Site Reliability Engineer Interview Questions (2027)
44 site reliability engineer interview questions organised by round: coding, Linux internals, live troubleshooting, reliability design and behaviour.
Site reliability engineer interviews usually test five areas in separate rounds: coding or scripting, Linux and networking internals, troubleshooting a live outage, system design for reliability, and behaviour around incidents and on-call. This page has 44 questions organised by round, with short answers and what the interviewer is checking in each.
Round structure varies by company, so check the hiring team's own description of its loop. The SLO, error budget and toil definitions below follow Google's public Site Reliability Engineering book and SRE Workbook, checked in November 2026.
Key takeaways
- SRE loops mix software and operations. Expect at least one coding or scripting round even for operations-heavy roles.
- Linux internals questions go deeper than for DevOps roles: processes, memory, file descriptors, TCP states and what happens when you load a URL.
- Troubleshooting rounds reward a method: scope the impact, check recent changes, form a hypothesis, test it, mitigate before you root-cause.
- Reliability design rounds ask you to put numbers on a system: requests per second, storage, replicas and failure domains.
- Behavioural rounds look for blameless postmortems, pushing back on toil and handling an error budget conversation with product owners.
Which SRE guide to use
This site has several SRE question banks with different jobs. This page is arranged by interview round so you can prepare for each stage of a loop.
| Guide | Best for |
|---|---|
| This page | Preparing round by round: coding, Linux, troubleshooting, design, behaviour |
| 100+ SRE interview questions | Broad coverage of SLOs, monitoring and incidents by level |
| Scenario-based SRE questions | Incident walkthroughs |
| Advanced SRE questions | Senior candidates: capacity, failure design, leadership |
New to the role itself? Read what SREs do in DevOps teams first.
Round 1: recruiter and SRE fundamentals
1. What does a site reliability engineer do?
An SRE keeps a service reliable by applying software engineering to operations work. That means defining SLOs, building monitoring and alerting, automating toil, running incidents and postmortems, and working with developers on how releases affect reliability.
2. How is SRE different from DevOps?
DevOps is a set of practices and culture for delivering software. Google describes SRE as one concrete way to implement DevOps, with specific tools such as SLOs, error budgets and caps on operational work. In many companies the job titles overlap.
3. Define SLI, SLO and SLA.
An SLI is a measured indicator, such as the share of requests served in under 300 ms. An SLO is the target for that indicator over a window, such as 99.9% over 30 days. An SLA is a contract with customers that has consequences if missed, and its target should be looser than the internal SLO.
4. What is an error budget?
It is the amount of unreliability the SLO allows: 100% minus the SLO. A 99.9% monthly SLO gives about 43 minutes of full downtime in a 30-day month. While budget remains, teams can ship faster; when it is spent, the agreed policy usually slows releases and puts reliability work first.
5. What is toil?
The SRE book defines toil as work that is manual, repetitive, automatable, tactical, without lasting value, and that grows with the service. Google aims to keep each SRE's toil below 50% of their time.
6. What are the four golden signals?
Latency, traffic, errors and saturation. The SRE book suggests that if you can measure only four things about a user-facing system, measure these.
Round 2: coding and scripting
Interviewers check that you can write working code for operational problems, read it aloud and handle edge cases. Python, Go or Bash are usually accepted.
7. Write a script that finds the 10 IP addresses with the most requests in an access log.
In Bash: awk '{print $1}' access.log | sort | uniq -c | sort -rn | head -10. In Python, read lines, split on spaces, count with collections.Counter and call most_common(10). Mention memory: for huge files, stream the file instead of loading it.
from collections import Counter
counts = Counter()
with open("access.log") as f:
for line in f:
counts[line.split(" ", 1)[0]] += 1
for ip, n in counts.most_common(10):
print(n, ip)
8. Parse log lines and compute the 99th percentile latency per endpoint.
Group latencies by endpoint in a dictionary of lists, sort each list and pick the value at index ceil(0.99 * n) - 1. Say that for streaming data you would use a histogram or a sketch such as t-digest rather than store every value.
9. Write a function that retries an HTTP call with exponential backoff.
Loop up to a maximum number of attempts, sleep for base * 2 ** attempt plus random jitter, and retry only on errors that are safe to retry, such as timeouts and 503. Stop retrying on 4xx client errors. Jitter avoids many clients retrying at the same moment.
10. Implement a rate limiter.
A token bucket works well: tokens refill at a fixed rate up to a capacity, and each request takes one token or is rejected. Store the token count and last refill time, and refill lazily when a request arrives. Discuss how you would share the state across instances, for example in Redis.
11. How would you check that 500 servers have the same config file?
Compute a hash such as sha256sum on each host through your configuration tool or SSH in parallel, then group hosts by hash and report outliers. Better still, manage the file with Ansible or another tool so drift is corrected automatically.
