Designing a real-time analytics pipeline for high-throughput event data: batch vs. streaming approaches?

Hey everyone,I'm looking for some insights on designing a real-time analytics pipeline. We're dealing with a high volume of event data (think millions of events per second) and need to perform near real-time aggregations and analysis. My main question revolves around the fundamental architectural choice: should we lean into a purely streaming approach (Kafka Streams, Flink, Spark Streaming) or is there a compelling case for a hybrid batch-streaming model (Lambda/Kappa architecture)?I'm particularly interested in the operational complexities, cost implications, and latency trade-offs associated with each. For instance, with a pure streaming solution, how do you handle late-arriving data or re-processing historical data without significant complexity? Conversely, does a batch layer in a hybrid approach introduce too much latency for what we consider "real-time"?Any practical experiences or recommendations, especially around specific technologies and data consistency challenges, would be greatly appreciated!