Service Level Objectives: How to Set Ones That Work

Service level objectives explained for engineers: how to pick SLIs, set targets, calculate error budgets, tell SLO from SLA, and avoid paging for noise.

A service level objective (SLO) is a reliability target for one measurable part of your service, over a fixed window of time. “99.9% of requests to the checkout API succeed over 30 days” is an SLO. It tells your team how reliable is reliable enough, and what to do when you fall short.

This guide covers the vocabulary, a method for choosing targets, worked examples with the downtime math, the difference between an SLO and an SLA, and the mistakes that make SLOs useless. It is written for small teams and solo engineers who carry the pager and don’t have a dedicated reliability group.

Uptime Beacon
Make reliability signals more trustworthy
Uptime Beacon confirms failures from a second location on paid plans, helping on-call teams avoid treating network blips as downtime.
Explore Uptime Beacon

TL;DR

  • An SLO is a target for an SLI (a measurement such as success rate or latency) over a time window. An SLA is a contract with consequences. Your SLO should be stricter than any SLA you sign.
  • Never target 100%. The gap between 100% and your SLO is your error budget, and it is what lets you ship changes.
  • Start with one or two SLOs per user-facing service, measured from the outside, and revise them after a quarter of real data.
  • Bad measurement ruins good targets. A false alarm counts as downtime that never happened, so confirm failures before they count.

Table of Contents

What a service level objective is

The vocabulary comes from Google’s SRE book, which defines an SLO as a target value or range of values for a service level, measured by a service level indicator. Three terms fit together:

  • SLI (indicator): a quantitative measurement of how the service behaves. Common ones are availability (the share of requests that don’t return errors) and latency (the share of requests completed within a set time).
  • SLO (objective): the target for that indicator, for example 99.9% over 30 days.
  • SLA (agreement): the contract with customers that says what happens, often a credit or refund, if you miss a target.

IBM describes an SLO as an agreed-upon performance target for a service over a period of time, one that helps stakeholders manage the health of specific services. The “agreed-upon” part matters. An SLO that only the on-call engineer knows about doesn’t change any decision. One that product, engineering and support all accept does.

SLO vs SLA: the difference that matters

People mix these up because the numbers look the same. The difference is consequence.

SLO SLA
Audience Your own team, sometimes shared with customers Customers, legal, finance
Nature Internal goal Contract
If missed You slow releases and fix reliability Credits, refunds or penalties
Typical strictness Tighter Looser, with margin

SLOs usually act as the internal goals that keep external SLA commitments safe, as the SRE book describes. If you promise customers 99.5%, aim internally for 99.9%. The gap is your warning zone: you can miss the SLO, learn from it and fix things, without paying anyone.

If you sell to businesses and you have never written an SLA, you still need an SLO. Customers feel your reliability whether or not a contract mentions it.

Choose SLIs your users would recognize

The usual failure is picking what is easy to measure instead of what users feel. CPU utilization is easy. Nobody files a ticket because CPU was at 80%. They file one because the page didn’t load.

For a typical web product or API, start with these:

  1. Availability: the proportion of requests (or checks) that succeed. For an HTTP endpoint, define success precisely: status code, response within a timeout, and ideally a content check.
  2. Latency: the proportion of requests faster than a threshold, such as 95% under 400 ms. Use a percentage of requests, not an average, because averages hide the slow tail.
  3. Freshness or completion for background work: did the nightly job run, and did it finish in time?

Two habits keep this honest. First, measure from outside your own network, because an internal health check can be green while users in another region see errors. Second, write the SLI in one sentence a new hire could implement: “A check is good if the endpoint returns 200 within 5 seconds and the body contains the expected field.”

Set the target: the math of nines

Every extra nine costs far more than the last one. The table uses a 30-day month (43,200 minutes), which is the window most teams use for rolling SLOs.

SLO Error budget Allowed downtime in 30 days
99% 1% about 7 hours 12 minutes
99.5% 0.5% about 3 hours 36 minutes
99.9% 0.1% about 43 minutes
99.95% 0.05% about 21.6 minutes
99.99% 0.01% about 4.3 minutes

The 99.9% row matches published guidance: an SLO of 99.9% leaves a 0.1% error budget, roughly 43 minutes of downtime a month. The other rows are the same multiplication: budget percentage times 43,200 minutes.

