Skip to the content.
View source on GitHub ↗

Amazon CloudWatch: The Operations Manager

The Business Goal

“Checkout is slow” is not yet an operational fact. Slow compared with what? In which Region? For every customer, or only one menu channel? Without a common view, each team measures a different restaurant.

The Story

The Operations Manager’s board has a few deliberate gauges: orders accepted, checkout latency, error rate, queue depth, Lambda duration and throttles, and DynamoDB capacity pressure. At 11:07, latency and Lambda duration rise together while order volume is ordinary. The board rules out a lunch rush and makes the kitchen station a better hypothesis.

She puts a bell beside the indicators that truly require action. A single late order does not ring it. Sustained p95 latency does.

Meet the AWS Service

CloudWatch metrics are time-ordered measurements. AWS services publish many metrics automatically; applications can publish custom metrics. A namespace groups related metrics, and dimensions distinguish useful slices such as function name, API stage, queue, table, or Region. A dashboard places selected metrics and alarms in one view. An alarm watches a metric or expression and can notify or trigger an approved action when it changes state.

How It Works

Choose signals that match a customer outcome and the dependency that could explain it.

Use dimensions carefully. A global average can hide one failing function version or API route. Publish custom metrics for business events only when they answer an operational question, such as OrdersCompleted or PaymentDeclines; do not make a unique dimension value for every order ID.

Alarms have states: OK, ALARM, and INSUFFICIENT_DATA. Design the missing-data behavior intentionally. An alarm can notify through Amazon SNS or initiate supported automated responses, but automatic action should be reversible and understood.

Architectural Mapping

Byte Burger board CloudWatch
Gauge metric
Label such as “drive-through” dimension
Wall display dashboard
Escalation bell alarm and action
“Too many ovens are occupied” service quota or saturation alarm

Painkiller

Alert on symptoms that need a person, not every measurement that exists. A dashboard helps a human understand; an alarm asks a human or system to act.

Knife Cut

CPU alone does not prove customer pain. A healthy-looking average also does not prove every customer is healthy. Pair infrastructure signals with request outcomes and percentiles.

The Masthead

CloudWatch makes Byte Burger observable from the floor up, but a gauge cannot explain the words inside a failed order.

A Note From the Author

See CloudWatch dashboards and CloudWatch alarms.

The Last Bite

The dashboard says the Lambda station slowed down at exactly the right time. Now the incident notebook must say what it was doing.

Next chapter: Amazon CloudWatch Logs: The Incident Notebook

Next: CloudWatch Logs becomes the record written during the shift.