Running a COUNT DISTINCT query on massive datasets can be a nightmare for data engineers. The author reflects on their own experience with a simple task—counting unique users—turning into a three-hour ordeal. Traditional methods like GROUP BY require keeping track of every unique value, which is impractical at scale. With billions of events daily, exact answers become resource-intensive and time-consuming, often leading to frustration as queries run for hours or even days. The piece introduces probabilistic data structures known as data sketches as a solution. These algorithms trade precision for speed, allowing for quick approximate answers with small memory footprints. Instead of storing all unique values, sketches summarize data in a compact form, making them ideal for streaming data where reprocessing isn’t feasible. The concept of sketching, which has roots in streaming algorithms, emerged from the work of Philippe Flajolet in the 1980s. His research paved the way for techniques that keep a small subset of data while still providing reliable estimates. Data sketches work by hashing incoming values into uniform random numbers, discarding most of the data while retaining important statistical information. For example, cardinality sketches like HyperLogLog allow for efficient COUNT DISTINCT calculations without the overhead of full dataset scans. Many modern data systems, including Spark and BigQuery, incorporate these sketch-based functions, enabling users to leverage their benefits without additional libraries. The author emphasizes that understanding and utilizing these sketches can drastically improve query performance and resource management in data engineering tasks.