Redis vs Memcached: Choosing an In-Memory Data Store

Introduction

Both sit in front of your database and cache frequently accessed data in memory. Developers often use them interchangeably without understanding the differences. The differences matter: Redis is a data structure server that also happens to support caching. Memcached is a caching engine with a simpler data model. That distinction shapes what you can build with them and how you debug them at 2am.

This is not a “both are good” comparison. I will tell you when each one makes sense.

Core Concepts

Memcached stores strings and nothing but strings. You give it a key, you get back a value. That is the whole API.

Redis supports strings, lists, hashes, sets, sorted sets, bitmaps, hyperloglogs, geospatial indexes, and streams. It can act as a cache, a session store, a message broker, a rate limiter, and a real-time analytics engine. Note that volatile-lru and allkeys-lru use an approximated LRU algorithm (sampled LRU), not true LRU — Redis picks a random set of keys and evicts the least recently used among them. This is a memory-efficient approximation. volatile-lru applies the same sampled eviction only to keys with a TTL set, while allkeys-lru applies it across all keys.

# Memcached: everything is a string
memcached.set("user:123", json.dumps(user_data))
user_data = json.loads(memcached.get("user:123"))

# Redis: native data structures
redis.hset("user:123", mapping=user_data)
user_data = redis.hgetall("user:123")

Performance depends on what you are doing with them.

---

Redis Data Structures

Strings

Both handle simple string values. Redis just has more ways to manipulate them.

# Memcached
set key "value"
get key

# Redis
set key "value"
get key

# Redis extras
append key " more"     # Append to existing
incr count             # Atomic increment
decr count             # Atomic decrement
setrange key 0 "re"    # Overwrite bytes
getrange key 0 3       # Substring retrieval

Lists

Memcached does not have lists. Redis does.

# Redis lists: ordered, push/pop from either end
redis.lpush("queue:jobs", "job1", "job2", "job3")
redis.rpop("queue:jobs")  # Returns "job1" (oldest)
redis.lrange("queue:jobs", 0, -1)  # Get all

# Common use: recently viewed items, job queues, activity logs
redis.lpush("user:123:views", product_id)
redis.ltrim("user:123:views", 0, 19)  # Keep last 20

Sets and Sorted Sets

Memcached has no sets. Redis has both.

# Redis sets: unique, unordered
redis.sadd("user:123:likes", "product1", "product2", "product3")
redis.smembers("user:123:likes")
redis.sismember("user:123:likes", "product1")  # O(1) check

# Redis sorted sets: scored sets for leaderboards, priorities
redis.zadd("leaderboard", {"player1": 100, "player2": 200, "player3": 150})
redis.zrevrange("leaderboard", 0, 9, withscores=True)  # Top 10
redis.zrank("leaderboard", "player2")  # Get rank

Hashes

Memcached has no hashes. Redis does.

# Redis hashes: objects without serialization overhead
redis.hset("user:123", "name", "Alice", "email", "alice@example.com")
redis.hget("user:123", "name")  # "Alice"
redis.hgetall("user:123")  # All fields

# vs Memcached requiring JSON serialization
memcached.set("user:123", json.dumps({"name": "Alice", "email": "..."}))

Eviction Policies

Both support similar eviction policies when memory is full.

# Memcached eviction
# -no-eviction: return error on out-of-memory
# -allkeys-lru: evict least recently used of all keys
# -allkeys-random: evict random
# -volatile-lru: evict LRU of keys with TTL
# -volatile-ttl: evict shortest TTL
# -volatile-random: evict random of keys with TTL

memcached -o expire_counter,merge_threshold,ev=volatile-lru

# Redis maxmemory policies
# allkeys-lru, allkeys-random, allkeys-lfu, allkeys-ttl
# volatile-lru, volatile-lfu, volatile-random, volatile-ttl
# noeviction

maxmemory 100mb
maxmemory-policy allkeys-lru

The policies are nearly identical. Redis adds LFU (Least Frequently Used) which Memcached does not have.

