OpenAI SWE Virtual Onsite - Full Interview Loop Experience and Questions

Question

## Round 1 - Coding

## Problem

**Format:** 5-round virtual onsite. Rounds include coding, system design, and behavioral.

**Coding Round 1:**
Given a large log file of API requests, each line formatted as `"TIMESTAMP METHOD PATH STATUS_CODE LATENCY_MS"`, parse and compute:
- P50, P90, P99 latency per endpoint
- Count of 5xx errors per endpoint per hour

**Coding Round 2:**
Design an in-memory task scheduler. Tasks have an ID, a scheduled timestamp, and a callback. Implement:
- `schedule(task_id, run_at, callback)` 
- `cancel(task_id)`
- `tick(current_time)` - fires all tasks due at or before `current_time`

**System Design:**
Design a distributed inference serving system for a large language model. Cover:
- Load balancing across GPU replicas
- Request batching for throughput
- Handling streaming responses (token-by-token)
- Model versioning and rollout

**Behavioral Focus:**
- A time you disagreed with a technical decision and how you handled it.
- How you've influenced a team to adopt a better engineering practice.

## Follow-ups

1. For the log parser: how do you compute percentiles efficiently without sorting the full dataset?
2. For the task scheduler: what data structure gives O(log n) insert and O(log n) extraction of the earliest task?
3. In the serving system: how do you handle a GPU replica that goes down mid-request?
4. How would you test the task scheduler for correctness across time zone boundaries?

OpenAI SWE Virtual Onsite - Full Interview Loop Experience and Questions

Interview Experience

Full Details

Topics

More from OpenAI