Redis TTL, Key Expiry & Memory Management

Understanding how Redis expires keys and manages memory under pressure is essential for designing reliable, cost-effective Redis deployments.

---

Setting TTL

# Set with TTL in one command (preferred — atomic)
SET session:abc123 "user-data" EX 3600      # Expire in 3600 seconds
SET session:abc123 "user-data" PX 3600000   # Expire in 3600000 milliseconds
SET session:abc123 "user-data" EXAT 1700000000  # Expire at Unix timestamp (seconds)
SET session:abc123 "user-data" PXAT 1700000000000  # Expire at Unix timestamp (ms)

# Add TTL to existing key
EXPIRE key 3600            # Seconds
PEXPIRE key 3600000        # Milliseconds
EXPIREAT key 1700000000    # Unix timestamp (seconds)
PEXPIREAT key 1700000000000  # Unix timestamp (ms)

# Inspect TTL
TTL key          # Returns seconds remaining; -1 = no expiry; -2 = key doesn't exist
PTTL key         # Returns milliseconds remaining

# Remove TTL (make key persistent)
PERSIST key

# Redis 7.0+ EXPIRETIME: get exact expiry timestamp
EXPIRETIME key    # Returns Unix timestamp when key will expire

---

How Redis Expires Keys

Redis uses a hybrid expiry strategy — it does NOT check every key for expiry on every access:

1. Lazy Expiry (Passive)

When a key is accessed, Redis checks if it's expired and deletes it on the spot:

Client: GET expired-key
Redis:  1. Check if key has passed TTL → YES, it's expired
        2. Delete the key
        3. Return (nil)

Problem: Lazy expiry alone means expired keys continue consuming memory until they're accessed.

2. Active Expiry (Periodic)

Every 100ms, Redis runs a background job that:

1. Samples 20 random keys with TTLs set
2. Deletes all expired keys from the sample
3. If >25% of sampled keys were expired → repeat immediately
4. Stop when <25% of sampled keys are expired or time budget exhausted

Active sampling cycle:
  Sample 20 random keys with TTL
  Delete expired ones
  If expired rate > 25%:
    → Run another cycle (up to ~25% of time budget)
  Else:
    → Wait until next 100ms cycle

Why this matters: A key with EXPIRE 300 is not guaranteed to be deleted exactly at 300 seconds. Under memory pressure or many expiring keys, deletion can be delayed by 0–200ms typically, sometimes more.

Active Expiry Budget

Redis limits active expiry to prevent blocking:

maxmemory-expire-cycle = ACTIVE_EXPIRE_CYCLE_FAST  # 1ms max per cycle (default)
maxmemory-expire-cycle = ACTIVE_EXPIRE_CYCLE_SLOW  # 25% of CPU time

---

Memory Eviction Policies

When Redis exceeds maxmemory, it must evict keys. The policy is configured via maxmemory-policy:

maxmemory 2gb
maxmemory-policy allkeys-lru

Policy	Scope	Algorithm	Use Case
noeviction	All	Reject writes with OOM error	DB use (no data loss)
allkeys-lru	All keys	Evict least recently used	General cache (recommend)
allkeys-lfu	All keys	Evict least frequently used	Non-uniform access patterns
allkeys-random	All keys	Random eviction	Uniform access patterns
volatile-lru	Keys with TTL	Evict LRU among keys with TTL	Mixed DB + cache
volatile-lfu	Keys with TTL	Evict LFU among keys with TTL	Mixed DB + cache
volatile-random	Keys with TTL	Random among keys with TTL	Simple mixed use
volatile-ttl	Keys with TTL	Evict keys closest to expiry	Prioritize short-lived keys

LRU vs LFU

LRU (Least Recently Used):

• Evicts the key that was accessed least recently
• Problem: a key accessed millions of times but not in the last minute gets evicted

LFU (Least Frequently Used):

• Counts access frequency with decay over time
• Better for cache with clear "hot" and "cold" data
• Available since Redis 4.0

# Redis uses approximate LRU/LFU (not exact) for memory efficiency
maxmemory-samples 5    # Sample 5 keys to pick eviction candidate (higher = more accurate)
maxmemory-samples 10   # Better approximation, slightly more CPU