Cache Invalidation Strategies

“Cache aside” (lazy loading) is the most common pattern, but there are several strategies with different trade-offs.

Write-Through Cache

Data is written to both cache and database synchronously. Reads always hit cache.

def set_user(user_id, data):
    # Write to cache AND database together
    redis.set(f"user:{user_id}", json.dumps(data))
    db.users.update(user_id, data)
    return data

def get_user(user_id):
    # Cache is always fresh
    cached = redis.get(f"user:{user_id}")
    if cached:
        return json.loads(cached)
    # Fallback only if cache miss
    data = db.users.get(user_id)
    redis.set(f"user:{user_id}", json.dumps(data))
    return data

Pros: Strong consistency — cache always matches database. Simple read path (always read from cache).

Cons: Write latency is higher (two writes). Cache can be stale if database write succeeds but cache write fails (use transactions).

Best for: Write-heavy workloads where data must always be current, configuration data, reference data.

Write-Behind Cache (Write-Back)

Data is written to cache only. Database is updated asynchronously.

def set_user(user_id, data):
    # Write to cache only — fast
    redis.set(f"user:{user_id}", json.dumps(data))
    # Async write to database via queue
    queue.enqueue("db:users:upsert", {"user_id": user_id, "data": data})
    return data

Pros: Very fast writes. Reduces database load during write spikes.

Cons: Risk of data loss if cache fails before database is updated. Requires additional infrastructure (write queue, retry logic). Cache can be inconsistent across nodes during propagation.

Best for: Write-heavy workloads where occasional data loss is acceptable (metrics, analytics, leaderboards), session data.

Cache-Aside (Lazy Loading)

The application manages cache explicitly — reads populate cache on miss, writes update database and invalidate cache.

def get_user(user_id):
    # Read from cache first
    cached = redis.get(f"user:{user_id}")
    if cached:
        return json.loads(cached)
    # Cache miss — load from database
    data = db.users.get(user_id)
    # Populate cache for next time
    redis.setex(f"user:{user_id}", 3600, json.dumps(data))
    return data

def set_user(user_id, data):
    # Write to database
    db.users.update(user_id, data)
    # Invalidate cache — do not update it
    redis.delete(f"user:{user_id}")
    return data

Pros: Simple. Cache never has stale writes (invalidated on update). Read-heavy workloads naturally populate cache. No write amplification.

Cons: Cache miss penalty — first read after invalidation or startup is slow. “Thundering herd” problem on popular keys.

Best for: Read-heavy workloads, most general-purpose caching, when database is source of truth.

TTL and Invalidation Details

TTL Selection Criteria

Choosing TTL is a trade-off between staleness and cache efficiency.

Data Type	Recommended TTL	Rationale
User sessions	24-48 hours	Sessions have natural expiry; 48h covers timezone gaps
Configuration data	5-15 minutes	Changes need to propagate; short enough to recover
API responses (public)	1-5 minutes	Fresh data important; short TTL limits staleness
Analytics / aggregates	15-60 minutes	Tolerates some staleness; longer = better hit rate
Product catalog	1-24 hours	Updates are infrequent; long TTL = better hit rate
Leaderboards	30-300 seconds	Needs near-real-time accuracy; short TTL required

Quick heuristic: Pick TTL based on how stale your data is allowed to be. If you cannot answer “how stale can this be?”, set it to 60 seconds or less.

Invalidation Patterns

Delete vs Expire:

# Delete: immediate removal
redis.delete("user:123")

# Expire: time-based removal
redis.setex("temp:data", 300, value)  # Auto-removes in 5 min

# Use expire for: temporary data, cached computations
# Use delete for: data that changed, explicit updates

Invalidation on update vs refresh on read:

# Option A: Invalidate on write (cache-aside)
def update_user(user_id, data):
    db.users.update(user_id, data)
    redis.delete(f"user:{user_id}")  # Next read fetches fresh

# Option B: Refresh on write (write-through variant)
def update_user(user_id, data):
    db.users.update(user_id, data)
    redis.setex(f"user:{user_id}", 3600, json.dumps(data))  # Write new value

# Option A (invalidate) is preferred because:
# - Avoids write amplification (only invalidates, does not rewrite)
# - Simpler: no need to serialize and store on every write
# - Cache stays consistent if write fails (delete does not happen)

Avoiding the Thundering Herd

When a popular cache key expires, many requests simultaneously hit the database.

