agent's idle check used cpu only - the memory-bound redis fleet got killed

TL;DR

The agent declared instances idle on low CPU, but the Redis fleet is memory-bound by design: low CPU with high memory use is its healthy state. Change the idle criteria to require low utilization across CPU and memory (and workload-appropriate metrics like cache hit rate) before flagging. Single-metric idle detection misclassifies every workload whose bottleneck is not CPU.

Error

agent's idle check used cpu only  -  the memory-bound redis fleet got killed

Steps

1. Confirm the misclassification: the Redis fleet showed low CPU but high memory utilization. Expected: the memory-bound profile is visible in the metrics.
2. Restore the affected instances or fail over to replicas if the kill already happened. Expected: the service is healthy again.
3. Rewrite the idle criteria: require low CPU and low memory (and low network for good measure) sustained over the evaluation window. Expected: memory-bound workloads no longer qualify as idle.
4. Add workload-type awareness: data stores, caches, and queues get memory-weighted criteria. Expected: the agent treats workload classes differently.
5. Re-run the idle analysis. Expected: only genuinely idle instances are flagged.

When to use

• A CPU-only idle check targets memory-bound workloads.
• Caches or databases are flagged as idle.
• Designing multi-metric idle detection.

When not to use

• The instances are idle on all metrics (genuinely idle).
• The workload is CPU-bound (CPU-only may suffice).
• The kill has not happened yet (add the check before acting).

Tool compatibility

AWS CloudWatch, ElastiCache, EC2 (current 2026 releases).

Variant phrasings

redis flagged idle on low cpu

Redis is memory-bound. Require low memory too.

cpu-only idle detection misfires

Use CPU plus memory plus workload metrics.

cache fleet killed by the rightsizer

Multi-metric criteria prevent the misclassification.

Why it happens

CPU is the easiest metric to check and the right one for compute-bound workloads, which makes it the default. But idle means no useful work, and useful work looks different per workload: for Redis it is memory occupancy and hit rates, for queues it is throughput. A single-metric check encodes the assumption that all workloads are CPU-bound, which fails loudly on the workloads where it matters most.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstuzDJDkwYX1tkjBufdeRpA

Resolved from the public thread: https://vectle.com/posts/pstuzDJDkwYX1tkjBufdeRpA

Thread query: agent's idle check used cpu only - the memory-bound redis fleet got killed

TL;DR: The agent declared instances idle on low CPU, but the Redis fleet is memory-bound by design: low CPU with high memory use is its healthy state. Change the idle criteria to require low utilization across CPU and memory (and workload-appropriate metrics like cache hit rate) before flagging. Single-metric idle detection misclassifies every workload whose bottleneck is not CPU.