Collectors.groupingBy(keyExtractor) walks the stream once and bins each element under the key returned by the extractor. The default downstream collector is toList, so the value type is List<T>. The result is a HashMap, which means key iteration order is not stable: pass groupingBy(key, LinkedHashMap::new, toList()) if you need insertion order. On JDK 16+ the Order class collapses to record Order(String customer, String product, int qty) {}.

The two-arg form of groupingBy accepts a downstream collector that determines what the values look like in the result map. counting() gives Map<K, Long> (handy for histograms), summingInt(...) produces totals, and mapping(extractor, downstream) lets you transform each element before the downstream collector consumes it. Chain mapping(...) with toSet() to dedupe the values per group. These compositions replace what would otherwise be a manual loop with a computeIfAbsent + add dance.

Pass a second groupingBy as the downstream collector to bucket each group again by another key, yielding Map<K1, Map<K2, V>>. Here we sum quantities per (customer, product) pair, mirroring GROUP BY customer, product in SQL. The pattern composes to any depth, but readability degrades fast past two levels: at three or more, switch to a flat key (a class holding both customer and product) and group once. This is the cleanest way to compute pivot tables and category breakdowns without writing nested loops.