The domain where you have the most real code already (arete, nextvibe, gees-arise — see projects/) — so the highest-signal reps here are deepening what exists, not starting fresh. Sections 01–07 hold in any language; every project below does too.

Difficulty: 🟢 starter (hours–days) · 🟡 intermediate (a week or two) · 🔴 ambitious / portfolio-grade. ⭐ = highest signal.

Every entry has a done when. If you can’t state when a project is finished, you’ll either stop early or never stop.

The ladder

  • 🟢 Add idempotency to an endpoint you already wrote — an Idempotency-Key header, a store of key→response, and correct behaviour on retry. Done when: the same request sent twice creates one record and returns the same response body both times. Exercises: API design, cross-cutting.

  • 🟢 Instrument one service properly — structured logs with a request ID propagated through every layer, RED metrics (rate/errors/duration), and one distributed trace end to end. Done when: you can take a slow request’s ID from a log line and see its full span breakdown. Exercises: observability.

  • 🟢 Write the failure tests you skipped — what happens on a DB timeout, a duplicate key, a malformed payload, a 500 from an upstream. Done when: every error path in one module has a test that asserts the status code and the response shape. Exercises: practices.

  • 🟡 ⭐ Rate limiter + circuit breaker, from scratch — token bucket in Redis (not a library), then a circuit breaker around one upstream call with half-open recovery. Done when: you can demonstrate the breaker opening under induced failure and recovering, with metrics showing both. Exercises: cross-cutting, building blocks.

  • 🟡 A job queue with real delivery semantics — outbox pattern, at-least-once delivery, dead-letter queue, and an idempotent consumer. Done when: you can kill -9 the worker mid-job and no message is lost or double-applied. Exercises: data and persistence, queues.

  • 🟡 Auth, implemented not imported — sessions and JWT, refresh-token rotation, and a working logout that actually revokes. Done when: you can explain why your refresh tokens rotate, and demonstrate that a stolen one is detectable. Exercises: auth, secure authentication.

  • 🟡 Migrate a monolith module to a service — pick one bounded context in a real project, extract it behind an API, and keep both running. Done when: the monolith calls it over the network and you’ve written up what got harder. Exercises: structuring, patterns.

  • 🔴 Load-test and fix your own p99 — pick a real endpoint, load it with k6 until p99 degrades, find the actual cause (N+1? pool exhaustion? GC? lock contention?), fix it, prove it with a before/after graph. Done when: you have both graphs and a written diagnosis. This is the single most interview-relevant backend rep — the backend bank is built from exactly this conversation.

  • 🔴 ⭐ Rebuild one service in a second language — take something small you wrote in Node and rewrite it in Go or Rust. Done when: both pass the same test suite and you can articulate what each language made easy and hard. Exercises: frameworks.

If you only do one

The p99 investigation. It’s the shortest path from “I built a backend” to “I can reason about a running system under load” — and it’s the conversation every backend interview eventually becomes.

↑ Back to Backend