Roadmaps / Linux / Interview topics
Reference
Interview topics
Each question has the follow-up an interviewer might ask and what a Staff answer adds; each phase adds a design prompt with a timed outline. A good design answer names the budget, the failure that matters, the measurement and the rollback.
Phase 1
| Question | Follow-up | Staff adds |
|---|---|---|
| What happens between pressing the power button and a login prompt? | Where would you look if it hangs after the kernel starts? | Which stage has a console and logs to read, and what you do on a remote host with no console |
| How do you make a service restart safely after a crash? | What if it crashes on start every time? | Start-limit settings, backoff, and an alert that fires before the restarts hide the failure |
| What is the difference between a process and a thread in Linux? | What does clone do? | Why a thread-heavy service needs TasksMax and a limit on file descriptors |
| What does a zombie process mean? | How do you remove one? | That the parent must call wait; fix the parent, not the zombie |
| How would you sandbox a service you do not trust? | What are the limits of that sandbox? | Name what the sandbox cannot stop (flaws inside the allowed set) and the cost in compatibility |
Principal probe: which unit default would you change for every service on a fleet, and what would that break? A strong answer picks one (for example Restart= with a start limit), shows the measured effect on one service, and names the service it would hurt.
Design prompt (35 minutes): a host runs one service that must restart within 5 seconds and not lose data. Outline: failure cases (5 min), unit and sandbox (10), logging and alerts (10), what you measure and the rollback (10).
Phase 2
| Question | Follow-up | Staff adds |
|---|---|---|
| The host is slow; what do you check first? | What if load is high and CPU use is low? | The USE method, PSI, and what each tool cannot see |
| How does the OOM killer choose a victim? | How do you protect one process? | Why oom_score_adj is a last resort and cgroup limits are the plan |
| What does a page cache hit rate tell you? | When is drop_caches useful? | Only for experiments; in production it hides the real working set |
What is the difference between write and fsync? | What does synchronous=FULL cost? | The measured ratio and when losing the last second of data is acceptable |
| How would you stop a noisy neighbour on a shared host? | What if two services must share a CPU? | cpu.weight against cpu.max, and why a quota can raise tail latency |
Principal probe: how would you set memory limits for services you have never measured? A strong answer observes in staging first, sets limits from the p99 working set with headroom, and alerts on oom_kill before enforcing.
Design prompt (35 minutes): twenty services share a 64 GB host and one leaks. Outline: limits and defaults (10), detection (10), what the owner sees (5), the policy and its cost (10).
Phase 3
| Question | Follow-up | Staff adds |
|---|---|---|
| A request takes exactly 5 seconds sometimes; why? | How do you prove it? | The resolver timeout, a tcpdump showing the retry, and the fix with a number |
| How does a client run out of ports? | Is raising the port range enough? | Connection reuse and pooling are the real fix; the range only moves the limit |
| What does a default-drop firewall break? | How do you roll out a change safely? | A console fallback, a timed rollback, and a test from outside |
| What is a network namespace? | How does a container use one? | That a container is a set of namespaces and cgroups, not a different kernel |
| Why drop capabilities? | Which would a web server need? | Often none above port 1023; CAP_NET_BIND_SERVICE only for low ports |
Principal probe: who owns the firewall policy when five teams share a host? A strong answer separates the host baseline (one owner, default drop) from per-service rules (each team, reviewed), with a tested rollback.
Design prompt (35 minutes): allow a new service to call one external API and nothing else. Outline: threat and failure cases (5 min), firewall and DNS (10), logging and alerts (10), rollout and rollback (10).
Phase 4
| Question | Follow-up | Staff adds |
|---|---|---|
| How would you find why a process uses 100% CPU? | What if it is in the kernel? | perf record -g, kernel against user time, and what a flame graph hides |
When do you use eBPF and when strace? | What does each cost? | strace slows the target heavily; eBPF has low overhead but needs kernel support and care |
| How do you upgrade a host with no downtime? | What if the new kernel does not boot? | A second instance or drain first, the old kernel entry as the way back |
| What is a good backup test? | How often? | A scheduled restore with a measured recovery time and recovery point |
| Walk through a postmortem you wrote. | What changed afterwards? | The measured outcome some months later, not only the actions |
Principal probe: how do you show a committee that a standards change reduced incidents rather than moved them? A strong answer compares incident counts and time to restore before and after for the same services over a stated period, and names what else changed.
Design prompt (45 minutes): own the host standards for 300 Linux servers. Outline: what you standardise (10 min), limits and sandbox defaults (10), upgrade and rollback policy (10), how you measure that it works (10), what you leave to teams (5).