# BAD: Many requests see cache miss, all hit database
def get_product(product_id):
    cached = redis.get(f"product:{product_id}")
    if not cached:
        data = db.products.get(product_id)  # ALL requests hit DB
        redis.setex(f"product:{product_id}", 300, json.dumps(data))
    return json.loads(cached) if cached else data

# GOOD: Single request refreshes, others wait
import threading
import time

def get_product_safe(product_id, lock_ttl=10):
    cached = redis.get(f"product:{product_id}")
    if cached:
        return json.loads(cached)

    lock_key = f"lock:product:{product_id}"
    # Try to acquire lock
    if redis.set(lock_key, "1", nx=True, ex=lock_ttl):
        # We got the lock — refresh from database
        data = db.products.get(product_id)
        redis.setex(f"product:{product_id}", 300, json.dumps(data))
        redis.delete(lock_key)
        return data
    else:
        # Another request is refreshing — wait and retry
        time.sleep(0.1)
        cached = redis.get(f"product:{product_id}")
        if cached:
            return json.loads(cached)
        return get_product_safe(product_id, lock_ttl)  # Retry

Alternative: Probabilistic early expiration (XFetch):

import hashlib
import random

def get_with_xfetch(key, beta=1.0):
    """XFetch: probabilistic early expiration to prevent thundering herd"""
    value = redis.get(key)
    if value:
        # Check if we should refresh early (probabilistic)
        ttl = redis.ttl(key)
        if ttl > 0:
            # Regenerate earlier if: random() < exp(-ttl/beta)
            if random.random() < math.exp(-ttl / beta):
                # Background refresh (in production, use separate thread/queue)
                return value, True  # "stale" flag to caller
    return value, False

# Usage: serve stale data while refreshing in background

Cache Warming and Cold Starts

When the Cache Starts Cold

When a cache starts empty (restart, deployment, failure), every request hits the database.

# BAD: Cold cache causes database overload at startup
# All 10,000 users hit DB simultaneously after Redis restart

# GOOD: Pre-warm cache before taking traffic
def warm_cache():
    """Run at startup before accepting traffic"""
    popular_keys = db.products.get_top_100()  # Identify hot data
    for product in popular_keys:
        redis.setex(f"product:{product.id}", 3600, json.dumps(product))
    # Now safe to take traffic

# Better: Progressive warming with rate limiting
def warm_cache_progressive():
    keys_to_warm = get_keys_by_priority()  # Sort by access frequency
    for i, key in enumerate(keys_to_warm):
        data = db.fetch(key)
        redis.setex(f"cache:{key}", get_ttl_for(key), data)
        # Rate limit: 1000 keys per second to avoid overwhelming DB
        if i % 1000 == 0:
            time.sleep(1)

Keeping Frequently-Used Data Hot

Monitor cache temperature — how often each key is accessed.

# Track key access frequency
def access_key(key):
    # Increment access counter (atomic)
    redis.hincrby("key:access", key, 1)
    return redis.get(key)

# Analyze access patterns weekly
def analyze_access():
    # Find keys not accessed in 7 days — low priority for warming
    # Find top 1000 accessed keys — priority for staying cached
    hot_keys = redis.zrevrange("key:access", 0, 999, withscores=True)

    # Set aggressive TTL on hot keys
    for key, score in hot_keys:
        current_ttl = redis.ttl(f"cache:{key}")
        if current_ttl < 3600:  # Less than 1 hour
            redis.expire(f"cache:{key}", 86400)  # Extend to 24 hours

