Lyft System Design Interview Questions

Combine distance, ETA, driver quality, surge zone, and driver intent (online status, destination filters) into a weighted score. Assign via a fair-queue dispatcher with caps to avoid overloading any driver and fallbacks if a driver declines or times out.

Define states such as requested, dispatched, accepted, en route to pickup, arrived, in trip, completed, canceled. Persist transitions as append-only events, enforce invariants (one active trip per driver), and emit updates to billing, safety, and notifications.

Use time-boxed holds on drivers, explicit cancel reasons, and automatic reassignment rules. Charge or waive fees based on policy, and publish events to reconcile payouts, adjust ETAs, and update marketplace metrics.

Train models on historical trips and real-time traffic. Serve low-latency predictions via a features cache (road segment speeds, time of day, weather) and update ETAs continuously as location and traffic change.

Ingest frequent GPS pings through a throttled gateway, snap to road, dedupe, and write to a time-series store plus an in-memory spatial index for dispatch. Downsample for analytics; keep high-rate streams for safety and ETA.

Either works; choose based on library and team familiarity. Geohash is simple for proximity buckets; quadtrees provide flexible partitioning in uneven density. Normalize on one, and add a small radius expansion to avoid boundary misses.

Use a hybrid. Rely on a third-party for global coverage and freshness, but keep a thin routing cache and fallbacks. Build internal abstractions so dispatch, ETA, and pricing are decoupled from any single provider.

Push CDN-cached tiles to clients, prefetch common zoom levels for hot cities, and respect provider quotas. Add backoff and local caches to ride-sharing clients to reduce duplicate fetches during interactive map use.

Apply Kalman or particle filters on the client and server, then project points to the most likely road segment. Detect teleports, clamp speed spikes, and flag inconsistent updates before they influence dispatch or ETA.

Use WebSockets for bi-directional control (acknowledgments, pings, cancellations). SSE fits one-way streams but offers less control. Either way, implement heartbeats, resumable streams, and exponential backoff.

Publish normalized events (location_update, dispatch_assigned, trip_state_changed) to a durable log. Consumers (ETA, pricing, safety, notifications) subscribe with their own offsets, enabling replay and backfills.

Enforce per-device rate caps, coalesce updates, and drop non-critical signals when saturated. In the backend, use bounded queues, shed lowest-priority topics first, and protect critical paths like dispatch and safety.

Maintain per-user budgets and collapse keys, dedupe server-side, and fall back to in-app banners if push fails. Record delivery receipts and retry on alternate channels for time-sensitive events.

Continuously estimate supply and demand per geo-cell, compute a multiplier with caps and fairness rules, and expose changes atomically. A/B test parameters, and throttle updates to avoid rider confusion.

Place dispatch partitions per region with local writes, replicate key metadata, and route riders to the nearest healthy region. Use idempotent assignment tokens and conflict resolution for cross-region failover.

Queue actions locally with monotonic timestamps, sync opportunistically, and reconcile conflicts server-side. Keep a minimal on-device map and last known assignments to navigate short outages.

Use adaptive sampling tied to motion activity, reduce frequency when stationary, and batch transmissions. Switch to lower-power sensors when possible and escalate only during an active trip or imminent dispatch.

Sign deep links with short-lived tokens, restore app state (trip ID, route, chat), and prefetch current trip data. Handle expired links with a safe fallback to the live trip screen.

Cache immutable assets (vehicle icons, strings) aggressively, store recent trip state with version tags, and reconcile via ETags or vector clocks. Prefer delta sync over full refreshes to save bandwidth.

Ask for p95/p99 latencies, dispatch acceptance targets, location update cadence, failover expectations, and data retention rules. Restate them back and use those numbers to justify design choices, capacity estimates, and degradation paths.