12. Write a script that alerts when disk usage on a host goes above 90%.
Read usage with shutil.disk_usage in Python or parse df -P, compare against the threshold and exit non-zero or send a message. In an interview, add that production alerting should come from your monitoring system, and the script is a fallback.
Round 3: Linux and networking internals
13. What happens when you type a URL into a browser and press Enter?
The browser checks its cache, resolves the name through DNS, opens a TCP connection (or QUIC for HTTP/3), performs a TLS handshake, sends the HTTP request, and the server responds through any load balancers and proxies. The browser then parses HTML and fetches more resources. Interviewers expect you to go deep on whichever layer they pick.
14. A process is using 100% CPU. How do you find out what it is doing?
Use top or pidstat to confirm, then perf top -p <pid> to see hot functions, or strace -p <pid> to see system calls. For a Java or Python service, take a thread dump or use a language profiler.
15. What does load average mean on Linux?
It is the average number of tasks running or waiting to run, including tasks in uninterruptible sleep (often waiting on disk I/O), over 1, 5 and 15 minutes. Compare it with the number of CPUs, and check I/O wait before you blame the CPU.
16. The disk is full according to df, but du shows much less used. Why?
A process still holds a deleted file open, so its space is not freed. Find it with lsof +L1 and restart the process or truncate the file through /proc/<pid>/fd.
17. An application reports "too many open files". What do you check?
Check the process limit in /proc/<pid>/limits and the current count with ls /proc/<pid>/fd | wc -l. Raise the limit in the systemd unit with LimitNOFILE if the usage is legitimate, or fix the leak if descriptors keep growing.
18. What is the OOM killer and how do you tell it killed your process?
When the kernel runs out of memory it picks a process to kill based on its OOM score. Check dmesg or journalctl -k for "Out of memory" lines. In containers, the cgroup memory limit triggers the same mechanism.
19. Many connections are in TIME_WAIT. Is that a problem?
TIME_WAIT is normal for the side that closes the connection first. It becomes a problem when a client opens many short connections and runs out of local ports. Use connection pooling or keep-alive. Inspect states with ss -tan state time-wait | wc -l.
20. How do you check whether a service can reach a database on port 5432?
Resolve the name with dig, test the port with nc -zv db 5432 or curl -v telnet://db:5432, and check routes and firewall rules if it fails. ss -tnp on the database host shows whether it is listening.
21. What is the difference between a process and a thread?
A process has its own address space; threads within a process share memory and file descriptors. On Linux both are tasks created by clone() with different sharing flags, which is why threads appear in top -H.
Round 4: troubleshooting a live outage
The interviewer plays the system. Ask questions, say what you would check and why, and mitigate before you dig for root cause.
22. Users report the website is slow. Walk me through your first five minutes.
Confirm impact from dashboards: which endpoints, regions and percentiles. Check recent deploys, config changes and dependency status. Look at the golden signals for each tier to find where latency starts. If a recent change lines up, roll it back first and investigate after.
23. Error rate jumps right after a deploy, but only in one region. What next?
Stop the rollout and roll back that region. Then compare config, feature flags and dependencies between regions; a region-specific secret, endpoint or capacity limit is a likely cause.
24. Latency rises slowly over a week and then the service falls over. What could cause that?
A memory leak, a growing table without an index, a cache with no eviction, or log files filling a disk. Look for resources that trend upward and correlate them with latency. Add saturation alerts so the next trend is caught early.
25. A downstream service is down and your service is now timing out too. How do you stop the cascade?
Set timeouts lower than your callers' timeouts, add a circuit breaker that fails fast, and serve a degraded response or cached data where possible. Limit retries so you do not add load to the failing service.
26. Only a small percentage of requests fail, seemingly at random. How do you narrow it down?
Break errors down by host, pod, zone, client version and endpoint. Random-looking failures often come from one bad instance behind a load balancer. Remove it from rotation and examine it separately.
27. DNS changes were made and some users still reach the old servers. Why?
Resolvers cache records for the TTL, and some ignore it. Lower the TTL well before planned changes, keep old servers serving until traffic drains, and check with dig against several public resolvers.
28. Your monitoring is down during an incident. What do you do?
Fall back to direct checks: logs on hosts, kubectl status, synthetic probes from outside and the cloud provider's status page. Afterwards, make monitoring independent of the systems it watches.
Round 5: system design for reliability
29. Design a URL shortener that stays available if one zone fails.
Run stateless API servers in at least three zones behind a load balancer, store mappings in a replicated database across zones, and cache hot keys. Estimate traffic: reads far exceed writes, so the cache does most of the work. Discuss what happens to writes during a zone loss.
30. How would you design monitoring for a new service?
Start from user journeys and define SLIs for them, set SLOs with product owners, alert on SLO burn rate rather than on every metric, and add dashboards for the golden signals per dependency. Logs and traces support diagnosis.
31. What is non-abstract large system design (NALSD)?
The SRE Workbook uses NALSD for design exercises that end in concrete numbers: how many machines, how much bandwidth and storage, and how the design behaves when parts fail. Interviewers want you to move from a sketch to estimates and back.
