A Framework for Any Design

Two candidates know the same material. One hears "design a pastebin" and jumps between caching, sharding and the API. The other clarifies, estimates, then designs, and has time left for depth. The second one passes, on order alone.

A fixed order means you never skip a step, the interviewer can follow you, and no thought goes on what to do next.

#The seven steps

Every case study on this site follows these steps, with these headings. Each step produces something the next one uses: the estimates decide the data store, the access patterns decide the keys, the hard parts decide the deep dives. Minutes are for a 45-minute slot with about 40 minutes of design, as in how the interview works.

That adds up to 5 + 4 + 3 + 3 + 10 + 10 + 5 = 40 minutes. In a 60-minute slot, give most of the extra time to steps 6 and 7.

#Worked example: a pastebin

A pastebin stores a block of text and gives it a link. Here is the whole framework, at spoken length.

#1. Clarify the problem

• Functional: create a paste (text up to 1 MB, optional expiry) and get a link; read a paste by its link.
• Non-functional: a created paste is never lost; reads are fast (under 100 ms) and highly available; links are hard to guess.
• Out of scope: accounts, editing, search, syntax highlighting.

Requirements and scope covers which questions to ask.

#2. Estimate the scale

Round a day to 100,000 seconds (really 86,400, so rates come out about 14% low). The estimation lesson explains the shortcuts.

• Writes: 5 million ÷ 100,000 = 50 a second. Assume peaks of 3 times: 150.
• Reads: 10 × 50 = 500 a second, 1,500 at peak.
• Bandwidth out: 500 × 10 KB = 5 MB/s on average, 15 MB/s at peak. Small.

#3. Define the API

POST /pastes
  body: { "content": "...", "expires_in_days": 7? }
  201 Created   { "id": "x7Kp2QaZ", "url": "https://paste.example/x7Kp2QaZ" }
  413 Content Too Large   over 1 MB
 
GET /pastes/{id}
  200 OK   { "content": "...", "created_at": "...", "expires_at": "..." }
  404 Not Found   no such paste, or it has expired

#4. Model the data

The metadata is a small row per paste (about 200 bytes: id, size, timestamps, a pointer to the content), looked up only by id. The content is up to 1 MB of text that never changes. So: metadata in a key-value or relational store keyed by id, about 5 million × 200 B = 1 GB a day; content in blob storage under the key pastes/{id}.

The id is 8 random characters from 62 letters and digits: 62⁸ ≈ 2.2 × 10¹⁴ possibilities. After five years (9 billion pastes) a new id clashes about 1 time in 24,000, so insert-if-absent and retry. That ratio cuts both ways: one random guess in 24,000 also finds some paste, so a scraper can harvest them. If that matters, use 10 characters (about 1 in 90 million) and rate-limit 404s.

#5. High-level design

#6. Deep dives

Pick the parts where the numbers are hardest. Here: where content lives, and id generation (see unique IDs).

What would change the choice: if the interviewer capped pastes at 4 KB and kept them for 30 days, the total would be about 5 million × 4 KB × 30 = 600 GB, and a single database would be simpler.

#7. Bottlenecks and failures

At 10 times the load (15,000 reads a second at peak), add paste servers, since they are stateless, and turn the cache into a cluster. Metadata then grows about 3.65 TB a year (10 GB a day), so partition it by id; random ids spread evenly. Then walk the failures: a dead cache node sends reads to the stores, the metadata database fails over to a replica, and a blob storage outage stops creates.

#In the interview

Say the plan once, then announce each transition:

• "Before I draw anything, I want to pin down features and scale."
• "Writes are 50 a second and storage grows 18 TB a year. That second number matters most, so I'll keep content out of the database."
• "The design works end to end now. The two parts I'd like to go deeper on are content storage and id generation. Which interests you more?"

Follow-ups you should expect:

• "Can we skip the estimates?" Yes, if the interviewer says so. Ask for the one number that matters ("roughly how much data?") and move on.
• "Skip to the hard part." Go. The framework is a default order, not a script to defend.

#Common mistakes

#What to practice next

Each step has its own lesson, starting with Requirements and Scope. To see the full framework at case-study depth, work through Design a URL Shortener.