Iterable Mobile SDK.

Deep links were failing intermittently on iOS — users tapped an email or SMS link and the app opened to the wrong place, or not at all. The maddening part: it was worse on good connections.

The SDK follows a link's redirect just far enough to capture attribution, then cancels it. Cancelling makes the network layer report an error — and the old code checked for that error first, bailing out and throwing away the destination URL it had already successfully captured a moment earlier. On a fast network the request finished quickly enough to trip the cancellation error every time; on a slow one it slipped through. Better connectivity, more failures.

The fix was to reverse the order: use the data if we have it, and only treat the error as fatal if we genuinely came back empty-handed. A handful of lines.

Their Android apps were generating ten times the request volume of their iOS apps, straining infrastructure. The culprit was a timer. Android's auth manager used a timer that runs on its own thread, independent of the app lifecycle — so it kept firing token refreshes even while the app was backgrounded. iOS timers are tied to the run loop and naturally pause when the app suspends.

I made the Android timer lifecycle-aware: clear it on background, re-evaluate on foreground. Ninety-five lines, and the platforms finally behaved the same way. The same customer's app-launch ANRs traced back to the same area — heavy crypto and storage work blocking the main thread at startup — which I fixed by moving initialization off the main thread entirely.

A customer's mobile analytics broke: millions of sessions logging zero seconds, push-open numbers inflated. It took the team weeks to find, partly because nobody could reproduce it internally — and partly because the root cause was intended behavior nobody had documented. "Silent" wake-up pushes were being counted as real opens. I owned the workstream to identify and correct the erroneously-tracked events and clean up the customer's data.

The deeper lesson stuck: the incident was a direct argument for building a real end-to-end test harness and real observability, so the next regression would be caught in CI, not in a customer's dashboard. Both became projects I led — and they're in the next section.