After "When the Cache Never Warms Up", I went back and tuned the test itself. I wanted memtier to show some real cache hit rate during the run.

So I added automatic max-request calculation, letting the ECS tasks generate enough load to fill memory and push AWS ElastiCache closer to the edge. In R/R, I still leave some randomness, because my main goal there is to compare the broader differences between Redis architectures and node types.

Once that started to look better, I ran into the next problem: no evictions. Looking back, that was fair enough for a fit-to-memory case. But eviction is still an edge case I want to see in the lab, so that is where the next iteration goes.

Avg Throughput

1,875ops/sec

Peak Throughput

7,051ops/sec

Throughput CV

50.2%

Avg Latency

1.29ms

Max Latency

1.82ms

Avg Bandwidth

580KB/s

Active Load Window

59min

Pre-fill Duration

-29min

Avg ECS CPU

25.6%

Peak ECS CPU

52.5%

ECS Mem Peak

5MB

Loadgen Tasks

3

Avg Memory

53.96%

Max Memory

92.04%

Mem Headroom

+8.0%

Engine CPU Peak

17.2%

Cache Hit Rate

37.9%

Total Evictions

0

Min Free Mem

114MB

Peak Key Count

267,404

First Eviction

None

Memtier Benchmark

benchmark window only

Infrastructure

full CloudWatch window

ElastiCache Deep-Dive

full CloudWatch window