Built from a real interview (August 2026) that went badly — every question in these two files was actually asked. That makes this the most useful bank in the vault, because it’s not a guess at what gets asked.

Files

  1. Production Debugging — p99 spikes, per-AZ metrics, AZ impairment vs bad deploy, N+1 regressions, retry safety, retry storms
  2. Node Runtime & API — Promise.all at scale, Buffer allocation, the event loop, streams, error handling

What that interview was actually testing

Not Node trivia. Production debugging judgment — can you narrow a search space instead of guessing? Look at the shape of it:

  • p99 spikes → what do you check? — bisect a latency distribution
  • AZ impaired or recent deploy? — bisect by metric dimension
  • N+1 caused a regression — bisect to a code change
  • retry safety, retry storms — don’t make it worse while you’re diagnosing
  • Promise.all with 200 requests — do you understand what your own code does to a downstream
  • new Buffer vs Buffer.alloc — do you know why an API was deprecated, not just that it was

That’s Rank II — the Diagnostician, almost item for item. It’s the rank the path calls “unglamorous and the one that separates people,” and it’s the one to drill.

The three highest-leverage things to fix

  1. Instrument event loop delay in a real service and watch it. Node’s signature failure — p99 spikes, flat p50, low CPU — is invisible until you measure this, and it’s the answer to the first question they asked.
  2. Tag metrics with az and version. That one design decision turns “is it the AZ or the deploy?” from an investigation into a dashboard query.
  3. Know the retry trio cold: full jitter, retry budgets, and retry at one layer only. Amplification is the thing that turns a blip into an outage.
↑ Back to Backend