Logs, Metrics and Alerts for Business Applications: Knowing Before Your Users Tell You
November 11, 2026 · 7 min read · Etzal Group Team
At a logistics company, the nightly job that sends delivery manifests to the warehouse fails on a Tuesday. Nothing announces it. On Wednesday morning the warehouse has no list and calls the office. The office calls IT. IT calls the vendor, who spends the afternoon working out what happened, because the application wrote nothing useful when it failed. A day is lost to a problem that began with one failed step at 01:00. Good logs, metrics and alerts would have put a message on someone's phone at 01:02 saying exactly which job failed and why.
Three things to have
- Logs: a record of what happened.
- Metrics: numbers showing how the system is behaving over time.
- Alerts: messages to a person when something needs attention.
Together they answer three questions: is it working, how well, and if not, why?
Logs: what to record
A log is useful when someone who was not there can understand what happened.
Record:
- Requests and their outcome. Which function, how long it took, success or failure.
- Errors, with context. What was being attempted, for which record, and the technical detail.
- Business events. Order created, invoice approved, payment received, export completed.
- Background jobs. Started, finished, how many records, how long, or why it failed.
- Calls to other systems. What was called, the response, the time taken.
- Security events. Sign-ins, failed attempts, permission changes, exports of sensitive data.
Each entry should carry:
- A precise timestamp, in one agreed time zone.
- A severity level: error, warning, information.
- An identifier that links all entries for the same request or transaction, so one user action can be followed across components.
- Who and what: the user and the record concerned, by identifier.
Never record:
- Passwords, tokens or keys.
- Full card numbers or identity numbers.
- More personal data than needed. Logs are read by more people, kept longer and protected less carefully than databases. Record identifiers, not contents.
Structure them. Entries written as named fields, not free sentences, can be searched and counted. "Find every failed payment for customer 4821 last week" should take seconds.
Collect them in one place. Logs left on individual servers are lost when a server is replaced, and cannot be searched together.
Decide how long to keep them. Long enough for investigation and any legal requirement, not forever.
Metrics: four to start with
For any application that serves users, four measurements describe its health.
- Traffic. How much it is being used: requests per minute, active users.
- Errors. The share of requests that fail.
- Speed. How long requests take, at the median and at the 95th percentile. Averages hide the slow cases users complain about.
- Saturation. How close resources are to their limits: processor, memory, disk, database connections, queue length.
Add, for background work:
- Job health. Did each scheduled job run, how long did it take, how many records did it process?
And for dependencies:
- External services. Success rate and response time of each system you call.
Keep history. The value of a metric is in comparison: with yesterday, with last month, with before the last release.
Business-level checks
Technical health is not business health. A system can respond quickly, with no errors, and still be doing nothing useful.
Measure what the business expects to happen:
- Orders created per hour. If it normally sees forty between 09:00 and 10:00 and today there are none, something is wrong, even with every server healthy.
- Messages sent, invoices issued, payments matched.
- Records waiting in a queue, and the age of the oldest.
- Time since the last successful sync with another system.
- Totals that must agree: orders in one system against orders received in another.
These catch the failures technical monitoring misses: a job that "succeeds" having processed zero records, an integration that returns empty results, a form that silently stopped submitting.
Alerts: the first five
An alert is a demand on someone's attention. Each one must require action.
Start with these.
- The application is unreachable. An external check, from outside your network, every minute.
- Error rate above a threshold for several minutes.
- Key operations are slow for several minutes.
- A scheduled job failed or did not run. Alert on absence: "the manifest job has not completed by 01:30" catches both failure and never starting.
- A critical business flow has stopped. No orders in an hour during business hours.
Then add infrastructure warnings: disk above 80 percent, certificates expiring within three weeks, backups not completed, queue growing.
Make each alert useful:
- What is wrong, in plain words.
- Where.
- Since when and how bad.
- What to do first, or a link to the procedure.
Send it to a person, not to a mailbox nobody reads. Decide who receives alerts in working hours and out of them, and what happens if they do not respond.
Prune. An alert that fires and needs no action should be changed or removed. Teams that receive fifty alerts a day read none of them.
A dashboard for people who are not engineers
One screen that a manager can read:
- Is the system up?
- Orders, invoices or jobs processed today against a normal day.
- Errors today.
- Background jobs: last run and result.
- Integrations: last successful sync.
- Open incidents.
If the person responsible for operations opens it every morning, problems are noticed that no alert was written for.
When something goes wrong
With logs, metrics and alerts in place, an incident follows a path:
- An alert says what is wrong.
- The metrics show when it started and what else changed at that moment: a release, a traffic spike, a dependency slowing.
- The logs, filtered by time and by the linking identifier, show the failing step and the error.
- The fix, and afterwards a new alert or check if this failure was not caught early enough.
Without them, step 2 is guesswork and step 3 is "can you try again while I watch?"
What to ask a vendor
When commissioning a system, put these in the requirements.
- Structured logs for requests, errors, business events, jobs and integrations, with a linking identifier.
- No secrets or unnecessary personal data in logs.
- Central log storage with search, and a stated retention period.
- The four core metrics, plus job and integration health.
- An external availability check.
- An agreed list of alerts, with recipients.
- A health dashboard for the client.
- A runbook: for each alert, what it means and what to do.
- An audit log of security-relevant actions, separate from technical logs.
Ask to see these in the demonstration. A system that can show you last night's jobs and yesterday's error rate is a system someone can operate.
What it costs
- Build effort. Small when designed in from the start. Significant when added to a finished system.
- Tooling. Log and metric services charge by volume. Log sensibly, sample high-volume routine events, and set retention deliberately.
- Attention. Someone must own the alerts and look at the dashboard.
Against that: the day lost at the logistics company, and every similar day.
Common mistakes
- Logging nothing useful, or everything.
- Free-text logs that cannot be searched.
- Secrets and personal data in logs.
- Logs kept only on the server.
- Watching servers and not the business flow.
- Alerts on causes nobody acts on.
- Alerts to a shared mailbox.
- No alert when a job fails to start.
- Averages instead of percentiles.
- Adding monitoring after the first serious incident.
We build logging, metrics, alerting and a client health dashboard into the systems we deliver, as part of Enterprise Software. If you learn about outages from your users, contact us.
Summary
Record structured logs of requests, errors, business events, jobs and integrations, linked by an identifier and free of secrets. Track traffic, errors, speed and saturation, plus job and integration health, and keep history. Add business-level checks such as orders per hour. Begin with five alerts on what users feel, send them to a named person and prune those that need no action. Specify all of it when commissioning a system.
Frequently asked questions
What is the difference between logs and metrics?
A log is a record of an individual event, such as a user signing in or an invoice failing to save, with details. A metric is a number measured over time, such as requests per minute or the share that failed. Metrics tell you that something is wrong and when. Logs help you find out what and why.
What should a business application alert on first?
Start with what users experience: the application is unreachable, the error rate is high, key pages are slow, a scheduled job has not run, and a critical business flow such as order creation has stopped. Add infrastructure alerts such as a full disk next. Alert on symptoms people feel before causes that might not matter.
Do we need a dedicated monitoring team?
Not for a mid-sized business system. You need the application built to produce useful logs and measurements, a small set of well-chosen alerts routed to a named person, and a simple routine for reviewing them. Much of this can be specified in the contract with whoever builds or runs the system.
Related service
Website Development
Fast, SEO-ready websites that turn visitors into customers.
