Availability Monitoring: Checking Whether a System Is Up and Usable
Learn how uptime is measured, how outside probes check a system, and how targets and alerts turn checks into action without false alarms.
In short
Availability monitoring is the habit of checking all the time whether a system is up and usable for the people who depend on it.
11 min readFree
What You Will Learn
A website can look fine from inside the company and still fail the people who pay for it. Availability monitoring is the habit of checking, all the time, whether a system is up and usable for its users. It also turns the question 'is it working?' into numbers. One common number is an uptime percentage, which is the share of time a system stays up. With numbers, a team can spot an outage quickly and compare the result with a goal.
By the end of this lesson you will be able to explain what availability means and how it is measured. You will also see how outside checks and health endpoints work, and how to decide when a failed check should become an alert.
Key words
- Availability
- The share of time a system is working and able to serve its users.
- Uptime
- The time a system is up and working during a set period.
- Downtime
- The time a system is down or cannot be used during a set period.
- Probe
- An automatic check that sends a request to a system and records the result.
- Health endpoint
- A special web address in an application that reports whether it is working.
- Synthetic monitoring
- Running scripted checks on a schedule, even when no real customer is online.
- SLO (service level objective)
- An internal target for a measure, such as 99.9% uptime over a month.
- Alert
- A message that tells people that something needs attention now.
- False alarm
- An alert that fires even though the system is actually fine.
What Does Availability Mean?
Availability is the share of time a system is up and able to do its job. Teams usually write it as a percentage. The basic formula divides uptime by the total of uptime plus downtime. Uptime is the time the system works. Downtime is the time it does not.
In words: take the time the system was working, divide it by the whole time, and multiply by 100. A system that works for 999 minutes out of every 1,000 minutes has 99.9% availability.
Measuring Uptime in Numbers
Teams often count the number of nines in a target. Two nines means 99%, three nines means 99.9%, and so on. Each extra nine cuts the allowed downtime to about one tenth of the year before. A 99.9% target allows about 8.76 hours of downtime in a year. A 99.99% target allows about 52.6 minutes. The code below does the same work for you, using 525,600 minutes in a 365-day year.
minutes_per_year = 365 * 24 * 60
for target in (99.9, 99.99):
allowed = minutes_per_year * (1 - target / 100)
print(f"{target}%: {allowed:.1f} minutes a year")Output
99.9%: 525.6 minutes a year 99.99%: 52.6 minutes a year
The two targets differ by a factor of ten in allowed downtime, not by a small amount. Also decide whether planned maintenance counts.
Server Responds vs Users Finish Tasks
A server that answers a network check is reachable. That is not the same as being usable. Three ideas explain the gap.
Reachable is not finished
A machine can reply to a network check while the checkout form throws an error. The monitor says up, but the customer cannot pay.
Tasks use many parts
Signing in may touch a web server, a database, a session store, and a sign-in service. If one part breaks, the task fails even though the front page loads.
Test the whole task
A useful check follows the path a real user takes and confirms the result. A check that stops at the home page measures very little.
How Availability Monitoring Works
Monitoring has two halves. One half sends checks at the system from outside. The other half decides what each result means. Both halves matter. Checks without a clear rule for 'down' produce noisy data, and a rule with no checks sees nothing at all.
What Are External Synthetic Probes?
Synthetic monitoring runs scripted checks on a schedule, from locations outside your own infrastructure, whether or not a real customer is online. The checks are called synthetic because they are made for testing, not caused by shoppers. Each single check is a probe. Probes matter most late at night and on weekends, when real traffic is low.
Ping check
Asks whether a machine answers at its network address. It shows that the machine is reachable, but it says nothing about the application running on it.
HTTP check
Requests a web address and reads the status code, such as 200 for success. Some checks also confirm that the page contains a required word.
Transaction check
Runs several steps in order, such as signing in and opening an order. It confirms that the task finished, not just that a page loaded.
Run probes from more than one location. Customers in different places reach a system through different network paths. A probe in one place can fail because of its own network connection, not because the service is down. Several locations help tell the two cases apart. Simple checks often run every one to five minutes, and heavier multi-step checks often run every five to fifteen minutes.
What Are Health-check Endpoints?
A health endpoint is a special address inside an application, often named /health. It reports the application's status to whoever asks. Microsoft's guidance describes two depths. A shallow check only proves that the application process can answer. A deep check also contacts the database and the other services the application needs to do useful work.
Shallow checks answer the question 'is the program alive?' Deep checks answer the question 'can it serve customers?' Deep checks cost something too. A probe that queries a database every minute uses connections that real shoppers may need. Keep these checks light, and when many probes run, consider caching the result for a short time.
Tip
A /health page that returns 200 as long as the program runs can stay green during a real outage. Make the endpoint use the same database and services that a real request uses, or point the monitor at a real endpoint instead.
Why Not Rely Only on Complaints
Internal signals, such as server load graphs and error logs, show how a system works inside. They do not prove that outside users can reach it. They can also stay quiet when a network route fails in front of the servers. User complaints arrive even later. A customer may wait hours before writing in, and many simply leave. Real user monitoring, which measures actual visits, is useful too, but it goes blind when few people visit. Scheduled probes keep watching in those quiet hours.
Acting on the Results
Measuring is only half the job. The other half is deciding what the numbers mean and what the team does next. Three decisions matter most: the target, the rule for down, and the rule for alerts.
How Do You Set Targets with SLOs?
Teams use three linked terms. They measure first, set a target second, and make a promise last. The steps below follow that order.
- 1
Pick the measure
Choose what you will count, such as the share of probes that pass. This measure is called an SLI, or service level indicator.
