Crack Your Next Interview with AI Assistance

Vertical vs. horizontal scaling

Vertical scaling makes one machine bigger. Horizontal scaling puts more machines behind a load balancer. The real trade is not throughput — it is a hard ceiling and a single point of failure on one side, against distributed state and the complexity that comes with it on the other.

What interviewers are listening for:

• Start vertical. It is one config change, and most services never outgrow a single large box.
• Go horizontal when you need availability, not just throughput — one machine is one outage.
• Name what actually blocks it: session state, local disk, in-process caches, sticky connections.
• Treat the database separately. App tiers scale out easily; the write path almost never does.

Start from the constraint, not the wishlist. With fixed capacity, prioritisation is a sequencing problem: what unblocks the most downstream value per engineering week, and what becomes more expensive the longer it waits.

A structure that holds up under follow-ups:

• Anchor on one metric the quarter is judged by — retention, activation, cost per order.
• Score each candidate on impact on that metric ÷ engineering weeks, then sanity-check the top of the list against risk.
• Separate reversible bets from one-way doors; ship the reversible ones first and learn.
• Name what you are explicitly NOT doing, and why — interviewers are testing whether you can drop things.

Choose the forest when the relationship is non-linear or full of interactions you cannot enumerate, and you have enough rows to support it. Choose logistic regression when you need a coefficient you can defend to a regulator, a risk committee, or a colleague.

The answer interviewers are listening for:

• Random forest handles non-linearity and interactions without you specifying them.
• Logistic regression gives calibrated probabilities and interpretable coefficients — often a hard requirement in credit, fraud and healthcare.
• Forests need more data and more compute, and are far easier to overfit on small, noisy tables.
• Say you would baseline with logistic regression first — it makes the forest’s lift measurable instead of assumed.

Do not discount, and do not defend the number. "Too expensive" is almost never about price — it is about value that has not been established yet, or a budget that sits with someone who is not in the room. Find out which one it is before you answer.

The response interviewers want to hear:

• Ask what they are comparing you against — a competitor, doing nothing, or building it in-house. Each needs a different answer.
• Reframe to cost of the problem: what the status quo costs them per month, against what you charge.
• Check whether you are talking to the budget holder. If not, the objection is a routing problem, not a pricing one.
• Trade, never concede: if the price moves, something moves with it — term length, seats, a case study, a faster start.

Answer in STAR, and pick a disagreement you lost as often as one you won — interviewers are testing whether you can commit after a decision goes against you, not whether you were right.

How to structure it:

• Situation and Task in two sentences — enough context that the stakes are clear, no more.
• Action is where the marks are: say how you raised it, what evidence you brought, and that you took it up privately first.
• Result should include the number — the ship date you protected, the incidents you avoided.
• Close with disagree-and-commit: what you did once the call was made, even when it was not your call.