Delivering accurate delivery promises in real time at national scale

Each promise must consider supplier, carrier, capacity, inventory, and cutoff times before the customer clicks “buy.”

Wayfair utilizes a dynamic supply chain, leveraging fulfillment from both dropship suppliers and their own CastleGate network. As a result, accurate fulfillment paths must be computed ahead of time with full supply chain context to provide customer promises that will remain consistent throughout the customer’s order journey.

Delivery routes vary by item, supplier location, customer ZIP code, carrier, cutoff times, and transportation capacity. The system must evaluate hundreds of millions of origin-to-destination permutations across hundreds of millions of product configurations, resulting in tens of millions of fulfillment plans calculated per second during peak system load. Each plan is working out the optimal supplier facility, the optimal carrier, the optimal route, and the estimated delivery date for the product.

A single late truck, broken warehouse conveyor, or carrier outage can change hundreds of delivery options. The system must react within minutes to prevent false promises and customer disappointment.

The architecture streams updates from across the network—new inventory, capacity shifts, disruptions—and identifies and recomputes only the affected promises at lightning speed.

With thousands of storefront API requests per second, delivery estimates must remain current, personalized, and responsive even as the underlying fulfillment reality changes in real time.

Supply chain events—inventory updates, truck departures, carrier capacity changes—flow through a stream-join engine based on Kafka and Aerospike. Any change triggers selective recalculation of affected delivery promises.

As events stream in, the stream-join engine consults and updates Aerospike stores to build complete joined events. This demands stronger write performance than a typical caching workload.

Wayfair groups ZIP codes into dynamic delivery zones to reduce storage and compute costs, while ensuring accuracy. The zone system balances performance with granularity.

With dozens of microservices working in parallel, any latency in Aerospike would cause queues to back up. Aerospike’s low-latency characteristics ensure smooth streaming and compute pipelines, avoiding the need of additional infrastructure overhead.

A machine learning model refines day-level delivery estimates. Aerospike enables sub-ms inference at scale by supporting the model’s real-time feature inputs.

Aerospike supports low-latency access to fulfillment and delivery promise data, even at peak scale and under heavy recomputation loads.

Wayfair handles thousands of delivery promise API calls per second and writes updates across a high-throughput pipeline involving dozens of microservices and Kafka topics.

What once took two days can now be recomputed in less than 60 minutes, restoring data integrity faster after major outages or data errors.

When a truck closes or capacity shifts, the streaming architecture adopted by Wayfair allows them to update the delivery estimates shown on the website within minutes, preventing bad customer experiences.

With access times under 1 ms, Aerospike makes it possible to support complex ML delivery models with real-time inference at customer-facing scale.