Netflix: Interview Perspective

Don't start drawing — open with the control-plane/data-plane split immediately, per Opening Move.

• "What if we also need to support live streaming — say, sports?" — Tests whether the candidate recognizes live fundamentally breaks the pre-encode-and-cache model: you can't pre-encode or pre-position content that hasn't happened yet, and latency tolerance drops sharply.
• "What if this needs to work well in a market with very poor average connectivity?" — Tests whether the candidate reasons about lower bitrate-ladder rungs and more aggressive adaptive behavior, not just "add more CDN."
• "How would offline downloads change this design?" — Tests whether the candidate identifies this as a control-plane-plus-local-storage problem, distinct from CDN streaming entirely.

Point to the Open Connect content in Architecture Deep Dive — this is a real, scale-justified decision, not a default.

• "At what scale does building your own CDN actually make sense?" — Tests whether the candidate reasons about a break-even point (bandwidth cost savings vs. infrastructure/ops investment) rather than assuming bigger is always better.
• "How would you decide between embedding hardware inside an ISP versus peering at an Internet Exchange Point?" — Tests understanding of that specific trade-off, not just awareness that Open Connect exists.

No — point to the per-title encoding content in Architecture Deep Dive.

• "What does per-title encoding actually cost you?" — Tests awareness of the real compute trade-off, not just the benefit.
• "How would you decide if a title needs per-shot encoding instead of just per-title?" — Tests whether the candidate can reason about diminishing returns rather than assuming finer-grained is always better.

Signal pipeline, then a hybrid approach — point to the Recommendation System layer for the full comparison.

• "How do you handle a brand-new user with no watch history?" — Tests cold-start awareness.
• "Should 'Continue Watching' use the same pipeline as 'Because You Watched X'?" — Tests whether the candidate distinguishes real-time session-driven rows from batch-computed ones, a genuinely nuanced production distinction.

Point to the Watch State Sync layer; contrast explicitly with why this doesn't need Uber-style continuous push infrastructure.

• "What if the same show is being watched on two devices at once?" — Tests conflict-handling reasoning (last-write-wins, and why that's fine here).
• "Does this need the same real-time push infrastructure as showing a driver's location on a map?" — A direct cross-system comparison question; tests whether the candidate over-engineers by reaching for a persistent connection when a simple read-through fetch is sufficient.

Point to the Regional Failure scenario for the reroute mechanism.

• "How is this different from a regional failure for a ride-hailing platform?" — Tests whether the candidate identifies why a stateless, cache-based system (Netflix's CDN) recovers more gracefully than a stateful one carrying real in-flight state (Uber's active trips) — the blast radius is fundamentally different in kind, not just degree.

Point to Adaptive Bitrate Streaming — fixed maximum quality ignores real, constantly-changing network conditions and would cause frequent buffering on anything less than an excellent connection.

• "How quickly should the player react to a drop in bandwidth?" — Tests whether the candidate understands the trade-off between reacting fast (risk of unnecessary quality thrashing) and reacting slow (risk of a stall).

Point to Pre-encoding vs On-demand Transcoding — pre-encoding wins at streaming scale generally, but the calculus is worth questioning at the long tail.

• "What would change your answer for extremely rare or low-traffic content?" — Tests whether the candidate can reason about the same trade-off at a different point on the popularity curve, rather than treating the decision as universal.

Point to Centralized vs Multi-region Control Plane — users already playing a title keep streaming (CDN doesn't need the control plane mid-playback), but browsing, search, auth, and starting anything new all fail.

• "Would a user notice anything if they were already watching something when this happened?" — Tests whether the candidate correctly separates what the control plane hands off once versus what it's needed for continuously — a strong answer says no, mid-playback is unaffected.

Tests recognizing the live-content break identified in the Opening & Scope category above, now applied mid-design — live breaks the pre-encode-and-cache model entirely, since nothing can be pre-positioned before it happens.

• "What would need to be true of the CDN and encoding pipeline to make this work?" — Tests whether the candidate can reason forward to near-real-time chunked encoding and much shorter cache lifetimes, rather than just naming the problem.

Tests prioritization under constraint — a strong answer identifies that per-title encoding, Open Connect, and elaborate recommendation infrastructure are all scale-justified investments that can be deferred, while the core control-plane/data-plane split and basic adaptive bitrate cannot.

• "At what point would you revisit each of those cuts?" — Tests whether the candidate ties each deferred piece back to the specific scale threshold that justifies it, rather than treating the MVP as permanent.