate   # Tokens added per second
        self.last_refill = time.time()
        self.lock = threading.Lock()

    def allow_request(self):
        with self.lock:
            now = time.time()
            # Refill tokens
            elapsed = now - self.last_refill
            self.tokens = min(
                self.capacity,
                self.tokens + elapsed * self.refill_rate
            )
            self.last_refill = now

            # Check if request allowed
            if self.tokens >= 1:
                self.tokens -= 1
                return True
            return False

# 10 requests/sec, burst up to 20
limiter = TokenBucket(capacity=20, refill_rate=10)

import time
from collections import deque

class SlidingWindowLog:
    def __init__(self, window_size, max_requests):
        self.window_size = window_size   # Window size (seconds)
        self.max_requests = max_requests
        self.requests = deque()

    def allow_request(self):
        now = time.time()
        # Remove old requests outside window
        while self.requests and self.requests[0] < now - self.window_size:
            self.requests.popleft()

        if len(self.requests) < self.max_requests:
            self.requests.append(now)
            return True
        return False

# Max 100 requests per minute
limiter = SlidingWindowLog(window_size=60, max_requests=100)

Sliding Window Counter (memory efficient):

Previous window count: 84
Current window count: 36
Window size: 60 seconds
Current position: 25% into window

Weighted count = 84 * (1 - 0.25) + 36 = 84 * 0.75 + 36 = 99

max_requests = 100 -> allowed (99 < 100)

Writes: 100M DAU * 0.1 URLs/day = 10M/day
Write QPS = 10,000,000 / 86,400 = ~116 QPS
Read QPS = 116 * 100 = ~11,600 QPS
Peak QPS = 11,600 * 3 = ~35,000 QPS

Storage (5 years):
10,000,000 * 365 * 5 = 18.25 billion records
Each record ~500 bytes -> 18.25B * 500B = ~9.1TB

+----------+     +------+     +---------------+
|  Client  |---->|  LB  |---->|  API Server   |
+----------+     +------+     +------+--------+
                                     |
                    +----------------+----------------+
                    v                v                v
              +-----------+  +-----------+  +-----------+
              |   Redis   |  |  MySQL    |  | Analytics |
              |   Cache   |  | (Sharded) |  |  (Kafka)  |
              +-----------+  +-----------+  +-----------+

import hashlib

def generate_short_url(long_url, counter):
    # Method 1: Base62 encoding (auto-increment ID)
    chars = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
    short = ""
    n = counter
    while n > 0:
        short = chars[n % 62] + short
        n //= 62
    return short  # e.g., "dK4x9Z"

    # Method 2: MD5 hash + Base62 (first 7 chars)
    hash_val = hashlib.md5(long_url.encode()).hexdigest()
    num = int(hash_val[:12], 16)
    short = ""
    for _ in range(7):
        short = chars[num % 62] + short
        num //= 62
    return short

# 62^7 = ~3.5 trillion unique URLs possible

Fan-out on Write (Push Model):
When user tweets -> pre-write to all followers' timelines

User A tweets -> Fan-out Service
                +--> Follower 1's timeline cache
                +--> Follower 2's timeline cache
                +--> Follower N's timeline cache

Pros: Reads are very fast (pre-computed)
Cons: Enormous write load for users with many followers

Fan-out on Read (Pull Model):
When timeline requested -> assemble tweets in real-time

User B requests timeline -> Query followed users
                           +--> User X's latest tweets
                           +--> User Y's latest tweets
                           +--> User Z's latest tweets
                           -> merge + sort -> return

Pros: Simple writes
Cons: Multiple DB queries on every read

Regular users (followers < 10,000): Fan-out on Write
Celebrities (followers > 10,000): Fan-out on Read

Timeline fetch:
1. Get pre-fanned-out tweets from cache
2. Query latest tweets from followed celebrities
3. Merge both results + sort by time

WebSocket-based Real-time Communication:

+----------+  WebSocket  +---------------+
| User A   |<==========>|  Chat         |
+----------+             |  Server      |
                         |  Cluster     |
+----------+  WebSocket  |              |
| User B   |<==========>|              |
+----------+             +------+-------+
                                |
                    +-----------+-----------+
                    v           v           v
              +----------+ +--------+ +----------+
              |  Redis   | | Kafka  | | Message  |
              | Presence | |        | |    DB    |
              +----------+ +--------+ +----------+

Method 1: Server timestamps (simple but clock sync issues)
Method 2: Snowflake ID (time + server + sequence)

Snowflake ID Structure (64-bit):
+---------------+----------+----------+--------------+
|  Timestamp    | Datactr  | Machine  |   Sequence   |
|   41 bits     |  5 bits  |  5 bits  |   12 bits    |
+---------------+----------+----------+--------------+
-> Time-sortable + globally unique
-> Can generate ~4 million IDs/sec

Heartbeat approach:
- Client sends heartbeat every 5 seconds
- Marked offline after 30 seconds without heartbeat
- Last heartbeat time stored in Redis

Status change notifications:
- Notify user's friend list on change
- For users with many friends: only notify currently online friends

Video Upload Pipeline:

+-----------+    +-----------+    +-------------------+
| Uploader  |--->|  Object   |--->|  Transcoding      |
|           |    |  Store    |    |  Pipeline          |
+-----------+    |  (S3)     |    |                    |
                 +-----------+    |  +---------------+ |
                                  |  | 240p          | |
                                  |  | 480p          | |
                                  |  | 720p          | |
                                  |  | 1080p         | |
                                  |  | 4K            | |
                                  |  +---------------+ |
                                  +---------+----------+
                                            v
                                  +-------------------+
                                  |   Deploy to CDN   |
                                  | (Global Edges)    |
                                  +-------------------+