System Design: Elevator Operation Logic

Onsite system design: design the operation logic for a building's elevator system, scaling from single-car scheduling to multi-car coordination and tall-building zone partitioning (sky lobby). Requirements Single-car scheduling: model up / down requests, current car position and direction, and pick the next stop. Multi-car coordination: assign incoming requests to one of N cars to minimize wait time. Tall building follow-up: partition into low / mid / high zones with an inter-zone sky lobby; design the assignment logic so a passenger in zone A bound for zone C is routed via the sky lobby. Notes Single car. Three classic algorithms with explicit trade-offs: FCFS: each request is served in arrival order. Simple, awful average wait time. SCAN / 'elevator algorithm': maintain a direction; service all requests in the current direction before reversing. Strong fairness, no starvation, used as the default. LOOK: variant of SCAN that reverses as soon as no further requests exist in the current direction. Slightly better average wait, same fairness. Multi-car assignment. Compute a per-car cost for each request (function of current direction, distance to the requested floor, queued stops). Assign the request to argmin car. Tie-break on least-loaded car for load balancing. Real systems re-evaluate assignments periodically (every 1–5 seconds) because new requests change the optimum. Modeling requests. Hall calls (request at a floor, direction known) vs car calls (destination floor pressed inside a car). The optimizer treats these differently: hall calls can be reassigned among cars; car calls are committed. Tall-building zoning. Partition floors into zones with dedicated express cars serving zone boundaries. The sky lobby acts as a transfer floor: 'zone A → sky lobby → zone C' becomes two scheduled hops. Surface the trade-offs: zone partitioning loses flexibility (a car serving zone A can't help zone B), but reduces average travel distance and crowd density per car. Edge cases. Fire / emergency mode (override to ground floor), VIP mode (commandeer a car), overload sensors (refuse new pickups), maintenance lockout. Mention without diving deep. Performance metrics. Average wait time, longest wait time (fairness), average ride time, throughput at peak hour, energy consumption. Industry benchmarks: peak-hour up-traffic of 12% of building population in 5 minutes is a standard target. Tie-back to leveling at Waymo: the round is graded on whether the candidate drives the algorithm choice with justification and only zooms into multi-car coordination once the single-car baseline is on the board. Preparation Memorize the SCAN / LOOK / FCFS taxonomy and be able to recite the average-wait-time argument for each. Pre-stage the multi-car cost function: cost(car, request) = distance + direction_penalty + queue_length. For the sky lobby follow-up, sketch a 30-floor building partition (e.g. 1–10 / 10–20 / 20–30) before the round and rehearse why the transfer floor sits at 10 and 20 specifically. Reserve 5 minutes for monitoring / observability — call out wait-time SLIs and the dashboards an operator would watch.