1
Convert a rate and know what the shortcut cost
Per-day to per-second and back with the 10⁵-second day, stating the 15.7% error it introduces and when that error matters.
Course · System Design
Capacity Estimation
Ten million users, a hundred actions a day, a kilobyte each: how much machine is that? The course teaches estimation as arithmetic with error bars — every factor is a range, every product is a wider range, and an estimate that arrives as a single number is hiding its width. It ends by checking an estimate the only honest way, by a second route that shares no assumption with the first.
1 day ≈ 10⁵ s 15.7% high, and worth it
three factors ±2× ⟹ product ±8×
storage = ingest × retention × replication, plus growth
N = ⌈peak / (capacity × ρ_target)⌉ an integer, with headroom named
What this course is about
- Subject
- System Design
- Lessons
- 10
- Format
- Interactive
Overview
What you will be able to do
By the end you can turn a population into a rate, a rate into storage and bandwidth, and a peak into a machine count — and say, for each, how wide the answer is and which assumption would move it.
2
Multiply factors that are ranges
Three factors each known to a factor of two give a product known to a factor of eight; report [lo, hi] and the width in powers of ten rather than a single number.
3
Size storage, bandwidth and memory
From an ingest rate, a retention window, a growth rate and a replication factor; in bits where the link is measured in bits; and for the hot fraction rather than the whole dataset.
4
Produce a machine count and then check it
An integer N from peak load, per-request cost and a named headroom — then a second route sharing no assumption with the first, and the ratio between them against the range.
Lessons
10 lessons
Ranges first, because everything after is a product of them; then the four things a rate sizes — storage, bandwidth, machines, memory — and last the check.
Orders of MagnitudeMultiply three factors that are ranges and report the interval and its width, not a single number.Not marked
Rates: the Day Has 10⁵ SecondsConvert a daily volume to a per-second rate both ways, and state the size and direction of the rounding error.Not marked
Read/Write RatioSplit a total request rate by a stated read-to-write ratio and report both rates as exact shares.Not marked
Storage from Ingest and RetentionSize storage after a retention window with a stated monthly growth rate and a stated number of copies.Not marked
Bandwidth: Bits, Bytes and OverheadCompute the megabits per second a request rate needs, with per-request headers included, and find the payload at which headers are half the traffic.Not marked
Latency Numbers on a Log ScaleRead the exact ratio between two operations off a logarithmic ruler and state which decade it falls in.Not marked
Peak to AverageCompute peak and mean request rates from a 24-bucket daily profile and report the provisioning multiplier between them.Not marked
From a Request's Cost to a Machine CountProduce an integer machine count from a peak rate, a per-request cost and a stated target utilisation, and say what the headroom is for.Not marked
Memory and the Working SetSize a cache in gigabytes for a target hit rate, given a dataset size and a stated hot fraction.Not marked
Triangulating an EstimateBuild two independent routes to one quantity and test their ratio against the width the inputs deserve.Not marked
Recommended background
Algebra: scientific-notation, integer-exponents, ratio-proportion-and-percent, logarithmic-scales, arithmetic-sequences-and-series, geometric-sequences-and-series.
Practice suggestions
Write the unit chain down before computing anything. Nearly every wrong capacity estimate is a unit error, and the lab prints the chain so that a missing /s is visible rather than merely wrong.
Keep the range to the end. The lab carries [lo, hi] through every step; the temptation is to collapse to the midpoint early, and the last lesson is about what that costs.
Do the second route before looking at the first answer again. Triangulation only works when the two routes are built independently, which is a fact about the order you do them in.
What it does not cover
Prices. What a machine or a gigabyte costs belongs to Scaling Laws and Cost; this course counts machines and gigabytes.
Latency. How long a request takes belongs to Latency and the Tail; this course counts how many of them arrive.
Why the headroom is what it is. The target utilisation is taken as given here and earned in Queues and Utilisation, where the knee explains it.