32. How do you roll out a risky change safely?
Use a canary on a small slice of traffic, compare its SLIs with the baseline, then increase in stages. Keep a fast rollback, use feature flags to separate deploy from release, and avoid deploying to all regions at once.
33. How many nines can you promise if your service depends on three others at 99.9% each?
If all three are in the request path and fail independently, the best case is about 0.999 cubed, roughly 99.7%, before your own failures. You raise it with redundancy, caching or making some dependencies optional.
34. How would you plan capacity for a known traffic peak?
Use past peaks and growth to forecast, load test to find the breaking point per instance, add headroom for losing a zone, and pre-scale before the event. Agree on a load-shedding plan in case the forecast is wrong.
35. What are the trade-offs of active-active versus active-passive multi-region?
Active-active serves traffic from all regions and fails over quickly, but needs conflict handling for writes and costs more. Active-passive is simpler for data consistency, but failover is slower and the passive region may not work when needed unless you test it.
36. How do you decide what to alert on?
Page only for symptoms that hurt users now or soon and that need a human. Everything else becomes a ticket or a dashboard. Each page should have a runbook link.
Round 6: behaviour, incidents and on-call
37. Describe how you run an incident.
Name an incident commander, a communications lead and people working the problem. Keep a timeline in one channel, update stakeholders on a fixed rhythm, mitigate first, and hand off cleanly if the incident runs long.
38. What makes a postmortem blameless?
It looks at how the system and process allowed the failure, not who made a mistake. People describe what they saw and why their action made sense at the time. Action items have owners and dates.
39. A product manager wants to ship a feature but the error budget is spent. What do you do?
Refer to the error budget policy the teams agreed earlier. Show the data, propose options such as shipping behind a flag to a small group, and escalate through the agreed path if needed. The policy should make this a routine conversation.
40. Tell me about toil you removed.
Describe the manual task, how often it happened, what you automated and the time saved. Use your own real example and real numbers. Interviewers ask follow-ups, so avoid anything you cannot explain in detail.
41. How do you keep on-call sustainable?
Track pages per shift, fix or delete noisy alerts, share rotations across enough people, and give time back after a hard night. The SRE book suggests a cap on incidents per shift so engineers can handle each one properly.
42. You made a change that caused an outage. How do you handle it?
Say so early in the incident channel, help roll back, and write up what happened honestly. Then look at why the change was possible: missing tests, a missing canary or an unclear runbook.
43. How do you work with developers who see SRE as a gate?
Offer tools rather than reviews: deployment pipelines with built-in canaries, SLO dashboards per team and templates for alerts. Join their planning when reliability work is discussed.
44. What would you do in your first 90 days as an SRE on a new team?
Learn the architecture and its failure history, join on-call as a shadow, review existing SLOs and alerts, and pick one small automation or alert fix to deliver early.
How to prepare round by round
| Round | Practise |
|---|---|
| Coding | Log parsing, retries, rate limiting in Python or Go; talk while you code |
| Linux | strace, lsof, ss, perf, /proc on a test VM |
| Troubleshooting | Break a small app on purpose and narrate your debugging |
| Design | Two NALSD-style exercises with real numbers |
| Behaviour | Three true stories: an incident, a toil fix, a disagreement |
For tools to learn, see SRE tools for beginners and on-call management tools. Our DevOps training covers Linux, monitoring and incident practice in the Kharadi, Pune classroom and live online, and the Prometheus and Grafana course goes deeper on alerting.
Frequently asked questions
What rounds are in an SRE interview?
Most loops include a recruiter screen, coding or scripting, Linux and networking, a live troubleshooting exercise, system design and a behavioural round. The order and number vary by company.
Do SRE interviews include coding?
Usually yes. Expect practical problems such as log parsing, retries or rate limiting in Python, Go or Bash rather than puzzle-style algorithms, though some companies use standard coding questions.
What is the difference between SLI, SLO and SLA?
An SLI is a measurement, an SLO is the internal target for it, and an SLA is a customer contract with consequences. The SLA target should be looser than the SLO.
How much downtime does a 99.9% SLO allow?
About 43 minutes in a 30-day month, or roughly 8.8 hours a year.
What is NALSD in SRE interviews?
Non-abstract large system design: a design exercise from the Google SRE Workbook that ends with concrete numbers for machines, bandwidth and storage, and checks behaviour under failure.
How should I answer an SRE troubleshooting question?
Scope the impact, check recent changes, follow the golden signals tier by tier, mitigate first, then find the root cause and describe how you would prevent it.
Which Linux commands come up in SRE interviews?
top, ss, lsof, strace, dmesg, df, du, dig and perf appear often, along with reading files under /proc.
References
What's Your Reaction?
Like
0
Dislike
0
Love
0
Funny
0
Angry
0
Sad
0
Wow
0