Read the bottom row carefully. At 99.99% you have about four minutes a month. A single slow deploy, a DNS hiccup or a person taking five minutes to wake up and open a laptop spends the whole budget. Most small teams don’t need that, and can’t staff for it.

A sensible starting point for a small SaaS is 99.9% for the main user-facing path and something looser, such as 99.5%, for secondary features like an admin panel. Pick the number your users would actually notice, then check whether your team can sustain it.

Set the target: the math of nines — overview diagram

Error budgets: the part that changes behavior

An SLO is only a number until you use the gap. The error budget is the difference between 100% and your SLO. Google’s workbook explains that SLOs are about balancing reliability and innovation: because you accept that the service won’t be perfect, the remaining budget becomes an objective guide for when to pause feature releases and focus on reliability.

A simple policy that fits a five-person team:

  • Budget healthy (more than half left): ship normally.
  • Budget below a quarter: only low-risk changes, and every incident gets a short written review.
  • Budget spent: freeze risky releases until the rolling window recovers, and put reliability work at the top of the sprint.

Write the policy down before you need it. The conversation “should we freeze?” is easy when the rule exists and painful when it doesn’t.

Three worked SLO examples

1. Public marketing site and login page

  • SLI: share of HTTP checks returning 200 within 5 seconds, run every minute from multiple regions.
  • SLO: 99.9% over a rolling 30 days.
  • Budget: about 43 minutes. One 20-minute outage uses nearly half.

2. Public API

  • Availability SLI: share of requests not returning 5xx errors.
  • Latency SLI: share of requests completed in under 500 ms.
  • SLOs: 99.9% availability, and 95% of requests under 500 ms, both over 30 days.
  • Why two: an API that answers correctly in 12 seconds is effectively down for the client calling it.

3. Nightly billing cron job

  • SLI: share of scheduled runs that report success within 30 minutes of their expected time.
  • SLO: 99% of runs over 90 days. With one run a day, that tolerates roughly one miss per quarter.
  • Why this window: with daily events, a 30-day window gives 30 data points and makes every single miss look catastrophic.

For more patterns, including queue processing, freshness and database targets, the SLO and SLI examples collection from Oreon walks through twelve with sample targets.

Measure it without lying to yourself

Here is the part most SLO guides skip. Your SLO is only as good as the data under it. Two failure modes are common.

False failures inflate your downtime. If a single failed check from one location counts as an outage, a brief network blip between the monitor and your server spends error budget you didn’t lose. Your numbers look worse than the real user experience, and you start freezing releases for nothing.

Alert noise destroys the response. The same blip that spends budget also pages someone at 3 a.m. After a few of those, people mute the channel, and the real incident gets the same shrug. We go deeper on this in our piece on SaaS uptime monitoring that pages only for real failures.

The fix is to require evidence before an event counts. Several failures in a row, ideally confirmed from a second location, separate a real outage from noise. Then decide, in writing, what counts as downtime for SLO purposes: confirmed incidents only, or any failed check? Both are defensible. Changing the definition halfway through a quarter is not.

Failure checks lead to confirmation and downtime accounting

One more thing to measure is how fast you recover. Your SLO tells you how much downtime you can afford. Recovery time tells you how many incidents that buys. Our write-up on mean time to recovery shows why one outage can produce three different MTTR figures, depending on where you start and stop the clock.

Our perspective: fewer alerts, better objectives

We build monitoring for engineers who carry the pager, and the pattern we see in the field is consistent: teams set a reasonable SLO and then wreck it with noisy measurement. So our design choice is deliberate. An incident in Uptime Beacon opens only after several failures in a row, and on paid plans the failure is confirmed from a second location before anyone is alerted. We check from five regions: EU Central, EU North, US East, US West and Asia Pacific.

Our founders have more than 30 years of combined experience running production systems in banking, fintech, telecom and SaaS, and the lesson from that background is simple. Reliability targets are decisions about people as much as machines. An SLO that wakes someone for nothing teaches them to ignore the pager. An SLO that quietly burns budget on phantom failures teaches the business to distrust the numbers.