- 2
Pick the target
Set an SLO, such as 99.9% over a month. It is the goal the team works to meet every period.
- 3
Pick the window
Choose the period, such as a rolling 30 days or a calendar month. The same downtime gives different percentages over different windows.
- 4
Know the promise
An SLA, or service level agreement, is the promise made to customers, sometimes with a penalty.
Deciding What Counts as Down
Write the rule for 'down' before an outage happens. For example, a probe may count as failed when it gets no reply within a set timeout, when the status code is not a success, or when a required word is missing from the page. A timeout is the longest wait before a check counts as failed. Pick timeouts on purpose. A timeout that is too short marks slow but healthy responses as outages. A timeout that is too long hides real outages.
Then turn the rule into a number. Count the minutes the rule marks as down, divide by the total minutes in the window, and subtract the result from 100 to get the uptime percentage.
How Do You Alert Without False Alarms?
One failed probe is weak evidence. Networks drop packets, and servers sometimes stall for a moment. If every failure pages an engineer, the team learns to ignore the pages. A better rule asks for more proof before it raises an alarm.
Consecutive failures
Alert only after several checks fail one after another, such as three in a row at one-minute intervals.
Several regions
Alert when a majority of at least three locations fail at the same time. A single bad network route then cannot wake the whole team.
Critical paths first
Alert on failure of a key path, such as checkout, instead of on every small probe. Microsoft's guidance favors alerts on meaningful changes to critical flows.
The function below uses the rule of consecutive failures. In the list, True means the check passed and False means it failed. The function answers yes only when the last few checks all failed.
def is_down(results, needed):
recent = results[-needed:]
return len(recent) == needed and not any(recent)
results = [True, False, True, False, False, False]
print(is_down(results, 3))
print(is_down(results[:5], 3))Output
True False
Worked Example: An Online Store's Login and Checkout
Imagine an online store that sells shoes. Its owners want to know whether shoppers can sign in and pay. They set up three probes, one in each of three regions. Every minute, each probe requests the /health endpoint. Every minute, each probe also signs in with a test account. The test account is isolated, so the check cannot create a real order or change real customer data.
- 1
Set the down rule
The store counts a minute as down when the test login fails or times out in two or more regions during that minute. A good /health reply alone does not prove that shoppers can sign in.
- 2
Count the failed minutes
Suppose that in one 30-day month the rule marks 20 minutes as down. A 30-day month has 43,200 one-minute checks, so the team divides the failed minutes by that total.
- 3
Compare with the target
The team's target is 99.9% for the month. That target allows 43.2 minutes of downtime. Twenty minutes is inside that budget, so the target is met with 23.2 minutes to spare.
checks_in_month = 30 * 24 * 60
failed = 20
uptime = 100 * (checks_in_month - failed) / checks_in_month
allowed = checks_in_month * (1 - 99.9 / 100)
print(f"uptime {uptime:.2f}%")
print(f"allowed {allowed:.1f} minutes")
print(f"target met: {uptime >= 99.9}")Output
uptime 99.95% allowed 43.2 minutes target met: True
This example uses one-minute counts only. A real team also keeps the exact times, the regions involved, and the cause of each failure, so it can learn from every outage.
Quick check
A store's home page returns HTTP 200 every minute, yet customers say they cannot pay. Does availability monitoring show the site as up, and what should change?
Common Mistakes to Avoid
Trusting one green light
A single status code is not proof that customers can pay. Check the task that matters most to the business.
Probing from one place
One location can mislead. Use more than one region before declaring a full outage.
Changing real data in checks
A probe that places real orders or sends real emails can cause harm. Prefer read-only checks, and use isolated test accounts when a check must change something.
Promising what you cannot measure
A target helps only if the monitoring can show whether it was met. Very high nines allow so little downtime that the measurement itself becomes unreliable.
Key Takeaways
Availability monitoring asks, all the time, whether users can finish real work. It turns the answer into uptime percentages, compares them with SLO targets, and sends alerts only when the evidence is strong. Start with a few checks on the most important path, run them from more than one place, and write down the rule for 'down' before the next outage.
Tip
Start small. A few reliable checks on critical paths are more useful than a long list of checks that often cry wolf.
Quick check
A team wants 99.95% uptime over a 30-day month. About how many minutes of downtime does that allow?
Sources
- Microsoft Azure Architecture Center: Health Endpoint Monitoring pattern
- Dash0: What Is Availability Monitoring?
- Dash0: What Is Synthetic Monitoring?
- ClickHouse: What is synthetic monitoring?
- Loadster: Synthetic Monitoring: Tools and Techniques
- OpenStatus: What Is Synthetic Monitoring?
- OpenObserve: How Many Nines Do You Actually Need? A Practical Guide
- Project Helena: What is Uptime? SLAs, Nines & Downtime Tables
- Wikipedia: High availability
- Low Level Design Mastery: Availability Patterns
- Hyperping: Nine Nines (99.9999999%) Uptime: What It Really Means
- Microsoft Azure Architecture Center: Best Practices for Monitoring and Diagnostics
- Dotcom-Monitor: API Health Monitoring: Catch Silent Failures Fast
Last updated on 18 November 2026
Related lessons
- System designMonitoring: Watching a Running System So Problems Surface Early
- System designUsage Monitoring: Learning What Users Really Do With Your System
- System designAvailability Patterns: Fail-Over and Replication to Keep Systems Running
- System designAvailability: How Much Time a System Is Up, and Why Parts Add Up
Learn with me
Every lesson is free. I add new ones as I learn. Have a question, or want help with your own work? Book a short call.