Performance Tuning

Introduction

Redis performance tuning starts with pipelining, connection pooling, right-sized values, and eliminating slow commands — not buying bigger hardware first. Default configs target dev laptops; production needs lazyfree, UNLINK, disabled THP, and co-located AZ placement.

Profile with redis-benchmark only as sanity check — real workloads need application-level tracing with realistic value sizes and concurrent clients.

Target p99 latency not average — one KEYS command spikes p99 for everyone.

Understanding the topic

Key concepts

• Pipelining — batch RTT.
• Connection pool max-active sizing.
• UNLINK async delete vs blocking DEL.
• lazyfree-lazy-user-del large key free.
• Disable transparent huge pages Linux.
• I/O threads help read/write not command CPU.

Step-by-step explanation

1. Measure baseline ops/sec and p99 latency.
2. Identify RTT-bound vs CPU-bound paths.
3. Apply pipeline/MGET to RTT issues.
4. Replace DEL with UNLINK on big keys.
5. Tune pool and co-locate network.
6. Re-measure under load test.

Syntax reference

Common commands

• UNLINK non-blocking delete large keys.
• activedefrag when mem_fragmentation_ratio high.
• LATENCY DOCTOR first tuning step.

CONFIG SET lazyfree-lazy-user-del yes
CONFIG SET activedefrag yes
UNLINK big:hash:key
INFO stats | grep instantaneous_ops
LATENCY DOCTOR

Informative example

Enable lazy free and measure before/after large hash delete:

redis-cli CONFIG SET lazyfree-lazy-user-del yes
redis-cli DEBUG SLEEP 0
# Before: time redis-cli DEL huge:hash
# After:  time redis-cli UNLINK huge:hash
redis-cli INFO stats | grep lazyfree_pending_objects

Spring: configure Lettuce pool max-active ~2× expected concurrent Redis users per pod.

Real-world use

Real-world use cases

• API p99 SLA breach on cache layer.
• Bulk delete migration without latency spike.
• Flash traffic event capacity planning.
• Reduce ElastiCache node size after optimization.
• Post-deploy regression detection.

Best practices

• Pipeline all batch read paths.
• Pool size from concurrency not guess.
• UNLINK not DEL for >1KB aggregates.
• Same AZ app and Redis.
• Disable THP; set overcommit_memory.
• Alert on slowlog growth.

Common mistakes

• Benchmark only SET GET 10 byte locally.
• Pool max-active=1 on high traffic pod.
• DEL million-field hash at peak.
• More Redis CPU without fixing hot key.

Advanced interview questions

Summary

Pipeline and pool sizing — first wins.