Monitoring metrics

A reference sheet of the metrics that matter at each tier. Use it as a starting point — add the metrics that predict failure for this particular workload, and remove the ones nobody acts on.

Web tier

Elastic Load Balancer

Metric	What it tells you
RequestCount	Traffic volume; the denominator for every rate metric
HTTPCode_ELB_4XX_Count	Requests the load balancer rejected, often malformed or unauthorised
HTTPCode_ELB_5XX_Count	Load balancer errors; usually capacity or target health
HTTPCode_Target_5XX_Count	Errors returned by the application
TargetResponseTime	Time spent behind the load balancer — the user's experience
HealthyHostCount / UnhealthyHostCount	Target availability
ActiveConnectionCount	Connection pressure
RejectedConnectionCount	Load balancer at connection limit

API Gateway

Metric	What it tells you
Count	Requests handled
4XXError	Client-side failure — auth, validation, rate limiting
5XXError	Integration or authoriser failure
Latency	End-to-end request time including integration
IntegrationLatency	Time spent in the backend — latency minus this is gateway overhead
CacheHitCount / CacheMissCount	Cache effectiveness

CloudFront

Metric	What it tells you
Requests	Volume at the edge
BytesDownloaded / BytesUploaded	Transfer volume and cost driver
4xxErrorRate / 5xxErrorRate	Error rate at the edge
CacheHitRate	Origin load and cost; a low rate points at cache policy
OriginLatency	Time to origin when the cache misses

Application tier

EC2

Metric	What it tells you
CPUUtilization	Compute saturation; scales with most workloads
MemoryUtilization	Requires the CloudWatch agent; often the real constraint
DiskReadOps / DiskWriteOps	I/O pressure
NetworkIn / NetworkOut	Traffic and potential cost
StatusCheckFailed	Instance or host failure

Auto Scaling

Metric	What it tells you
GroupDesiredCapacity	Intended capacity
GroupInServiceInstances	Capacity actually available
GroupOutOfServiceInstances	Capacity lost to health failures
GroupPendingInstances	Capacity not yet ready — a scaling lag signal

Lambda

Metric	What it tells you
Invocations	Volume
Duration	Execution time; the cost driver alongside memory
Errors	Handler failures
Throttles	Concurrency limit reached — requests are being dropped
ConcurrentExecutions	Account and function pool pressure
IteratorAge	Consumer is falling behind a stream
DeadLetterErrors	Failed delivery to a DLQ

ElastiCache

Metric	What it tells you
CurrConnections	Connection pool health
CacheHits / CacheMisses	Cache effectiveness
Evictions	Memory pressure causing silent data loss
CPUUtilization	Node saturation
MemoryUsage / MemoryFree	Headroom
ReplicationLag	Read replica freshness

EKS

Metric	What it tells you
pod_cpu_utilization	Pod CPU relative to request
pod_memory_utilization	Pod memory relative to limit
pod_number_of_container_restarts	Crash loop or OOM kill
kube_pod_status_phase	Pending, Running, Failed
kube_deployment_status_replicas_available	Serving replicas
kube_deployment_status_replicas_unavailable	Deployment health
node_filesystem_avail_bytes	Disk pressure
kube_job_failed	Batch failure

Database tier

RDS

Metric	What it tells you
CPUUtilization	Instance saturation
FreeStorageSpace	Storage exhaustion risk
DatabaseConnections	Connection pool pressure
ReadIOPS / WriteIOPS	I/O load against provisioned capacity
ReadLatency / WriteLatency	Storage-level performance
DatabaseLoad	Active sessions — the most useful single RDS metric
ReplicaLag	Read replica staleness
FreeableMemory	Memory pressure
BurstBalance	Burst-capable volume exhaustion

DynamoDB

Metric	What it tells you
ConsumedReadCapacityUnits / ConsumedWriteCapacityUnits	Throughput used
ReadThrottleEvents / WriteThrottleEvents	Requests rejected for capacity
ThrottledRequests	Combined throttle signal
SuccessfulRequestLatency	User-facing latency
SystemErrors	Service-side failures

Aurora

Metric	What it tells you
AuroraReplicaLag	Replica freshness
BufferCacheHitRatio	Whether the working set fits in memory
DDLThroughput / DMLThroughput	Write load
RollbackSegmentHistoryListLength	Long-running transaction pressure

Cross-cutting signals

Signal	Source	Why it matters
Deployment events	CI/CD → CloudWatch annotations	Correlates change with effect
Queue depth and age	SQS	Consumers falling behind
DLQ depth	SQS	Silent message loss
Certificate expiry	ACM	Outage on a known date
Snapshot age	AWS Backup	Recovery point risk
Budget threshold	AWS Budgets	Cost anomaly

How to use this sheet

1. Select the metrics for the tiers the workload actually has.
2. Baseline for one week under normal load.
3. Set thresholds relative to the baseline and to user impact, not to generic defaults.
4. Alarm on the two or three that indicate a problem, with composite alarms to reduce noise.
5. Dashboard the rest for diagnosis, not alerting.
6. Review monthly — delete what nobody looks at.

Related standards

• Monitoring
• Workload health KPIs
• Runbooks