Amazon ElastiCache

Engine-level Redis depth (data structures, persistence, clustering internals, distributed locks, cache-stampede mitigation) lives in Redis. Operational depth (failover, Serverless, parameter groups) is in ElastiCache — Failover, Serverless, and Operations. This file is the AWS-managed-service overview.

Managed Redis / Memcached. Saves you from operating a cache cluster. Two engines:

• ElastiCache for Redis / Valkey — feature-rich; pub/sub, sorted sets, persistence, streams.
• ElastiCache for Memcached — multi-threaded simple KV; rarely the right choice in 2024+.

Almost always: Redis (or Valkey, the open-source AWS-backed fork after Redis Inc.’s license change).

What ElastiCache gives you

• Provisioning, patching, monitoring.
• Multi-AZ replication with automatic failover.
• Backup/restore (RDB snapshots).
• Encryption in transit (TLS) and at rest.
• IAM-based authentication.

For Redis fundamentals — data structures, eviction, caching patterns — see Redis.

Deployment topologies

Single node

A single primary, no replica. Cheap, no HA. Use only for dev/staging.

Replication group (cluster mode disabled)

One primary + up to 5 replicas. Automatic failover. Single primary — write throughput limited by one node. Reads can scale across replicas.

Cluster mode enabled

Sharded. Data partitioned across N “node groups” (shards), each with primary + replicas. Hash slots distribute keys. Use when you outgrow a single primary’s write throughput.

# Cluster mode requires a cluster-aware client
from redis.cluster import RedisCluster
r = RedisCluster(host="prod.xxx.use1.cache.amazonaws.com", port=6379, ssl=True)

Trade-off: cluster mode complicates multi-key operations. MGET / pipelines across multiple slots need extra logic; transactions can’t span slots.

Endpoints

# cluster-mode-disabled — separate read/write endpoints
write = redis.Redis(host="prod.xxx.use1.cache.amazonaws.com", port=6379, ssl=True)
read  = redis.Redis(host="prod-ro.xxx.use1.cache.amazonaws.com", port=6379, ssl=True)

Versioning and feature flags

ElastiCache lags upstream Redis by months. Check the version map before relying on a specific feature (Redis Streams, JSON, RedisGraph, etc.).

Newer choice: ElastiCache Serverless (2023+). No node sizing — usage-based pricing per byte stored + per request. Good for unpredictable load.

Sizing

Memory is king. Watch:

• BytesUsedForCache — actual data size.
• DatabaseMemoryUsagePercentage — running close to 100% leads to evictions.
• Evictions — keys being dropped because the cache is full.
• CurrConnections — Redis has a default limit of 65k; each app pod opens N.

CPU rarely matters except for Lua scripts or KEYS * in production (don’t).

Persistence options

• RDB snapshots — point-in-time dump. Daily by default; restore creates a fresh cluster from a snapshot.
• AOF — append-only log of every write. ElastiCache exposes limited AOF — usually you’d use snapshots.

If your cache is only a cache (rebuildable from source of truth), don’t persist. If it’s a primary store (rate-limit counters, sessions), persist + replicate + back up.

Connection patterns for Python

import redis
from redis.connection import SSLConnection

# Cluster-mode-disabled
pool = redis.ConnectionPool(
    host="prod.xxx.use1.cache.amazonaws.com",
    port=6379,
    connection_class=SSLConnection,
    max_connections=50,
    decode_responses=True,
)
r = redis.Redis(connection_pool=pool)

For cluster mode, use redis.cluster.RedisCluster (sync) or redis.asyncio.cluster.RedisCluster (async, redis-py 5+).

Pool sizing on k8s

N pods × max_connections per pod = total. ElastiCache caps at 65k. With 100 pods at 100 connections each you’re at 10k — fine. With 10000 lambdas? You need an in-front proxy or fewer connections per worker.

Auth and encryption

aws elasticache create-replication-group \
  --replication-group-id orders \
  --engine redis \
  --transit-encryption-enabled \
  --at-rest-encryption-enabled \
  --auth-token "long-random-string" \
  ...

Client supplies the auth token as password. IAM auth (newer) replaces shared tokens with short-lived IAM-derived tokens — same model as RDS IAM auth.

Cache-aside is the dominant pattern

def get_user(user_id: int) -> User:
    key = f"user:{user_id}"
    raw = r.get(key)
    if raw:
        return User.parse_raw(raw)
    user = db.query(User).get(user_id)
    if user:
        r.setex(key, 3600, user.json())  # 1h TTL
    return user

Plus invalidation on write:

def update_user(user_id: int, **fields):
    db.update_user(user_id, **fields)
    r.delete(f"user:{user_id}")  # invalidate

Caveats:

• Cache stampede when a hot key expires — see Redis.
• Stale on race — concurrent update + read can re-cache stale data. Mitigate with write-through or short TTLs.

Common gotchas

• KEYS * in production. Blocks the single-threaded server. Use SCAN.
• Hot key (sharded mode). All traffic for one key hits one shard. Replicate the value to N suffixed keys.
• TLS without ssl=True in client. Connections refused. Easy to miss in dev → prod transition.
• Memcached vs Redis interchangeability. Memcached is just a key/value cache — no pub/sub, no sorted sets, no streams, no persistence. Don’t pick it unless you know why.
• Default maxmemory-policy: noeviction in ElastiCache. Hitting the limit returns errors instead of evicting. Set to allkeys-lru for a generic cache.
• Snapshot restore creates a NEW cluster. No in-place rollback.

Cost optimization

• Reserved nodes for steady workloads (~50% off).
• Right-size memory — cache.r6g.large (13GB) costs 4× cache.t4g.small (1.5GB); a too-large instance with 10% memory used is waste.
• Serverless for spiky / dev / staging — pay per byte + request.
• Move to Valkey — same engine post-fork, AWS bills it slightly differently; some price savings appearing.

Common interview pattern: sessions / rate limit

# Rate limit: max 100 requests per minute per user
def check_rate(user_id: int) -> bool:
    key = f"rate:{user_id}:{int(time.time() // 60)}"
    count = r.incr(key)
    if count == 1:
        r.expire(key, 60)
    return count <= 100

# Session token
r.setex(f"sess:{token}", 3600, json.dumps(session_data))