Part 4 of 8 · 2 min read
Deep dive
The middle thirty minutes.
Choosing the deep dive. Pick the part where the numbers or the correctness are hardest. Examples: the fan-out in a feed, the lock in a booking, the ordering in a chat, the hot key in a counter. If unsure, ask: "I think the hardest part here is X. Shall I go deep there?"
Shape of a deep dive.
- State the problem with a number.
- Give the simple answer, and why it fails.
- Give the better answer.
- Give the cost of the better answer.
- Say how you would check that it works (a test or a metric).
Example. "A counter row updated 5,000 times a second becomes a hot spot (problem). A simple row update with a lock fails because the lock serialises writes (simple). Use 16 sharded counters and add them on read (better). The cost: reads add 16 numbers, and a read is slightly out of date (cost). I would watch the lock wait time and the total error against a nightly exact count (check)."
Follow-ups. Typical ones: "What if the traffic is 10 times higher?", "What if this server fails?", "How would you add feature Z?", "How do you test this?" Answer in three steps: restate, change the design, name the new cost.
Hints. When the interviewer says something like "What about the database?", it is a hint. Do not defend the old design. Say "Good point, let me look at that" and work through it. Taking a hint well scores better than ignoring it.
Mistakes. If you notice an error ("I said 500 a second, but it is 5,000"), correct it at once and say what changes. Calm corrections are a strength.
When you are lost. Say what you know, name the unknown, state an assumption and move on. Then return if time allows.
Common mistake
Drawing a large design in minute 5 and then having no time to go deep. Start with the simple design and let numbers tell you what to add.
Check yourself: 1. What are the five steps of a deep dive? 2. How do you take a hint? 3. What is the right response to your own arithmetic error?