CloudWatch Metrics

CloudWatch Metrics is a time-series data store for numerical measurements. Every AWS service emits metrics by default (basic monitoring), and you can emit custom metrics from your applications.

Core Concepts

Metric Structure

Namespace: AWS/EC2
MetricName: CPUUtilization
Dimensions: [InstanceId=i-xxxxx]
Value: 72.5
Unit: Percent
Timestamp: 2024-01-15T10:30:00Z
Period: 60 seconds

AWS Service Namespaces

Namespace	Common Metrics
AWS/EC2	CPUUtilization, NetworkIn, NetworkOut, DiskReadBytes
AWS/RDS	CPUUtilization, DatabaseConnections, FreeStorageSpace
AWS/Lambda	Invocations, Duration, Errors, Throttles
AWS/ALB	RequestCount, TargetResponseTime, UnHealthyHostCount
AWS/S3	BucketSizeBytes, NumberOfObjects, AllRequests
AWS/DynamoDB	ConsumedReadCapacityUnits, ConsumedWriteCapacityUnits

Basic vs Detailed Monitoring

	Basic (default)	Detailed (costs extra)
Resolution	5 minutes	1 minute
Cost	Free	$0.30/metric/month
Data retention	15 days	15 days
Use	Cost optimization, low-traffic	Production, real-time

Retrieving Metrics

GetMetricStatistics (single metric)

aws cloudwatch get-metric-statistics \
  --namespace AWS/EC2 \
  --metric-name CPUUtilization \
  --start-time 2024-01-15T00:00:00Z \
  --end-time 2024-01-15T12:00:00Z \
  --period 300 \
  --statistics Average,Maximum \
  --dimensions Name=InstanceId,Value=i-xxxxx

GetMetricData (multiple metrics, metric math)

aws cloudwatch get-metric-data \
  --metric-data-queries '[
    {
      "Id": "cpu",
      "MetricStat": {
        "Metric": {
          "Namespace": "AWS/EC2",
          "MetricName": "CPUUtilization",
          "Dimensions": [{"Name": "InstanceId", "Value": "i-xxxxx"}]
        },
        "Period": 300,
        "Stat": "Average"
      }
    },
    {
      "Id": "mem",
      "MetricStat": {
        "Metric": {
          "Namespace": "CWAgent",
          "MetricName": "mem_used_percent",
          "Dimensions": [{"Name": "InstanceId", "Value": "i-xxxxx"}]
        },
        "Period": 300,
        "Stat": "Average"
      }
    }
  ]' \
  --start-time 2024-01-15T00:00:00Z \
  --end-time 2024-01-15T12:00:00Z

Metric Math

[
  {
    "Id": "total_cost",
    "Expression": "m1 * 0.10 + m2 * 0.05",
    "Period": 300
  }
]

Common expressions:

• m1 + m2 — sum of two metrics
• SEARCH('{AWS/EC2}, Average)', 300) — search for metrics matching a pattern

Custom Metrics

Emit custom metrics from your application using the AWS SDK:

import boto3
 
cloudwatch = boto3.client('cloudwatch')
 
# Put a single data point
cloudwatch.put_metric_data(
    Namespace='MyApp',
    MetricData=[{
        'MetricName': 'OrderCount',
        'Value': 150,
        'Unit': 'Count',
        'Timestamp': datetime.utcnow(),
        'Dimensions': [
            {'Name': 'Service', 'Value': 'Checkout'},
            {'Name': 'Region', 'Value': 'us-east-1'}
        ]
    }]
)

StatsD Protocol (via CloudWatch Agent)

The CloudWatch Agent can receive StatsD metrics on UDP port 8125:

# Send a metric via netcat
echo "app.orders:150|c" | nc -u -w0 127.0.0.1 8125

{
  "metrics": {
    "namespace": "MyApp",
    "metrics_collection_interval": 60,
    "append_dimensions": {
      "InstanceId": "${aws:InstanceId}"
    },
    "statsd": {
      "service_address": ":8125",
      "metrics": [
        {
          "metric_name": "orders",
          "unit": "Count"
        }
      ]
    }
  }
}

