How Netflix interviews — the keeper test

1. The shape of a Netflix loop

Netflix runs the smallest loop of any company in this course. Three or four rounds for senior engineers. Five at the very top. The loop is short because Netflix doesn't want to spend the team's time on a candidate the recruiter and hiring manager aren't already serious about.

A standard onsite for L5 senior is one coding round, one system design, one behavioral with the hiring manager, and one cross-team round. L6 staff adds a second design or a deep technical conversation with a director. L7 principal adds an executive round.

The behavioral round is the one most candidates underprepare for. At every other FAANG, the behavioral is a screen for obvious red flags. At Netflix, the behavioral round is the round most likely to get you rejected even after the technical rounds went well.

2. The keeper test

Every Netflix manager runs the keeper test on every direct report, every quarter. The question is plain: "if this person told me today they were leaving for a similar role at another company, would I fight to keep them?"

If the answer is no, the manager is expected to part ways with the report and offer a generous severance. There is no performance improvement plan culture. There is no holding on to mediocre talent because firing is awkward.

For an interviewer, the keeper test runs in reverse: "if this candidate joined and I worked with them for six months, would I fight to keep them?" Hire-strong means yes. Hire-weak means probably. No-hire means no.

This pressure shows up in how interviewers grade. They aren't looking for someone who can do the job. They're looking for someone who would raise the bar of the team they're joining. A competent answer is a no-hire here. A competent answer at Google is a hire-weak.

3. Freedom and responsibility

Netflix's published culture deck describes the company as freedom and responsibility. Engineers are given large amounts of autonomy and expected to make decisions a director would make at most other companies.

In the technical rounds, this leaks in two ways:

You're expected to drive. The interviewer will give you the question and then sit back. If you stop and ask "what should I do next?" every two minutes, you fail the autonomy signal. Most candidates from Google and Microsoft find this awkward at first.
You're expected to defend your choices. When you pick a primitive, the interviewer might say "why not the other one?" They aren't pushing because you're wrong. They're checking if you can articulate the cost. If you fold immediately and switch to their suggestion, you fail the responsibility signal.

The senior pattern: pick, defend, name the inversion case, move on.

4. What Netflix grades you on

Four signals, weighted heavier on the right side than other companies:

Technical depth at one seam. Netflix engineers are expected to go very deep on the specific system they own. A breadth-first answer that touches ten primitives lightly grades worse than a depth-first answer that picks two and defends them to three pushes.
Reliability instinct. Netflix has shipped Chaos Monkey, Hystrix, the Simian Army. The expectation is that you think about failure before you think about happy path. "What happens when this dependency is slow?" is asked in every design round.
Cost awareness. Netflix runs on AWS at a scale where compute and network costs are eight-figure monthly line items. Designs that ignore cost grade worse here than at any other company in this course.
Culture fit on freedom and responsibility. This is the behavioral round. Stories where you made a big call without asking permission score well. Stories where you waited for direction score badly.

5. The level bar

How the bar shifts:

L4 / L5 (senior) — Netflix's L5 bar is roughly Google's L6. Expect to defend three pushes on the deep dive, name primitives correctly, and articulate cost on every choice. Reliability is graded explicitly.
L6 (staff) — design at the platform level. You're not just designing the system. You're designing the system that other teams build on top of. The chaos engineering question hits hard at this level.
L7+ (principal) — design at the org level. You're shaping how the company spends money. Cost articulation moves from "we'd pick the cheaper primitive" to "here's how this design saves $4M/year at our scale."

In this chapter, units target L5 or L6 by default. The take-home in 8.7 is L5.

6. The Netflix-flavored mistakes

Watch for:

Skipping the failure model. "And what if this service is unreachable?" is asked on every design. If you reach minute 25 without naming a failure mode unprompted, you've already lost depth points.
Ignoring cost. Designs with 50 microservices each running in 3 regions are theoretically fine. They cost $50M/year. Netflix interviewers expect you to mention cost as a constraint without being asked.
Drawing the diagram without driving the conversation. Netflix interviewers will let you go silent. If you let the silence stretch, they assume you don't know what to do next.
Folding under pushback. "Why not Cassandra here?" is not a hint that you should use Cassandra. It's a check that you can defend your choice. Defend, then offer the inversion case.
Treating chaos engineering as a feature. It's a continuous practice at Netflix, not a project. If you describe it as something you'd build "once the system is stable," you've signaled you don't understand the culture.

7. The seven scenarios in this chapter

Each unit dissects one question end to end. The chapter covers:

8.1 Netflix's video CDN — Open Connect, ISP-embedded appliances, tier model. The signature Netflix question.
8.2 Chaos engineering harness — Chaos Monkey, the Simian Army, fault injection as a continuous feature.
8.3 Profiles and per-profile state — one account, many family members, shared subscription with separate state.
8.4 Continue Watching row — cross-device sync of position-within-title, the universally beloved feature.
8.5 A/B testing infrastructure — experiment definitions, sticky assignment, exposure logging, significance.
8.6 Downloads for offline playback — DRM license model, expiration, sync on reconnect.
8.7 Take-home — fault injection middleware — TypeScript code task. The grader runs npm test.

8. The next page

The first scenario: Netflix's video CDN. The system that lets a single show ship to 200M households on the same night without melting the internet.