Atlassian ML System Design: Design Jira Recommender

Problem overview

You are asked to design a recommendation system for Jira that surfaces relevant issues, projects, and teammates to millions of users. The system must provide personalized suggestions based on user behavior (views, comments, assignments), project metadata, and recent activity while integrating seamlessly with dashboards, notifications, and search.

What you'll be evaluated on

Start by clarifying scope: which entities to recommend (issues, projects, people), real-time vs batch freshness, and privacy constraints. Then present a two-stage architecture: candidate generation (recall) and ranking (personalized scoring). Explain data sources (issue text, metadata, activity logs, social graphs), feature engineering (recency, role, similarity, embeddings), and model choices (matrix factorization, session-based models, dense embeddings + shallow ranker). Cover online components: low-latency feature store, online scoring service, cache, and event-driven updates for real-time personalization.

Flow of the interview

You’ll typically walk through: requirements and tradeoffs; high-level architecture; ML pipelines (offline training, online serving); latency, scalability, and fault-tolerance strategies; and evaluation (offline metrics, A/B testing, feedback loop). Expect questions that probe cold-start, diversity, bias mitigation, and maintainability.

Skill signals to demonstrate

You should show knowledge of recommender systems (candidate generation vs ranking), embedding methods, feature stores, streaming/event-driven pipelines, real-time inference, A/B testing, evaluation metrics (precision@k, NDCG), and operational concerns like instrumentation, rollback, and model lifecycle.