Our advice, for what it is worth:

  • Start with one SLO per critical user path. Add more only when a real incident shows a gap.
  • Choose a target slightly easier than you wish, hit it for a quarter, then tighten it.
  • Alert on confirmed user impact, not on every wobble.
  • Review the target after every serious incident. If it was too loose, say so. If it was too strict, say that too.

We also think the check history behind an SLO should be easy to question. That is why every plan, including Free, can connect an AI assistant over MCP to read monitors, incidents and check history, and to diagnose a failing monitor region by region. Ask “which regions failed during last Tuesday’s incident, and for how long?” and get an answer from the data, rather than from a spreadsheet you build by hand at the end of the quarter.

— Aaditya Parashar

Keep SLOs alive: a quarterly routine

The SLO Development Life Cycle handbook treats SLOs as living artifacts: research them, implement them, then adjust based on experience with real service performance. A lightweight version for a small team:

  1. Week 1 of the quarter: review the last 90 days. Did you meet each SLO? How much budget did each incident use?
  2. Check relevance: did any customer complaint happen while the SLO said everything was fine? If so, the SLI is measuring the wrong thing.
  3. Adjust one thing: tighten a target you beat easily, loosen one you can’t sustain, or add an SLI for a gap.
  4. Share it: a one-page summary for the whole company. Reliability is a product decision, not only an engineering one.

Alex Hidalgo’s The SLO Book is the deeper reference if you want to go beyond this routine, since much SLO advice stops before the implementation details.

Common mistakes

  • Targeting 100%. It is unrealistic, and Google’s guidance names it as the most common pitfall because it stifles releases. Without an error budget, every change looks like a threat.
  • Too many SLOs. Twenty targets means none get attention. Three that people recall in standup beat twenty in a wiki.
  • Measuring only from inside. Your load balancer’s view is not your customer’s view.
  • Ignoring non-HTTP things. Expired certificates, lapsed domains and silent cron jobs are outages your users feel. Watching them is cheap: see our guides on certificate expiry monitoring, domain expiry monitoring and cron job monitoring.
  • Copying another company’s numbers. A target that suits a bank’s payments rail may be absurd for your internal tool.

Try it on your own services

If you want real data under your first SLO, you can start without a card. The Free plan gives you 5 monitors checking every 5 minutes from EU Central, one alert channel (email or Slack) and one status page, plus MCP access for your AI assistant. Paid plans add second-location confirmation of failures, and Starter and above include maintenance windows, so planned work doesn’t count against your budget. Starter costs €14.99 per month, Pro €39.99 and Business €89.99. Extra monitors and expiry watches are available as add-ons, and Enterprise pricing is not published.

Compare the plans on the Uptime Beacon homepage, or see how we handle API and cron monitoring for SaaS teams. Every result is stored on a control plane and database in the EU, which helps if your customers ask where monitoring data lives.

FAQ

What is a good SLO for a small SaaS?

For the main user-facing path, 99.9% over a rolling 30 days is a common starting point, which allows about 43 minutes of downtime a month. Secondary features can run looser, such as 99.5%. Start there, hit it for a quarter, then adjust based on what your customers actually notice.

Is an SLO the same as an SLA?

No. An SLO is an internal target that guides engineering decisions. An SLA is a contract with customers that carries consequences, often financial, if you miss it. Set your SLO stricter than your SLA so you get a warning before you owe anyone anything.

How many SLOs should I have?

Fewer than you think. One or two per critical user journey, such as availability and latency for your API, is enough to begin. Add another only when an incident reveals something your current SLOs can’t see.

What happens when the error budget runs out?

You follow the policy your team wrote in advance. Typically that means pausing risky releases and prioritizing reliability work until the rolling window recovers. Google’s SRE guidance uses the budget as an objective signal for exactly this decision, so it doesn’t become an argument.

Can I measure SLOs with an uptime monitor?

Yes, for availability and response-time SLIs on endpoints, ports and hosts, plus heartbeats for scheduled jobs. The key is to count only confirmed failures, so a network blip doesn’t spend your budget. Define what counts before the quarter starts, and keep the definition fixed.

Sources

Put your site on it. Sleep through the blips.

Add your first monitor now. If your site is fine, you will not hear from us.

Monitor it free
Sign up with GoogleOpening Google…
Free plan, no cardSign up with Google or email