Roadmaps / Embedded 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 Principal probe and 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 |
|---|---|---|
| Yocto or Buildroot? | When would you switch? | The cost of the switch, team skills and the support window, with your own build numbers |
| How do you make a build reproducible? | What can still differ? | Pins, mirrors and timestamps, and how you detect a drift before a release |
| What goes wrong with a cache? | How do you trust it? | Keys, invalidation, and building releases from a clean cache |
| Why a read-only root? | Where does state live? | Data partition layout, wear, and how you migrate state in an update |
| How do you keep an image small? | What do you lose? | Debuggability against size, and a policy for debug images |
Principal probe: who owns the layer pins and what does it cost when nobody does? A strong answer names one owner, a pin-review cadence, and the delay in days between an upstream fix and your build.
Design prompt (35 minutes): build an image for 5,000 sites that you can rebuild in five years. Outline: sources and pins (10 min), the layer structure (10), the checks in CI (10), what you archive (5).
Phase 2
| Question | Follow-up | Staff adds |
|---|---|---|
| How does the kernel find a driver for a device? | What if two match? | Compatible lists, probe order and deferral |
| Config fragment or full config? | How do you upgrade the kernel? | The effort per release and a regression test set |
| A driver does not bind; how do you debug it? | What if there is no console? | dmesg, the sysfs driver directory and the device tree as booted |
| How do you reduce boot time? | What is the floor? | A measured breakdown, not a guess |
| Out-of-tree or upstream driver? | Who maintains it? | The cost of carrying patches for ten years |
Principal probe: when would you refuse a vendor kernel, and what would you do instead? A strong answer weighs the patch count, the vendor's rebase record and the exit cost, and says what you would carry yourself.
Design prompt (35 minutes): bring up a new sensor board with a vendor kernel. Outline: what you check first (5 min), the device tree and driver plan (10), the test (10), the maintenance plan (10).
Phase 3
| Question | Follow-up | Staff adds |
|---|---|---|
| How do you make an update survive a power cut? | What about the bootloader itself? | The ordering argument and the trial count, including the bound |
| A/B or recovery partition? | What does each cost? | Flash budget, download cost and failure modes |
| How do you roll back a bad update? | What if it boots and is broken? | A health check that tests the product, and who sets the threshold |
| What does signing protect against? | What does it not? | Key storage, rotation and what you do when a key leaks |
| RAUC, SWUpdate or Mender? | When would you pick each? | Fleet management, who operates it and the exit cost |
Principal probe: how many trials would you need before you tell the business "no bricks", and what else would you want? A strong answer gives the per-window bound, the field weighting, and the failure classes a QEMU kill cannot show.
Design prompt (40 minutes): design the update path for a device with 4 GB of flash and an intermittent link. Outline: slot layout and budget (10 min), update flow and failure cases (10), signing and keys (10), the test plan (10).
Phase 4
| Question | Follow-up | Staff adds |
|---|---|---|
| What is in an SBOM? | What is not? | Uses and limits; keeping it per release |
| A CVE appears in a library you ship; what do you do? | How fast? | Triage statuses, a timeline and a rehearsed drill |
| How do you do a staged rollout? | What stops it? | Health thresholds fixed in advance and who can override |
| What does the CRA mean for your product? | What if you ship to non-EU customers? | Support period, reporting and documentation, and the gaps you found |
| How do you support a device for ten years? | What breaks first? | Kernel and toolchain cadence, key expiry and hardware end of life |
Principal probe: how do you show a committee that your process reduced exposure rather than only produced documents? A strong answer compares time from awareness to signed bundle, and the share of fleet on the fix after 14 days, before and after, for the same products.
Design prompt (45 minutes): own the security process for 20 product lines on Yocto. Outline: what you standardise (10 min), the CVE and SBOM pipeline (10), reporting and staffing (10), measures that it works (10), what you leave to teams (5).