Eviction Policy Decision Guide

Is Redis used purely as a cache (no persistence needed)?
  → allkeys-lru or allkeys-lfu

Is Redis a primary store with some cached keys?
  → volatile-lru or volatile-lfu (only evict TTL keys, protect persistent data)

Is OOM rejection acceptable (you want to know when you're over capacity)?
  → noeviction (forces application-level capacity management)

Uniform access patterns (all keys equally likely)?
  → allkeys-random (fastest, no tracking overhead)

---

TTL Antipatterns and Gotchas

1. TTL Reset on Write

SET key "value" EX 3600   # TTL = 3600s
# ... 1800 seconds later ...
SET key "new-value"        # ❌ TTL is REMOVED! key is now persistent

# Fix: always re-set TTL on update
SET key "new-value" KEEPTTL   # Redis 6.0+: preserve existing TTL
# or
SET key "new-value" EX 3600   # Explicitly re-apply TTL

2. Stampede Attack on TTL

When many keys with the same TTL expire simultaneously, all clients make requests to the backend at once — overwhelming the database.

00:00:00 — SET product:* EX 3600 (1000 products cached)
01:00:00 — All 1000 expire simultaneously
01:00:00 — 1000 threads all hit the DB at once!

Fix: Jittered TTL

import random

def cache_product(product_id, data):
    base_ttl = 3600
    jitter = random.randint(0, 300)   # 0–5 min jitter
    redis.set(f"product:{product_id}", data, ex=base_ttl + jitter)

3. EXPIRE on Non-Existent Key

EXPIRE nonexistent-key 3600   # Returns 0 — key doesn't exist (no error, silently fails)
# Always verify the key exists before applying EXPIRE, or use SET ... EX

4. Volatile Policy with No TTL Keys

maxmemory-policy volatile-lru
# If all keys have no TTL → Redis CANNOT evict anything → OOM error on writes
# Always ensure cached keys have TTLs when using volatile-* policy

---

Memory Optimization Techniques

1. Object Sharing

Redis shares integer objects from 0 to 9999 — no memory allocation needed:

SET a 1000   # Points to shared integer object
SET b 1000   # Points to SAME shared integer object
# Memory cost: 1 object, not 2

2. Compact Encodings

Keep collections within encoding thresholds to use listpack:

# Good: listpack (compact array)
HSET user:1 name "Alice" age "30"        # 2 fields < 128 → listpack

# Bad: full hashtable
HSET user:1 f1 v1 f2 v2 ... f200 v200  # 200 fields → hashtable (much larger)

3. Key Compression

# Long key names waste memory
SET user:profile:john.doe@example.com:settings ...   # 43-byte key

# Shorter key names
SET up:12345:s ...   # 10-byte key (use numeric IDs, not emails)

4. Memory Analysis

MEMORY USAGE key          # Memory used by one key
MEMORY DOCTOR             # Redis's memory health diagnosis
MEMORY STATS              # Detailed memory statistics

# Find biggest keys (scan-based, safe for production)
redis-cli --bigkeys        # CLI tool — finds top keys by memory

# Object encoding inspection
OBJECT ENCODING key        # See ziplist, hashtable, intset, etc.
OBJECT FREQ key            # LFU frequency counter
OBJECT IDLETIME key        # Seconds since last access (for LRU)

---

Redis Keyspace Notifications

Subscribe to expiry and other key events:

# Enable in redis.conf
notify-keyspace-events Ex   # E=keyevent, x=expiry events
notify-keyspace-events KEA  # All events

# Subscribe to expired key events
SUBSCRIBE __keyevent@0__:expired

# Subscriber receives: key name when it expires
# Use case: triggers on session expiry, cleanup tasks, delayed processing

// Spring Data Redis keyspace notification listener
@Component
public class KeyExpirationListener extends KeyExpirationEventMessageListener {
    public KeyExpirationListener(RedisMessageListenerContainer container) {
        super(container);
    }

    @Override
    public void onMessage(Message message, byte[] pattern) {
        String expiredKey = message.toString();
        log.info("Key expired: {}", expiredKey);
        // Trigger cleanup, send notification, etc.
    }
}

Caveat: Keyspace notifications use Pub/Sub — they are at-most-once. If Redis is under load, some expiry notifications may be dropped.