Matching and dispatch: test the timing, not just the outcome
Uber's own developer documentation is specific about one timing detail that matters for testing: after a trip is created, the driver's status needs to update quickly — within seconds — or the flow can hit a timeout error, which is why its own guidance is to trigger the driver-accept state change immediately after trip creation rather than leaving a gap. Separately, an engineering write-up on ride-sharing matching design sets a concrete target worth borrowing for any dispatch system: a p95 of around 5 seconds to assign a provider, described as roughly the point where the experience starts to feel broken to the person waiting. Both point to the same test case: measure status-propagation and assignment latency directly, under load, rather than only checking that a match eventually happens.
Driver-status timing via Uber's Guest Rides API build guide FAQ; the p95 assignment-latency target via a ride-sharing matching-engine design write-up.
Double-assignment is a concurrency bug, and it needs a concurrency test
A matching engine that tracks provider status as a simple flag is vulnerable to two near-simultaneous requests both reading "available" and both assigning the same provider before either write lands. The fix pattern described in the same matching-engine write-up is a conditional update — the assignment only succeeds if the provider's row is still in the expected state when the write happens, and a failed conditional update means the system retries with the next candidate instead of double-booking. The test case this implies is deliberate: fire two assignment requests for the same provider at close to the same instant and confirm exactly one wins, rather than trusting that it's statistically unlikely to happen in production.
Conditional-update assignment logic via the same ride-sharing matching-engine design write-up.
GPS accuracy depends on the device, not just the signal
Location-testing guides are consistent on one point: GPS and location accuracy vary by device, chipset, and OS version, and that variation is worse in dense urban environments — exactly where a lot of on-demand demand concentrates. That's the practical argument for testing location-dependent flows on a representative set of real devices in real conditions, not only against a simulated or mocked location feed, and for including a stale-or-missing-fix case in the test plan rather than assuming a fix is always current.
Device- and environment-dependent GPS accuracy via a last-mile delivery QA write-up on GPS drift and connectivity issues.
Push notifications fail silently, and the cause is often the device, not the app
A background push notification that never arrives is a recurring, well-documented failure mode on Android, and it's frequently caused by the device's own battery-management behavior rather than a bug in the app. dontkillmyapp.com, a community-run tracker of per-manufacturer background restrictions, documents OnePlus's "App Auto-Launch" feature as one that can stop an app from running in the background entirely unless it's explicitly exempted, and a separate sleep-tracking feature on some OnePlus devices that disables network connections during a learned "asleep" window — which blocks notification delivery outright during that window. The test implication: run the notification suite against backgrounded and battery-saver device states on a representative OEM set, not only against a foregrounded app with the screen on.
OnePlus background-restriction behavior via dontkillmyapp.com's OnePlus page.
Office Wi-Fi is the one network condition least like real use
Mobile-testing guidance keeps repeating the same correction for a reason: testing exclusively on a strong office Wi-Fi connection is one of the most common gaps in a mobile test plan, because it's also the condition real users hit least often on an on-demand app used on the move. Standard guidance is to deliberately cover switching between Wi-Fi and mobile data mid-session, throttled 2G/3G conditions, and packet loss, using network-shaping tools built for exactly this, and to build the device matrix from actual usage analytics rather than a theoretical list of OS versions.
Network-condition and device-matrix guidance via BrowserStack's mobile app testing checklist.
Peak load is tested as a recurring drill, not a one-time check
Uber's own engineering blog describes running Hailstorm, an internal load-testing tool that uses test accounts to simulate extra traffic on the system, as part of weekly system-wide drills — alongside individual teams independently load-testing their own services on a self-service platform. The pattern worth borrowing at any scale is the cadence: treat peak-readiness as something re-verified on a schedule as the system changes, not a box checked once before the first launch.
Weekly load-testing drills via Uber's own engineering blog on New Year's Eve traffic preparation.