What is Apache Druid?

The Problem It Solves: Traditional data warehouses like Snowflake batch load data every few minutes or hours. They're too slow for real time analytics. Streaming systems like Kafka handle ingest well but struggle with complex aggregations. You need something that can answer questions like "show me website clicks by country in the last 15 minutes" with subsecond latency while millions of new events pour in every second. This is the exact gap Druid fills. Core Architecture: Druid models data as fact tables of events, always with a timestamp and dimensions (attributes like country, device type) plus metrics (numbers like click count, revenue). It ingests streams continuously from sources like Kafka, builds columnar indexes in memory, and periodically writes immutable compressed segments to cheap storage like S3. Here's what makes it different: queries run on these columnar segments, not directly on the stream. This separation allows you to scan billions of rows with subsecond latency. Data is partitioned by time (usually hourly or daily), so time filtered queries only touch relevant segments. Common Use Cases: Druid shines in scenarios requiring both high ingest throughput and interactive queries. Real time fraud detection systems that analyze transaction patterns as they occur. Ad campaign monitoring dashboards that show click through rates updated every few seconds. Game telemetry where you need to aggregate player actions across millions of concurrent users. Operational business intelligence where executives need up to the minute metrics on sales or user behavior.