Original questions, not reported employer questions. Q09–Q15 are derived from documented cases or architecture; Q16 applies an operations framework to a hypothetical scenario. Answer aloud and state your assumptions before you open the guide; guides are editorial suggestions, not official answer keys.
Q09, Ramp case discussion, Scope & customers
A customer asks for a feature that will take three days. How do you decide whether to build it?
Case-derived question. Source: Forward Deployed Engineering ↗
Response guide
An original interview-style question derived from a published delivery example; not reported as an actual hiring question.
Ask about the underlying blocker, deadline, existing alternatives, and acceptance criteria. Compare a workaround with a reusable feature and include maintenance cost.
Follow-up to expect: When would a workaround create unacceptable long-term risk?
Q10, Cognition case discussion, SPE ownership
Employees are rapidly building internal apps, but no team owns a safe deployment environment. What do you do first?
Case-derived question. Source: Nick Wong, Special Projects Engineer ↗
Response guide
An original question based on Nick Wong’s published account; not a reported Cognition interview question.
Identify the workflows, sensitive data, users, and ownership gap. Build a paved path with clear permissions, maintenance ownership, and a narrow first use case.
Follow-up to expect: What evidence would justify turning the internal capability into a customer product?
Q11, Tax AI case discussion, Evaluation
Experts keep correcting your AI output. How would you turn those corrections into a reliable improvement process?
Case-derived question. Source: Building self-improving tax agents with Codex ↗
Response guide
An original question derived from the published Tax AI engineering case.
Capture context and traces, distinguish genuine errors from policy changes, create representative held-out cases, and test fixes against regressions before rollout.
Follow-up to expect: How do you keep the improvement loop from overfitting to the latest corrections?
Q12, Sonos case discussion, Adoption & value
Your support agent resolves chats quickly, but customers still struggle to use the product. What should you measure next?
Case-derived question. Source: How Sonos elevates the listener experience with Sierra ↗
Response guide
An original question grounded in the published time-to-music and customer-effort discussion.
Follow the full customer journey. Examine repeat contacts, actual task completion, cohort differences, and time to first value. Validate the data with users.
Follow-up to expect: How would you distinguish genuine resolution from a customer giving up?
Q13, Zanzibar architecture discussion, Security & data
A user loses access to a folder, then new content is added. How do you stop stale permission checks from exposing it?
Case-derived question. Source: Zanzibar: Google’s Consistent, Global Authorization System ↗
Response guide
An original design question based on the paper’s new-enemy problem; not a reported hiring question.
Model content and ACL versions, inherited permissions, and read consistency. Include caches and exports. State the revocation guarantee and test forbidden reads after changes.
Follow-up to expect: Where is an authorization check too stale to be acceptable?
Q14, Stripe engineering discussion, Migration
Move live customer subscriptions into a new data model without losing updates. How do you stage and reverse the change?
Case-derived question. Source: Online migrations at scale ↗
Response guide
An original question derived from the published online migration; not a Stripe interview report.
Separate write duplication, backfill, validation, read cutover, and retirement. Inspect concurrent updates, parity failures, load, and the conditions under which rollback remains possible.
Follow-up to expect: A backfill races with a newer write. Which version wins and how would you detect a regression?
Q15, GitHub incident discussion, Incident judgment
Failover has produced writes in two locations and increased application latency. What do you protect first?
Case-derived question. Source: October 21 post-incident analysis ↗
Response guide
An original incident discussion based on the 2018 postmortem; not a GitHub interview report.
State the data-integrity invariant, freeze unrelated changes, inspect replication, and choose a containment path. Explain the consequences of degraded service versus unsafe failback.
Follow-up to expect: Which queued jobs can be replayed safely, and which require reconciliation?
Q16, Production operations discussion, Unit economics
The model bill falls, but review and support increase. How do you decide whether the deployment improved?
Case-derived question. Source: Eliminating Toil · The Site Reliability Workbook ↗
Response guide
An original synthesis applying the published toil framework to FDE delivery; the scenario is hypothetical.
Use total cost per accepted task, include failures and human work, and hold task mix and quality constant. Separate realized savings from theoretical capacity.
Follow-up to expect: What changes if usage doubles but the support team cannot grow?