System design cheat sheet

Numbers worth knowing

Memory read

About 100 nanoseconds

SSD read

About 100 microseconds

Same-data-centre round trip

About 0.5 milliseconds

Disk seek (spinning)

About 10 milliseconds

Round trip across continents

About 150 milliseconds

1 day

About 86,400 seconds, or roughly 100,000

1 million requests a day

About 12 requests a second

These are rough orders of magnitude. They are for estimates, not exact values.

Availability

99.9% (three nines)

About 8.8 hours of downtime a year

99.99%

About 53 minutes a year

99.999%

About 5 minutes a year

Scaling

Vertical

A bigger machine. Simple, but it has a limit

Horizontal

More machines behind a load balancer

Load balancing

Round robin, least connections, or hash of the user

Caching

Keep hot data close: browser, CDN, application, database

CDN

Serve static files from a place near the user

Queue

Do slow work later so the request returns fast

Databases

Replication

Copies of the data: read scale and failover

Sharding

Split data across machines by a key

SQL

Strong structure, joins, transactions

NoSQL

Flexible shape or huge scale; fewer joins

Index

Faster reads, slower writes

Trade-offs

CAP

During a network split, choose consistency or availability

Strong vs eventual consistency

Always the latest value vs the same value after a short wait

Latency vs throughput

Time for one request vs how many per second

Push vs pull

Send updates to clients vs clients ask for them

Interview steps

1. Clarify

Features, users, scale, what is out of scope

2. Estimate

Requests per second, data size, bandwidth

3. Draw the big picture

Clients, API, services, database, cache, queue

4. Go deep

The hardest part: data model, scaling, failures

5. Review

Bottlenecks, single points of failure, monitoring