Flipkart's advertising system processes a staggering one million events per second, which creates significant financial stakes for advertisers. When users interact with ads, any delay in processing can lead to overspending or lost revenue. To tackle this challenge, Flipkart engineered a highly scalable, stateful architecture that ensures precise budget management in real time. The article details the design of this system, focusing on key processes like event-time semantics, deduplication, and watermarking. The ad data lifecycle starts when a user engages with ads, generating telemetry data that the system must process quickly to charge advertisers accurately. Flipkart categorizes this telemetry into two main types: impressions, which are billable views, and engagements, which represent direct user actions like clicks. The architecture includes a two-stage processing topology to ensure idempotency, meaning that even if an event is repeated, it only results in one charge for advertisers. Normalization is vital for this system. Incoming data is transformed into a Key-Value pair format, which includes a unique identification for each event. The article explains how Flipkart incorporates a Lambda Architecture to balance speed and completeness. The speed layer, powered by Apache Flink, handles real-time processing, while the batch layer reconciles and verifies data accuracy against immutable logs. This dual approach helps to ensure that advertisers are charged accurately without duplicates or errors.