Back-of-the-envelope maths

  • Lesson 3 of 40
  • 10 min
  • Beginner
  • 9 steps

Before choosing technology, work out the scale. A few rough sums tell you whether you need one server or a hundred, and how many terabytes.

Daily users
10M
Average
2,300 req/s
Peak
7,000 req/s
Servers
10
Storage / year
730 TB
10M daily usersOneserver?One server?Web tier · 10 serversStorageCDNCDNCache~21 GBCacheenough?The briefphoto app: 10M daily active userseach makes ~20 requests a day1M photos uploaded a day, 2 MB eachcan one server cope? how much disk?80%Requests per second10M users × 20 = 200M requests/day1 day = 86,400 s (≈ 100,000)200M ÷ 86,400 ≈ 2,300 req/s100%Peak trafficin the evening, ~3× the average2,300 × 3 ≈ 7,000 req/s70%Servers1 server ≈ 1,000 req/s (measured)7,000 ÷ 1,000 = 7 servers+ ~30% headroom10 serversStorage1M photos × 2 MB = 2 TB a day× 365 days ≈ 730 TB a year× 3 copies ≈ 2.2 PBplan ~2 PB for the first yearimages · ~16 Gbit/scache fillBytes per secondin: 2 TB/day ÷ 100,000 s ≈ 20 MB/sout: 10M × 100 views × 200 KB= 200 TB/day ≈ 2 GB/s≈ 16 Gbit/s → serve it from a CDNHot data in RAMprofiles10M × 2 KB = 20 GBnew photos1M × 1 KB = 1 GB≈ 21 GB → one 64 GB cache serverHandy numbers1 day≈ 100,000 s1M req/day≈ 12 req/s1 KB × 1M= 1 GB1 MB × 1M= 1 TBThe estimateaverage≈ 2,300 req/speak≈ 7,000 req/sservers10data≈ 730 TB/year (× 3)
  • request
  • response
  • waiting in queue

Step 1 of 9

Problem

Is one server enough?

You are asked to design a photo-sharing app. Before choosing any technology you need the scale: are we talking about one server or a hundred?

You do not need exact numbers, only the right order of magnitude. A few sums that fit on a napkin get you there in minutes.

Step 2 of 9

Technique

From daily users to requests per second

Start with what you know: how many users there are and how much they do. 10M users × 20 requests = 200M requests a day.

A day has 86,400 seconds; round it to 100,000. On a napkin speed matters, not decimals: 200M ÷ 100,000 ≈ 2,000, and the exact sum gives about 2,300 req/s.

What changed

  • Average load ≈ 2,300 req/s

Step 3 of 9

Concept

Size for the peak, not the average

Traffic is not flat: people use the app in the evening and it spikes with any event. A typical peak is 2 to 3 times the daily average.

At 3× that is about 7,000 req/s, and a single server is clearly overwhelmed.

What changed

  • Peak ≈ 3 × average ≈ 7,000 req/s
  • One server: overwhelmed

Step 4 of 9

Solution

How many servers?

A load test says one server handles about 1,000 req/s. The peak is 7,000, so we need 7, plus headroom for bursts and for when one goes down.

Call it 10. Notice the starting figure is a measurement, not a guess.

What changed

  • Web tier: 10 servers
  • Each one at ~70% at peak

Step 5 of 9

Technique

How much storage?

Photos are what weighs most: 1M uploads × 2 MB = 2 TB a day, about 730 TB a year.

Data is usually stored three times so a dead disk loses nothing, so plan for about 2 PB in the first year. That rules out «a single database» straight away.

What changed

  • Storage tier
  • 2 TB/day → ~730 TB/year
  • × 3 copies → ~2 PB

Step 6 of 9

Technique

How much bandwidth?

Uploading is the small part: 2 TB a day is about 20 MB/s in. Viewing is the big one: if each user views 100 photos a day as 200 KB thumbnails, that is 200 TB a day.

200 TB ÷ 100,000 s ≈ 2 GB/s, about 16 Gbit/s on average and more at peak. That is a job for a CDN, not for ten web servers.

What changed

  • The CDN serves the images
  • Egress ≈ 16 Gbit/s on average

Step 7 of 9

Technique

Does it fit in memory?

Feeds keep reading photo metadata and profiles. Does the hot part fit in a cache? Profiles: 10M × 2 KB = 20 GB. Today's new photos: 1M × 1 KB = 1 GB.

About 21 GB: it fits in the memory of a single cache server (plus a second one in case it fails). Nothing exotic needed.

What changed

  • Cache ≈ 21 GB in RAM

Step 8 of 9

Concept

Numbers worth knowing

A handful of numbers lets you estimate almost anything. Learn the order of magnitude, not the decimals:

Read from memory: ~100 ns · Read from an SSD: ~100 µs · Round trip within a data centre: ~0.5 ms · Round trip across an ocean: 75–150 ms · Seconds in a day: ~100,000.

Step 9 of 9

Summary

Summary

Users × actions ÷ 100,000 s = average requests per second.

Multiply by 2–3 for the peak and size for that.

Servers = peak ÷ measured capacity of one, plus headroom.

Storage = items a day × size × 365 × copies.

Also estimate outgoing bytes and memory: they often decide the design.

You will learn

  • Go from daily users to requests per second
  • Size for the peak, not the average
  • Work out servers from a measurement
  • Estimate storage, bandwidth and memory
  • Reference numbers for estimating