Warming Patterns in Practice

# Pattern 1: Scheduled pre-warming before high-traffic events
def warm_for_black_friday():
    """Run 30 minutes before expected traffic spike"""
    # Pre-compute and cache popular product pages
    top_products = db.get_products_by_category("popular", limit=500)
    for product in top_products:
        redis.setex(f"product:{product.id}", 7200, compute_product_page(product))
    # Pre-warm user sessions that will be active
    active_user_ids = db.get_users_logged_in_recently(limit=10000)
    for user_id in active_user_ids:
        redis.setex(f"session:{user_id}", 172800, load_user_session(user_id))

# Pattern 2: Proactive caching on database write
def write_with_proactive_cache(user_id, data):
    # Write to database
    db.users.update(user_id, data)
    # Proactively cache the result
    redis.setex(f"user:{user_id}", 86400, json.dumps(data))
    # Also cache related data
    redis.setex(f"user:{user_id}:profile", 86400, json.dumps(data["profile"]))

# Pattern 3: Background refresh for critical keys
def start_background_refresh(key, compute_fn, ttl=300):
    """Refresh key in background before expiry"""
    def refresh():
        while True:
            value = compute_fn()
            redis.setex(key, ttl, value)
            time.sleep(ttl * 0.8)  # Refresh at 80% of TTL
    thread = threading.Thread(target=refresh, daemon=True)
    thread.start()

Persistence

Memcached: Pure Memory

Memcached is pure memory. It never touches disk. When it restarts, everything is gone.

# Memcached has no persistence options
# Restart = empty cache

This sounds like a drawback, but for pure caching it is fine. Your source of truth is the database anyway.

Redis: Optional Persistence

Redis persists to disk. You can survive restarts without losing data.

# RDB snapshots: point-in-time dumps
save 900 1    # Save if 1 key changed in 900 seconds
save 300 10   # Save if 10 keys changed in 300 seconds
save 60 10000 # Save if 10000 keys changed in 60 seconds

# AOF (Append Only File): every write logged
appendonly yes
appendfsync everysec  # fsync every second (balance of speed/safety)

# Or no persistence at all (pure cache mode)
save ""

Redis persistence is configurable. You can turn it off for pure caching or enable it for durability.

Performance and Clustering

Performance

Raw performance depends on your workload. Here is a general comparison:

Operation	Memcached	Redis
GET/SET (simple)	Very fast	Fast
MGET/MSET (batch)	Faster	Slower (per-key overhead)
INCR (atomic counter)	Fast	Very fast
Sets/Lists/Hashes	Not supported	Depends on operation
Memory efficiency	Better (simple values)	Depends on data structures

For simple string caching, Memcached often uses less memory per key. For complex data structures, Redis’s overhead is usually worth it.

Redis uses single-threaded execution (one command at a time per connection, but multiple connections). Memcached is multi-threaded. On a single instance, Redis can saturate network bandwidth. Memcached scales better on multi-core for raw throughput.

# Redis pipelining: batch commands to reduce round trips
pipe = redis.pipeline()
for key in keys:
    pipe.get(key)
results = pipe.execute()  # One round trip for all

Clustering and Distribution

Memcached

Memcached has no native clustering. You shard across instances manually using consistent hashing.

import hashlib

class ConsistentHash:
    def __init__(self, servers):
        self.servers = servers
        self.ring = {}
        self.sorted_keys = []

        for server in servers:
            for i in range(150):
                key = f"{server}:{i}"
                hash_key = int(hashlib.md5(key.encode()).hexdigest(), 16)
                self.ring[hash_key] = server
                self.sorted_keys.append(hash_key)

        self.sorted_keys.sort()

    def get_server(self, key):
        hash_key = int(hashlib.md5(key.encode()).hexdigest(), 16)
        for sorted_key in self.sorted_keys:
            if hash_key <= sorted_key:
                return self.ring[sorted_key]
        return self.ring[self.sorted_keys[0]]

It works. You are just managing the sharding yourself.

Redis

Redis has built-in clustering with automatic sharding.

# Redis Cluster configuration
cluster-enabled yes
cluster-config-file nodes.conf
cluster-node-timeout 15000

# Automatic sharding, replication, and failover
# Your application sees a single logical database

Redis Cluster partitions keys across nodes automatically. It also supports replication for read scaling and failover.

