Driver Dispatch Under Timeouts

A search problem with a clock on it — expanding rings, batched offers, and a retry loop for when the answer is 'not yet' rather than 'no'.

Rider App → Dispatch API → MySQL

The request opens a stream

A ride request is created and acknowledged immediately, but the response isn't closed. Dispatch can legitimately take minutes, so the connection stays open and the app is told what actually happens — accepted, queued, or exhausted. The alternative is a single reply and a client left polling an operation whose duration nobody can predict in advance.

Why not the alternatives?

Broadcast to every driver in the city: Fastest possible assignment and a notification storm for a ride one person can take. It also destroys proximity ordering — a driver twenty minutes away competes with one around the corner, and riders wait longer for worse pickups.
Offer to one driver at a time: No wasted notifications, and an unassigned rider watching a spinner while the system works through unresponsive drivers one acceptance window at a time. The batch is the compromise: bounded waste, bounded wait.
Query once, then work through the list: Cheaper in database terms and increasingly wrong. A sweep lasts long enough for drivers to move, go offline, or take other rides, so dispatching from a snapshot means dispatching to where drivers used to be.
Treat an exhausted sweep as failure: Simple, and it discards the distinction the system exists to preserve. Nothing about a failed sweep says the next one fails too, and availability turns over on a timescale shorter than a rider's patience.
Let the queue's automatic retry handle re-attempts: Automatic retry replays a job that errored. A dispatch re-attempt isn't an error — the previous sweep completed correctly and found nobody — and it must re-validate that the request still exists before acting. Conflating the two means retrying rides that were already cancelled.

Lessons

'Not found' and 'not yet' are different answers. Most of the value in this pipeline came from refusing to collapse them into one.
Freshness beats efficiency when the underlying data is in motion — re-querying every ring looked wasteful and was the difference between dispatching to drivers and dispatching to their last known positions.
An operation that legitimately takes minutes isn't a slow request; it's a process with progress. Give it a channel to say so, or the client invents polling and guesswork.
The failure path holds the money. Releasing funds correctly across every exit is harder, and matters more, than the assignment logic everyone focuses on.
Logic that runs in two execution contexts will drift apart. Shared behaviour has to be shared code — a copy with a comment explaining that it's a copy is a bug with a delay on it.