My Interview Experience for Principal Engineer - Interviews in AI Era!

Summary

Piyush Garg describes a recent, unnamed interview for a senior engineering role and argues that AI has changed how coding interviews work. Rather than focusing mainly on how quickly candidates can type, the interview tested their engineering judgment, technical knowledge, and ability to use AI in a controlled way.

Interview Format and Examples

The interview took place in a cloud IDE containing a production-style TypeScript application, with an AI assistant available. The interviewer presented written tickets describing bugs or customer problems.

Rate-Limiting Bug

A ticket reported that a limit was being exceeded during traffic spikes. Instead of immediately asking AI to fix it, Garg traced the relevant service and its use in the codebase. He identified the issue as a fixed-window implementation where a sliding-window approach was needed. After outlining the intended solution in a comment or pseudocode, he used Copilot to generate the implementation.

Duplicate Queue Entries

Another ticket concerned duplicate entries in a database-backed queue. Garg reasoned that simultaneous requests could bypass an application-level check, making this a concurrency problem. He proposed adding a database-level unique index on the user ID and handling the resulting error during enqueueing.

The follow-up discussion covered the costs of indexes and how databases enforce uniqueness.

Database Theory and Experience

The interview also explored deadlocks, including their causes and types, a real-world example, and a short exercise to write code that could produce a deadlock. AI was not allowed for this exercise.

Distributed Rate Limiting

For horizontally scaled servers, the discussion turned to using a shared store such as Redis. The interviewer asked what should happen if Redis failed. Options included taking down a critical service or temporarily falling back to the database.

Garg emphasizes weighing the risk of overloading the database and triggering cascading failures, and making the fallback policy explicit.

Main Takeaway and Preparation Advice

AI can generate code, but it cannot replace a candidate’s understanding of the system and its trade-offs. Strong candidates should investigate the code themselves, explain their reasoning, identify likely causes, give AI specific technical requirements, and review the generated result. They should also be prepared to discuss consequences, not just propose an implementation.

Preparation should include practical knowledge of concurrency, databases and indexes, deadlocks, rate-limiting algorithms, distributed systems, and failure handling. These interviews may involve less manual typing, but Garg says the expected depth and engineering bar are higher.

Main Speaker

• Piyush Garg, sharing and analyzing his own interview experience.