Capacity Estimation: Memory-per-Key and Cluster Slot Planning

Memory per key differs significantly between Redis and Memcached.

Redis memory breakdown per key (string value):

Component	Size
Key pointer	~56 bytes (SDS allocator)
Value storage	Actual value size
Redis object overhead	~16 bytes
Dictionary entry (if in hash)	~32 bytes
Total minimum per key	~72 bytes + value

Memcached memory breakdown per key (string value):

Component	Size
Key	Key length
Value	Actual value size
flags byte	1 byte
CAS token (optional)	8 bytes
Expiry time	4 bytes
Overhead per item	~25 bytes
Total minimum per item	~25 bytes + key + value

For a cache with 1 million keys, each storing a 200-byte value:

• Redis string: ~72M overhead + 200M data = ~272M total
• Memcached: ~25M overhead + key_space + 200M data = ~240M + key_space

Memcached wins on simple string workloads by 10-20% memory efficiency. Redis pays the overhead for richer data structures.

Redis Cluster slot planning: 16,384 slots divided across N master nodes. For a 6-node cluster (3 masters + 3 replicas), each master owns ~5,461 slots. Slot ownership determines which node stores which keys. The formula: slot = CRC16(key) % 16384. When planning capacity, ensure each master has headroom — if one master owns 5,461 slots and your average key is 1KB with 100K keys per slot, that node needs roughly 5GB. Plan for 2x headroom.

Memcached cluster sizing: No slots — consistent hashing distributes keys. Target 150-200 virtual nodes per physical node for even distribution. With N nodes and V virtual nodes each, the coefficient of variation (CV) of key distribution should stay below 0.3. Formula: CV ≈ 1/√(N × V). For CV < 0.3 with V = 150, you need N ≥ 12 nodes for even distribution. Fewer nodes means higher variance in distribution.

Comparative Analysis Tables

Cache Invalidation Strategy Comparison

Strategy	Write Latency	Read Consistency	Data Loss Risk	Complexity	Best For
Write-Through	High (sync)	Always fresh	None	Low	Write-heavy, consistency-critical
Write-Behind	Low (async)	Eventually consistent	High	High	Write spikes, high-throughput
Cache-Aside	Low (1 write)	Strong (on invalidation)	Low	Medium	Read-heavy, general purpose

Redis vs Memcached: When to Use Which

Factor	Redis	Memcached	Winner
Data structures	Strings, Lists, Sets, Hashes, Sorted Sets	Strings only	Redis
Memory efficiency	Higher per-key overhead (~72 bytes + value)	Lower per-key overhead (~25 bytes + value)	Memcached
Persistence	RDB snapshots, AOF logs	None (pure memory)	Redis
Clustering	Built-in cluster with slots	Client-side consistent hashing	Redis
Threading model	Single-threaded (no locks)	Multi-threaded (global lock per operation)	Memcached (throughput), Redis (consistency)
Atomic operations	INCR, SETNX, Lua scripts	CAS tokens only	Redis
Pub/Sub	Native support	Not supported	Redis
Latency	Sub-millisecond, predictable	Sub-millisecond, predictable	Tie
Operational complexity	Higher (config, persistence)	Lower (stateless)	Memcached
Production maturity	Very mature at scale	Mature	Tie

Eviction Policy Comparison

Policy	Redis Support	Memcached Support	Behavior
LRU (Least Recently Used)	allkeys-lru, volatile-lru	allkeys-lru, volatile-lru	Evict least recently accessed
LFU (Least Frequently Used)	allkeys-lfu, volatile-lfu	Not supported	Evict least frequently accessed
TTL	volatile-ttl	volatile-ttl	Evict shortest TTL first
Random	allkeys-random, volatile-random	allkeys-random, volatile-random	Random eviction
No eviction	noeviction	noeviction	Return error on OOM

Redis LFU advantage: LFU tracks frequency, not just recency. For data that is accessed frequently for a period then becomes cold, LFU prevents it from being evicted by a single recent access spike. Memcached does not have this capability.

