Learners moving through a DevOps curriculum eventually reach a fork: tooling for delivery on one branch, and the discipline of keeping production dependable on the other. SRE, short for site reliability engineering, sits on that second branch, and interviewers increasingly test candidates on it. The nine ideas below are the ones worth knowing well before you apply for a reliability-flavoured role. Facts and numbers come from the Ops Atlas entry on SRE, so you can read the full context there when you want to go deeper.
1. Reliability is measured from the user’s side
An indicator is the count of good events divided by the count of valid events, taken as close to the user as practical. Candidates often stumble here by quoting CPU load or server uptime. The Atlas separates that sort of telemetry from the user’s perception and notes that the four golden signals, helpful as they are, do not equal customer happiness. In an interview, describe a user-facing journey and how you would count its successes.
2. Speed is a proportion, not an average
Report latency as a fraction: the requests completed inside a threshold you pick. An average hides the slow tail where unhappy users live. A related detail that impresses examiners: a quickly returned error counts as a failure in availability, yet it is excluded from the latency figure so that it cannot flatter the speed score.
3. The right target is never 100 percent
The reasoning is economic. The source says every additional increment may cost about a hundred times the previous one, and users usually cannot tell the difference. Learn the typical alternatives and be ready to justify one. The entry also supplies the “extra nine” rule: a dependency should be built to a target one nine higher than the service that relies on it.
4. The error budget turns targets into decisions
Subtract the objective from 100 percent and you have the allowance for failure. Practise the conversion: a 99.9 percent goal across a 28-day window allows about 40 minutes of bad service. For request counts, 50 million valid requests with a 99.9 percent goal give 50,000 allowable failures. Be able to do this on paper, because it is a classic whiteboard question.
5. A policy gives the budget teeth
A number on a wiki is not a practice. The Atlas gives a sample rule: if the trailing four-week allowance is exhausted, stop every release other than urgent and security-related ones. It also lists “no signed policy” among the anti-patterns. When asked how to introduce SRE to a sceptical organisation, say that you would draft a one-page policy and get a senior stakeholder to sign it before building anything else.
6. Alerts follow burn rate
The alternative to pages for every metric is alerting on how fast the budget is draining. For a 99.9 percent objective, the workbook pairs a one-hour window with a five-minute check at a burn of 14.4 for pages, and a six-hour window at a burn of 6, while tickets come from slower rates of 3 and 1. You do not need to memorise every figure; you need to explain why short and long windows are combined and why slow burns do not wake people. This is also a good moment to read our guide on the DevOps vocabulary learners should commit to memory, which complements this list.
7. Toil is a measurable enemy
Toil is manual, repetitive, automatable, tactical work with no lasting value that scales with the service. Show that you can quantify it: if six engineers each work 80 hours and 190 hours are spent on chores, the share is just under 40 percent. The Atlas references a cap of under half of engineer time and reports survey medians near a third. Strong candidates also mention that a cap needs consequences, for example routing tickets back to the product group or putting developers on call again.
8. Postmortems are about conditions and actions
Blameless does not equal unaccountable. Good reviews examine the circumstances that made an error easy while remaining strict about action items, owners and due dates. The entry lists postmortems that never close their actions as an anti-pattern, and offers a trigger: any single event that burns more than a fifth of the budget deserves a formal review with a top-priority action attached. If you can cite that sort of rule, you sound like someone who has run the process.
9. Team shape and on-call design matter
Finally, know the organisational side. The Atlas sketches six ways to organise a team, among them the all-purpose “kitchen sink” group, an infrastructure group, a tools group, and product-aligned, embedded and consulting variants. Each has typical risks. It also gives staffing heuristics, including a rota of around eight people for one site covered continuously, and no more than a couple of incidents landing on any half-day shift. Link these to the career question of whether you want to work on shared platforms or alongside a product group. For a view of how SRE fits with other infrastructure disciplines in large companies, see our article on combining cloud, DevOps and reliability engineering in enterprise IT.
Bonus: the reading list
When recruiters ask what you have studied, a short, accurate list helps. The Atlas names the original SRE book from 2016, which is free to read online, the companion workbook from 2018, a 2017 paper in a computing magazine on the calculus of service availability, and a freely licensed guide to setting objectives. It also mentions an open YAML specification for service targets. Reading even two of these closely will set you apart from candidates who only skimmed blog posts.
A final tip on delivery: practise saying each concept in two sentences, then in ten. Interviewers often start with the short version to test clarity and follow with a probing question on the long version. Candidates who have rehearsed both levels sound calm, and calm is a reliability trait in its own right.
How to practise
- Choose any small application you control and write one indicator for it.
- Compute the budget for a 28-day window and keep it in a spreadsheet.
- Draft a one-page policy and share it with a mentor for comments.
- Log your repetitive tasks for a fortnight and automate the largest one.
- Write a two-page review of a past mistake, with owners and dates for each action.
Employers often ask whether a certificate is enough. Our discussion of what certifications are actually worth in technology careers argues that credentials help most when paired with evidence like the project above.
Closing advice
None of these nine ideas requires special software. They require clear thinking, honest arithmetic and the habit of writing things down. Master them in order, rehearse the explanations aloud, and the reliability questions in your next interview will feel familiar rather than daunting.

