Swiggy: the p90 the average is hiding

Constraints

1. 1Delivery latency is the whole number of minutes between ordered_at and delivered_at.
2. 2Use the nearest-rank definition: sort a city's latencies ascending and take the value at position CEIL(0.9 * n) for p90.
3. 3Median is the value at position (n + 1) / 2 under integer division — with 9 deliveries that is position 5.
4. 4The cities have different numbers of deliveries — 9, 8 and 7 — so the positions must be computed per city, never hard-coded.
5. 5avg_minutes is rounded to 1 decimal place; the percentiles are actual observed values, so they stay whole numbers.
6. 6Order by p90_minutes descending, then city.

Worked example

Bengaluru's nine latencies sorted are 22, 26, 28, 29, 32, 35, 41, 55 and 95 minutes. The median is position 5, so 32. p90 is position CEIL(0.9 * 9) = 9, so 95. The average is 40.3, which describes no delivery that happened: it is 8 minutes above the median and 55 below the worst. A single 95-minute delivery drags it up far enough to look like a mild problem, while the customer who waited an hour and a half is invisible in it.

What this tests

Building a percentile from positions when the engine provides none, and reading the difference between a mean and a tail — the reason latency is reported at p90 or p99 rather than on average.