Oct 08, 2026 by Offer

Millennium System Design Interview 2027: Median Across Servers

For the Millennium median 10TB 10 servers question, use distributed binary search on the value range: each round, every server counts values below the midpoint, you sum the counts, and narrow toward the median. Data stays put; only tiny messages move. Commonly reported by candidates.

What This Question Assesses

This tests distributed-systems thinking: you cannot move 10TB to one machine, so the algorithm must come to the data. The interviewer wants to see you recognise the constraints (network is the bottleneck, data stays put), propose a correct distributed algorithm, and analyse its communication cost. Candidates who suggest "just sort it" or "sample it" miss the point — the question is about exact computation under distribution.

Millennium Median 10TB 10 Servers: How to Answer

Step 1 — State the constraint. "10TB cannot be centralised — moving it would saturate the network. The data stays on the 10 servers; only small messages move."
Step 2 — Propose distributed binary search. "We binary-search the value range [low, high]. Each round, the coordinator broadcasts a midpoint m; each server returns the count of local values ≤ m. If the total count ≥ 5TB worth of elements, the median is ≤ m — search lower; else search higher."
Step 3 — Analyse the cost. "Each round moves 10 small integers — negligible. About 50 rounds for full 64-bit precision, fewer if we accept approximate precision or quantise."
Step 4 — Discuss alternatives and trade-offs. "T-digest or Q-digest sketches give approximate medians in one pass; exact distributed selection algorithms exist but are more complex. For an exact median with minimal communication, binary search on values is the clean answer."

An example line: "I would keep the data in place and binary-search the value range instead — each server counts values below the midpoint, I sum ten integers per round, and fifty rounds pin down the exact median with almost no network traffic."

Millennium Median 10TB 10 Servers: Common Mistakes

Proposing to centralise the data. Moving 10TB defeats the purpose and shows you missed the core constraint. Always start from "data stays put."
Confusing median of medians with the true median. The median of per-server medians is not the global median — it is a common wrong answer. Name it only to reject it.
Ignoring the communication analysis. The algorithm is only half the answer; quantifying rounds and message sizes is what makes it a system design answer.

Distributed median questions test whether your algorithms survive contact with real infrastructure — lead with the constraint, then the algorithm.

Keep Reading

FAQ

Why not just sample the data? Sampling gives an approximate median with confidence bounds, which may be acceptable in practice. But the question asks for the median — the binary search gives it exactly, so lead with exact and offer sampling as a pragmatic alternative.

How many rounds does the binary search need? One per bit of precision: ~32 for 32-bit integers, ~64 for 64-bit. In practice you stop earlier once the range is narrower than your required precision.

What if the data is not numeric? The approach generalises to any totally ordered domain where you can test "≤ midpoint" — for strings, binary search over the lexicographic range works the same way.

Preparing for Millennium's interview? Our 2027 Millennium Caliper Assessment and Quantitative Assessment Exact Questions and Answers has practice questions and answers — $79 one-time, instant download.