Connection Management Trade-offs

Aspect	Single Connection	Connection Pool	Presized Pool
Setup cost	High (connect latency)	Medium (pool creation)	Low
Concurrent requests	Poor (blocks)	Good	Best
Resource usage	Low (1 socket)	Medium (N sockets)	Medium
Complexity	Simple	Moderate	Simple
Best for	Scripts, short-lived	Web applications	High-throughput

When to Use / When Not to Use

Memcached makes sense for simple string caching — HTML fragments, API responses, session data — where maximum memory efficiency matters and you do not need complex data structures. It scales horizontally via consistent hashing, and the operational surface is small. Use it for database query results that fit naturally in key-value form, and for things that do not change often and benefit from sub-millisecond access.

Redis makes sense when you need lists, sets, sorted sets, or hashes; when you want optional persistence; when you need atomic counters for rate limiting or distributed locks; when you need pub/sub for real-time features or chat; when you want built-in clustering; or when you are building leaderboards, job queues, or caching with complex data access patterns. Lua scripting adds atomic multi-step operations without race conditions.

---

A Practical Decision Framework

Do you need anything beyond simple string key-value?
  YES -> Redis
  NO  -> Does memory efficiency matter more than features?
          YES -> Memcached
          NO  -> Redis (for easier operations and clustering)

If you are not sure, start with Redis. The extra memory usage is negligible for most workloads. If you later find memory is tight and profiling shows Memcached is meaningfully better, switch.

---

Production Failure Scenarios and Trade-off Analysis

Failure	Impact	Mitigation
Redis/Memcached OOM	Cache returns errors; application falls back to database	Monitor used_memory/maxmemory ratio; set alerts at 70% threshold
Redis fork for RDB save	Brief blocking during fork; memory doubles during copy-on-write	Schedule RDB saves during low-traffic; use AOF instead for persistence
Memcached restart	All data lost immediately (no persistence)	Design for cold cache; implement application-level cache warming
Redis replica lag	Reads from replica may return stale data	Monitor replication_backlog_histlen; read from primary for consistency-critical data
Connection pool exhaustion	Requests timeout waiting for connection	Size connection pool appropriately; implement request queuing with timeout
Single-threaded Redis blocking	Long commands block all other commands	Avoid KEYS, SMEMBERS on large sets; use pipeline/batch operations
Memcached multi-thread contention	High CPU under heavy load	Scale horizontally with consistent hashing; consider Redis for complex workloads

Detailed Failure Scenarios

Case 1: Redis OOM During Peak Traffic

What happened: A Redis instance reached maxmemory during a flash sale. Eviction policy was noeviction. Redis started returning errors instead of serving requests.

Root cause: The maxmemory-policy was set to noeviction (return error on OOM) instead of allkeys-lru. Additionally, the application was not handling cache errors gracefully — it failed fast instead of falling back to the database.

Impact: 12% of requests failed during a 45-minute window. The database was under-provisioned for the fallback load and also started timing out.

Lesson learned: Always use an eviction policy that allows Redis to keep serving. Implement circuit breakers so the application falls back to the database gracefully when cache errors spike. Monitor evicted_keys and used_memory metrics.

Case 2: Memcached Restart Storm

What happened: A Memcached node was restarted for a configuration update. Within 90 seconds, the database was overwhelmed by cold-cache requests from all application servers simultaneously.

Root cause: No cache warming strategy. All 50 application instances started with empty local caches and hit the database for the same popular keys simultaneously. The database had no protection against this concurrent access pattern.

Impact: Average response time jumped from 15ms to 8,400ms. Database CPU hit 100%. The restart took 3 minutes longer than expected because the database was too overloaded to respond to health checks.

Lesson learned: Implement cache warming before taking a cache node out of rotation. Use consistent hashing with virtual nodes so individual key popularity does not spike on single nodes after restart. Consider using a local L1 cache (in-memory LRU) in front of Memcached to absorb cold-start load.

