Andrei Manakov

ShareChat is one of the largest social media platforms in India, with over 180 million monthly active users. We had a high-throughput real-time stream (>200K RPS) processing using a Node.js + Redis-based deduplication with a 24-hour window. In this talk, I'll walk you through how we transitioned to an Apache Flink-based solution, the challenges we faced, and the strategies that led to a 7x cost reduction. Topics Covered: 1. State Management at Scale: - Our early attempts to structure Flink state efficiently to handle massive-scale deduplication. - Lessons learned in making the job manageable and performant despite the huge state size. 2. Autoscaling Challenges: - How we leveraged the Flink Kubernetes Operator to enable autoscaling. - Why autoscaling initially increased duplication—and how we solved it. 3. When Async API Matters in Apache Flink: - Understanding the role of Async I/O in Flink. - How it impacts performance and resource efficiency in real-time streaming. 4. How We Achieved 7x Cost Savings