High-Resolution Metrics

Standard metrics: 1-minute resolution. High-resolution: 1-second resolution (costs 10x more).

cloudwatch.put_metric_data(
    Namespace='MyApp',
    MetricData=[{
        'MetricName': 'Latency',
        'Value': 45.2,
        'Unit': 'Milliseconds',
        'Timestamp': datetime.utcnow(),
        'StorageResolution': 1  # 1-second resolution
    }]
)

GetMetricWidgetImage

Generate a PNG of a metric graph without using dashboards:

aws cloudwatch get-metric-widget-image \
  --metric-widget '{"metrics": [["AWS/EC2","CPUUtilization","InstanceId","i-xxxxx"]]}' \
  --output-format png

Integration with Other Services

Service	How It Uses Metrics
CloudWatch Alarms	Trigger alerts when thresholds are breached
CloudWatch Dashboards	Visualize metrics in real-time
CloudWatch Contributor Insights	Identify top contributors to a metric
Auto Scaling	Scale EC2/ECS based on metrics
EventBridge	Trigger rules based on metric math results

Limits

Resource	Limit
Metrics per call (PutMetricData)	20
Dimensions per metric	30
Metric name length	255 characters
Namespace name length	255 characters
GetMetricData metrics per call	500
GetMetricStatistics period minimum	60 seconds (1 second for high-res)

References

• Homepage: https://aws.amazon.com/cloudwatch/
• Documentation: https://docs.aws.amazon.com/cloudwatch/
• Pricing: https://aws.amazon.com/cloudwatch/pricing/

Pricing Examples

Scenario 1: An application with 50 custom metrics (order count, latency, error rate) emitted every minute. 50 metrics × 60 minutes × 24 hours × 30 days = 2.16M metric data points/month. At $0.30/million=$0.65/month. CloudWatch Logs ingestion for application logs (1GB/month) = $0.50/month.TotalCloudWatch:$1.15/month.

Scenario 2: A production system using detailed monitoring on 50 EC2 instances. 50 × $0.30/month=$15/month. Without detailed monitoring, you’d only see 5-minute averages — missing 1-minute spikes in CPU. At $0.02/vCPU-hour(spotinstance),a4-vCPUinstancerunningat100$0.0013. Detailed monitoring costs $0.30/instance/month. The cost is justified by the visibility.

Nuggets & Gotchas

• PutMetricData has a 20-metric limit per call — use batching for high-volume: If you emit 100 custom metrics per minute, batch them into 5 calls of 20 metrics each. For ultra-high-volume applications (10K+ metrics/minute), use the metric math API or the CloudWatch Agent’s batch feature.
• Basic monitoring (5-min) is free but has a 15-day retention — you can’t query older data: If you need historical data beyond 15 days, you must either use detailed monitoring (1-min resolution, 15-day retention) or stream metrics to S3 via Kinesis for long-term storage.
• Metric math expressions are evaluated independently of source metrics’ retention periods: If you use SEARCH() to aggregate metrics, the resulting metric math metric inherits the minimum retention period of the source metrics. For 5-minute basic monitoring, the math result is also 5-minute.
• Dimensions are case-sensitive — “InstanceId” not “instanceid”: If you query a metric with dimensions and get no results, check the dimension name casing. The CloudWatch API treats InstanceId and instanceId as different dimensions.
• **EC2 basic monitoring (5-min) is free but detailed monitoring (1-min) costs $0.30/instance/month:\ast \ast Beforeenablingdetailedmonitoringonallinstances,considerthatfora100-instancefleet,tha{t}^{\prime }s$30/month. For production, detailed monitoring is worth it. For dev/test, basic monitoring is sufficient.