Case 3: Redis Pipeline Blocking on Large Set

What happened: A developer ran redis-cli --bigkeys on a production Redis instance during peak hours to find memory-heavy keys.

Root cause: The --bigkeys flag performs a full SCAN and evaluates every key’s memory footprint. On a 50GB Redis instance with millions of keys, this command consumed 15 seconds of CPU and blocked all other commands during that window.

Impact: P99 latency spiked from 5ms to 12,000ms. The application saw 200+ connection timeouts. The on-call engineer spent 45 minutes diagnosing why Redis was suddenly unresponsive.

Lesson learned: Never run memory introspection commands (--bigkeys, MEMORY USAGE on unknown keys, KEYS *) on production Redis instances. Use SCAN with COUNT limits for any introspection, and always run it during maintenance windows. For memory analysis, use Redis MEMORY STATS and INFO memory instead.

Common Pitfalls and Anti-Patterns

1. Using KEYS Command in Production

The KEYS command scans all keys and blocks Redis. Never use it in production.

# BAD: KEYS blocks Redis for seconds
all_keys = redis.keys("user:*")

# GOOD: Use SCAN for production
cursor = 0
while True:
    cursor, keys = redis.scan(cursor, match="user:*", count=100)
    process(keys)
    if cursor == 0:
        break

2. Storing Large Values Without Compression

Large values consume memory disproportionately and slow down operations.

# BAD: Storing large uncompressed data
redis.set("page:123", large_html_content)  # 500KB+ per page

# GOOD: Compress large values
import zlib
compressed = zlib.compress(large_html_content.encode())
redis.setex("page:123", 3600, compressed)

3. Not Using Connection Pooling

Each operation creating a new connection adds overhead.

# BAD: New connection each time
def get_user(user_id):
    r = redis.Redis(host='localhost', port=6379)  # Connection every call
    return r.get(f"user:{user_id}")

# GOOD: Reuse connection
pool = redis.ConnectionPool(host='localhost', port=6379, max_connections=50)

def get_user(user_id):
    r = redis.Redis(connection_pool=pool)
    return r.get(f"user:{user_id}")

4. Ignoring Memcached Persistence Limitations

Memcached has no persistence. Data is lost on restart.

# BAD: Assuming Memcached persists data
memcached.set("session:123", session_data)
# ... server restarts ...
session = memcached.get("session:123")  # None - data gone

# GOOD: Design for cold start
session = memcached.get("session:123")
if not session:
    session = load_from_database()  # Always have fallback
    memcached.set("session:123", session, time=3600)

5. Using Redis Single Instance for Everything

Redis single-threaded nature means CPU-bound operations block everything.

# BAD: CPU-heavy operation in Redis
# This blocks all other commands
redis.sort("large-set")  # O(N log N) - blocks Redis

# GOOD: Move CPU work to application
data = redis.lrange("large-list", 0, -1)
sorted_data = sorted(data)  # Application handles sorting

Quick Recap

Redis offers data structures — lists, sets, hashes — that Memcached cannot match. Memcached is more memory-efficient for simple strings. Redis persistence (RDB/AOF) survives restarts; Memcached does not. Redis Cluster provides automatic sharding; Memcached requires client-side sharding. Both support LRU/LFU eviction but Redis LFU is more sophisticated. Redis single-threaded is a feature — no race conditions — but CPU-heavy operations block everything.

Best Practices Summary

Redis: use connection pooling (never a new connection per request); set maxmemory and an eviction policy like allkeys-lru for most caching workloads; never run KEYS, SMEMBERS, or SORT on large sets; use pipelining for batch operations; enable slow log monitoring at 10ms threshold; use hashes for objects instead of JSON strings; set reasonable TTLs; rename dangerous commands in production (rename-command FLUSHDB ""); monitor memory fragmentation (mem_fragmentation_ratio > 1.5 indicates problems); use Lua scripts for atomic multi-step operations.

