Memory Optimization

Introduction

Redis memory is your primary bill and failure mode — maxmemory triggers eviction or OOM. Right-sizing types (hash not JSON blob), eviction policy choice, and active defrag can cut ElastiCache costs 20–30% without code rewrites.

Run redis-cli --bigkeys and MEMORY DOCTOR monthly. Fragmentation ratio above 1.5 means RSS exceeds logical memory — activedefrag or restart during maintenance.

Every key without TTL is a future incident.

Understanding the topic

Key concepts

• used_memory vs used_memory_rss — fragmentation.
• maxmemory + maxmemory-policy.
• allkeys-lru vs volatile-lru vs noeviction.
• Encodings — intset, listpack, embstr savings.
• activedefrag — online defragmentation.
• MEMORY USAGE key — bytes per key.

Step-by-step explanation

1. Allocator (jemalloc) tracks logical usage.
2. Updates cause holes — RSS grows.
3. maxmemory hit triggers eviction algorithm.
4. noeviction returns errors on write.
5. Defrag relocates keys compacting memory.

Syntax reference

Common commands

• --bigkeys off-peak only.
• noeviction for queue workloads needing no silent loss.
• allkeys-lru when most keys should be evictable.

INFO memory
MEMORY DOCTOR
MEMORY USAGE session:abc
redis-cli --bigkeys -i 0.1
CONFIG SET maxmemory 4gb
CONFIG SET maxmemory-policy allkeys-lru

Informative example

Find top memory consumers and compare hash vs JSON string:

redis-cli --bigkeys
# Sample output shows biggest string/hash/zset

MEMORY USAGE user:42:string_json
MEMORY USAGE user:42:hash_fields
# Hash often smaller when multiple fields updated

Migrate hot JSON keys to hash field-by-field in background job. Monitor used_memory drop.

Real-world use

Real-world use cases

• Reduce AWS ElastiCache bill.
• Prevent OOM before product launch.
• Choose HLL over SET for UV.
• Eviction tuning for pure cache tier.
• Fragmentation fix after big delete.

Best practices

• Set maxmemory 75–80% RAM on self-hosted.
• TTL every ephemeral key.
• allkeys-lru for cache; noeviction for job queues.
• Enable activedefrag when supported.
• Right type and encoding awareness.
• Alert at 80% used_memory.

Common mistakes

• noeviction + no maxmemory on cache cluster.
• Giant keys never split or trimmed.
• Ignoring fragmentation until OOM kill.
• volatile-lru but keys lack TTL — never evict.

Advanced interview questions

Summary

Memory = cost and reliability boundary.