Skip to content
Tutorials Center
My learning

Topics

Lesson 63 of 65 ยท Practice and Interviews

Back-of-Envelope Estimation Practice

Practice system design estimation: powers of two, latency numbers, the seconds-per-day trick, and five worked problems for QPS, storage, bandwidth and servers, with Python.

Back-of-envelope estimation means getting a rough answer fast, using simple arithmetic and a few memorized numbers. In a design interview, and in real planning, it tells you whether you are building something that fits on one machine or something that needs a fleet. The goal is the right order of magnitude (is it 100 or 100,000?), not the exact digit.

This lesson gives you the numbers to memorize, a method, and five worked problems with Python checks. Do each problem on paper first, then compare.

Numbers to memorize

Powers of two

Power Approximate value Name
2^10 1 thousand 1 KB
2^20 1 million 1 MB
2^30 1 billion 1 GB
2^40 1 trillion 1 TB
2^50 1 quadrillion 1 PB

Latencies (rough, modern hardware)

Operation Time
CPU cache or memory read 1 to 100 ns
SSD random read about 100 microseconds
Read 1 MB from memory about 100 microseconds
Round trip within a data center about 0.5 ms
Read 1 MB from SSD about 1 ms
Disk seek (spinning) about 10 ms
Round trip across continents about 100 to 150 ms

The lesson of the table: memory is fast, the network inside a data center is cheap, and crossing oceans is expensive. Avoid chatty cross-region calls.

Time: a day has 86,400 seconds, which you can round to 100,000 (a 15 percent error is fine). A month is about 2.5 million seconds; a year is about 30 million.

Availability as downtime per year: 99% is 3.65 days; 99.9% is 8.8 hours; 99.99% is 53 minutes; 99.999% is 5 minutes.

Python Check the memorized tables

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The method

  1. State assumptions out loud (users, actions, sizes). Round them.
  2. Convert to per-second by dividing daily totals by 100,000.
  3. Multiply for peak (2 to 5 times) and for sizes.
  4. Sanity check against known ranges: one server, one database, one data center.
  5. Write units next to every number, and say the answer in a sentence.

Problem 1: tweets per second and storage

A service has 300 million daily users; each posts 0.5 tweets per day on average, each tweet is 300 bytes of text plus metadata. How many posts per second and how much storage per year?

Python Problem 1: posting traffic and text storage

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

About 1,700 writes per second (5,000 at peak) is easy for a sharded database. Text alone is only about 27 TB per year. Media (images, video) will be the real storage problem.

Problem 2: image storage and CDN bandwidth

Users upload 10 million photos per day, average 2 MB after processing, plus three resized copies of 100 KB total. Each photo is viewed 50 times on average. Storage per year, and egress bandwidth?

Python Problem 2: photos

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

About 8 PB per year and tens of Gbps of egress. The answer is: object storage with tiering, resized derivatives (serve small versions most of the time), and a CDN. No single data center link should carry this alone.

Problem 3: how many servers?

A web tier must handle 20,000 requests per second at peak. Each request uses 20 ms of CPU. A server has 16 cores and should run at 60 percent CPU. How many servers?

Python Problem 3: server count

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

400 cores of work, 10 usable cores per server, so 42 servers. Multiplying CPU time by QPS gives core-seconds per second, which is just the number of busy cores.

Problem 4: cache memory

A site has 100 million items; each cached record is 1 KB. The hottest 10 percent of items serve 90 percent of the reads. How much memory to cache them, and how many 64 GB nodes?

Python Problem 4: cache sizing

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

About 10 GB: it fits on a single node, so you add a second for availability rather than for capacity. Always check whether the "big" number is actually small.

Problem 5: Little's law and connections

A chat service has 50 million concurrent users, each holding one WebSocket. A chat server can hold 200,000 connections. How many servers? If each user sends 1 message per minute averaging 200 bytes, what is the inbound message rate?

Python Problem 5: connections and messages

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

250 servers just to hold the connections, and 830,000 messages per second: connection count, not bandwidth, is the first constraint. This matches the shape of the chat system design.

Tips

  • Round aggressively: 86,400 becomes 100,000; 2.3 becomes 2.
  • Write the units each time (bytes versus bits is a classic slip; network speeds are bits).
  • Be consistent about decimal versus binary prefixes; the difference (about 7 percent at GB scale) rarely matters.
  • Estimate in layers: users, then actions, then requests, then bytes.
  • Compare to known capacity: if the result fits on one big machine, say so.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Roughly how many seconds are in a day, for quick math?

    86,400 rounds to 100,000, which makes daily-to-per-second conversion easy.
  2. 2. Which is faster: a round trip inside a data center or across continents?

    About 0.5 ms versus 100 to 150 ms.
  3. 3. A request needs 20 ms of CPU and arrives 1,000 times per second. How many cores are busy?

    1,000 x 0.02 s = 20 core-seconds per second, so 20 cores.
  4. 4. 99.99% availability allows about how much downtime per year?

    0.01 percent of 525,600 minutes is about 53 minutes.

Summary

  • Memorize powers of two, the main latency numbers, 100,000 seconds per day and availability tables.
  • Method: state assumptions, convert to per-second, apply peak factors, sanity check, write units.
  • Core-seconds, bytes per second and connections per server cover most interview estimates.
  • Always ask whether the large-looking number is actually small for modern hardware.

Next: System Design Trade-offs Cheat Sheet