Memcached: use consistent hashing for sharding with 150-200 virtual nodes per physical node; use consistent key naming with service prefixes (users:123, sessions:abc); store serialized data efficiently with MessagePack or Protocol Buffers instead of JSON for 20-30% smaller payloads; set appropriate chunk size (default 1MB may waste memory for small values); monitor evictions — high rates indicate cache is too small or TTLs are misconfigured; prefer UDP for get operations in read-heavy workloads; use multi-get for batch operations.

Observability Checklist

Security Checklist

• Enable authentication — Redis 6+ supports ACLs. Memcached supports SASL authentication. Never run without auth in production.
• Bind to internal IPs only — bind 127.0.0.1 or bind 10.0.0.0/8 to prevent unauthorized access. No public IP exposure.
• Encrypt in transit — Use TLS for Redis and Memcached if crossing network boundaries. Redis 6+ has native TLS support.
• Limit commands — Rename dangerous commands: rename-command FLUSHDB "" rename-command CONFIG "" rename-command KEYS "".
• Set maxmemory — Prevent cache from consuming all available RAM and causing system instability.
• Use firewall rules — Restrict access to cache ports (6379 for Redis, 11211 for Memcached) to application servers only.

Metrics to Track

Redis:

# Core metrics via Redis INFO
INFO memory  # used_memory, maxmemory, mem_fragmentation_ratio
INFO stats   # total_commands_processed, keyspace_hits, keyspace_misses
INFO replication  # master_link_status, slave_read_only, replication_lag
INFO clients  # connected_clients, blocked_clients

# Calculate hit rate
# hit_rate = keyspace_hits / (keyspace_hits + keyspace_misses)

Memcached:

# Stats command
stats
# Items: curr_items, total_items, evictions
# Memory: bytes, limit_maxbytes
# Hit rate: get_hits, get_misses

# Calculate hit rate
# hit_rate = get_hits / (get_hits + get_misses)

Logs to Capture

import structlog
import time

logger = structlog.get_logger()

class CacheMetrics:
    def __init__(self, redis_client):
        self.redis = redis_client
        self.start_time = time.time()

    def track_operation(self, operation, key, hit=True):
        logger.info("cache_operation",
            operation=operation,
            key=key,
            cache_hit=hit,
            latency_ms=self._measure_latency()
        )

    def log_memory_pressure(self):
        info = self.redis.info('memory')
        used = info['used_memory']
        maxmem = info['maxmemory']

        if maxmem > 0 and used / maxmem > 0.8:
            logger.warning("cache_memory_critical",
                used_mb=used / 1024 / 1024,
                max_mb=maxmem / 1024 / 1024,
                fragmentation=info.get('mem_fragmentation_ratio', 1))

Interview Questions

Further Reading

Official Documentation

• Redis Documentation — Official Redis docs, including data types, persistence, and clustering guides
• Memcached Wiki — Memcached official wiki with usage examples and configuration options
• Redis Cluster Tutorial — Scaling Redis with cluster mode

Related Articles

• Cache Eviction Policies — How LRU, LFU, and other policies work in Redis and Memcached
• Caching Strategies — How to use these tools effectively
• Distributed Caching — Scaling Redis and Memcached across multiple nodes

External Resources

• Redis Persistence Demystified — Understanding RDB and AOF trade-offs
• Memcached vs Redis: Choosing an In-Memory Store — Instagram’s migration case study
• Redis Lua Scripts Best Practices — Atomic operations with Lua
• Cache Warming Patterns — AWS guidance on cache strategies

Performance Tuning

• Redis Latency Problems Troubleshooting — Diagnosing and fixing Redis latency issues
• Memcached Tuning Guide — Configuration tuning for different workloads

---

Conclusion

Memcached is simpler and more memory-efficient for pure string caching. Redis is more capable. For basic caching, they are comparable. But Redis’s data structures unlock patterns that would be painful or impossible with Memcached.

I default to Redis for new projects. The operational simplicity of having one system for caching, sessions, pub/sub, and rate limiting usually beats the memory efficiency gains of Memcached.

That said, if you are caching primarily string data and memory is tight, Memcached still earns its place.