AWS monitoring should tell you, fast, when customers cannot do what they came to do. Start with a few business and user-facing signals, set CloudWatch alarms that people trust, keep logs for a set period, and turn on a CloudTrail trail. Then name an owner for every alert.
This guide is for owners, operations leads and IT managers who run a customer portal, internal app or API on AWS, whether it was built in house or by a vendor. It covers what to watch, how to set alarms, which logs to keep, and what to ask whoever runs your stack.
What should AWS monitoring tell you?
Many teams turn on the default dashboards and stop there. Those graphs show CPU, memory and network traffic. They are useful, but they do not tell you whether an order went through or a customer form was saved.
AWS's own Well-Architected reliability guidance lists this as a common mistake: relying only on the metrics AWS services provide and skipping metrics specific to your workload. The same guidance notes that a business metric, such as orders processed per minute, can reveal a problem faster than CPU use.
So we frame monitoring around three questions:
- Can customers and staff complete the main tasks right now?
- Is anything about to run out, such as disk space, database connections or an API rate limit?
- Who changed what in the account, and when?
Which signals should you watch first?
Start small. We suggest five to eight signals for a first release, chosen from this list:
- Task success. Completed checkouts, submitted forms or processed jobs per minute, sent from your app as custom metrics.
- Error rate. The share of requests that return server errors at the load balancer or API gateway.
- Response time. A high percentile, such as p95, rather than the average, which hides slow requests.
- Saturation. Database connections, queue depth, disk space and memory on the busiest part of the system.
- Scheduled jobs. Did the nightly import or billing run finish, and when?
- Synthetic checks. A scripted login or test order every few minutes.
Synthetic checks matter more than most teams expect. AWS describes CloudWatch Synthetics canaries as scripts that follow the same steps a customer would, so you can "discover issues before your customers do." That works at 3 a.m., when there is no real traffic to trip an alarm. Keep these scripts short and use a test account, so they do not create real orders or put load on the system.
How should you set CloudWatch alarms that people trust?
An alarm nobody believes is worse than no alarm. People learn to ignore it, and then they ignore the real one. Three settings make the difference.
First, decide how many bad readings count as a problem. CloudWatch lets you set an alarm to fire when, for example, 2 out of 3 data points breach the threshold. This filters out one-off spikes without hiding a real outage.
Second, decide what missing data means. CloudWatch alarms have three states: OK, ALARM and INSUFFICIENT_DATA. You choose how each alarm treats missing data: as good (notBreaching), as bad (breaching), by keeping the current state (ignore), or with the default (missing), which moves the alarm to INSUFFICIENT_DATA when every point in the window is missing. For a heartbeat from a scheduled job, missing data usually means the job did not run, so treat it as breaching. For a metric that only reports when errors happen, treat missing data as good.
Third, split alerts into two groups. Pages wake someone up and need action now. Tickets can wait for business hours. If an alert does not lead to a clear action, remove it or move it to a dashboard.
What logs should you keep, and for how long?
Two kinds of records matter most: application logs and an audit trail of account changes.
For application logs, check retention. By default, CloudWatch Logs keeps log data indefinitely. That sounds safe, but storage costs grow, and you may hold personal data longer than your own policy allows. Set a retention period for each log group based on how long you need it for troubleshooting, contracts and any rules that apply to your industry.
For account changes, set up AWS CloudTrail properly. CloudTrail event history shows the past 90 days of management events in a Region. AWS states that for an ongoing record you should create a trail or an event data store. Without one, the record of who changed a security group or deleted a resource can be gone when you need it for an investigation.
Also decide what never goes into logs: passwords, full card numbers, API keys and, in most cases, health details. Setting that rule before launch is far easier than cleaning logs after.
Who should own the alerts?
Monitoring most often fails at the handoff, not in the tooling. Before go-live, write down:
- Who receives each page, and who is the backup.
- What the first responder should check, in a short runbook linked from the alert.
- When to call the vendor, and how fast they must answer.
- Who can approve a rollback or a restore.
If a vendor built your app, put these items in the support agreement. Ask them to show a real alert firing in a test environment and the runbook steps that follow it. Our guide on what to require for custom software testing covers asking for proof the same way, and our API integration team can connect alerts to the ticketing or chat tools your staff already use.
What does a first 30-day monitoring plan look like?
- Week 1. List the three to five tasks that matter most to customers and staff. Add a custom metric for each.
- Week 2. Set alarms for task success, error rate, response time and saturation. Choose missing-data handling for each one.
- Week 3. Add synthetic checks for login and the most important task. Set log retention and turn on a CloudTrail trail.
- Week 4. Run a drill. Break something on purpose in a test environment, confirm the right person gets paged, and time how long it takes to find the cause.
After that, review alerts once a month. Remove the ones that fired without needing action, and add one for any incident that customers noticed before you did.
Need help with AWS monitoring?
We build and run business applications on AWS with task-level metrics, tested alarms, set log retention and named owners, so problems reach the right person before customers notice.
Explore Application DevelopmentFrequently asked questions
CloudWatch collects metrics and logs and runs alarms, so it tells you how your app is behaving. CloudTrail records API activity in your AWS account, so it tells you who changed what and when. Most teams need both.
A handful. Five to eight alarms tied to customer tasks, errors, response time and capacity is a sound first set. Add more only when an incident shows a gap.
It depends on your contracts and the rules in your industry. CloudWatch Logs keeps log data indefinitely by default, so set a retention period for each log group on purpose instead of leaving the default.
Not always. CloudWatch covers metrics, logs, alarms and synthetic checks. A third-party tool can help when you run across several clouds or want one view of many services. Decide based on who will watch it, not on the feature list.

