How DoorDash Optimized Item Availability at Scale Using Elasticsearch

DoorDash's homepage item carousels needed to filter millions of items by availability in under 300ms. We couldn't call the menu service at request time (too much fan-out, too slow), so we indexed availability directly in Elasticsearch. We went through three schema iterations: nested documents (600ms), Gojek-style encoded time slots as terms (350ms but 6x storage), and finally range fields backed by BKD trees (250ms, baseline storage). The range approach won on both latency and storage.