How to Debug a Crash That Only Happens Online

A crash that only happens online is the kind that eats days, because the normal debugging loop breaks down: you cannot reliably make it happen, so you cannot watch it fail. The reason is almost always the same — timing, dropped packets, or state that desyncs under real latency. This guide is about recovering the conditions from a real occurrence so you can debug a crash that only happens online on demand: capture the failure with the network state on each client to find where it diverges.

Why a crash that only happens online is hard

The difficulty with a crash that only happens online is that timing, dropped packets, or state that desyncs under real latency. It is not actually random; it is deterministic given the right conditions. The problem is that you do not have those conditions in front of you, so on your machine the failure simply refuses to appear when you are watching.

This is why trying harder by hand rarely works. You can replay the game your way a hundred times and never line up the exact circumstances. What you need is not more attempts but the actual conditions of a real occurrence, captured the moment it happened.

Connecting failures to the build that caused them

Regressions are the cruelest class of bug because they punish your most engaged players — the ones who already own the game and updated to your newest patch. A change meant to improve things quietly breaks something else, and without build-level tracking you have no way to link the dip in retention to the release that caused it.

The fix is to attach a build identifier to every captured failure. Then a new signature that appears the day you ship a patch is unmistakable, and you can roll back or hotfix while only a few players are affected instead of discovering the problem weeks later in your reviews.

What good context actually looks like

The difference between a bug you fix in five minutes and one you chase for a week is almost always context. A bare error message tells you something went wrong; a useful report tells you where, on what, after what sequence of actions, in which build. Stack trace, device model, OS version, available memory, and the breadcrumb trail of recent events are the fields that turn guessing into reading.

When that context is captured automatically and consistently, reproduction stops being the bottleneck. You can often see the cause directly in the trace, and when you cannot, the breadcrumbs show you the exact path to walk to reproduce it yourself.

Turning a pile of crashes into a ranked worklist

Raw crash data is overwhelming if every occurrence is its own line. The trick is grouping: identical failures, fingerprinted by their stack trace, collapse into one issue with a count. Suddenly the question “what should I fix first?” answers itself, because the bug hitting the most players sits at the top with the biggest number next to it.

That ordering is what makes a small team effective. You are never going to fix everything, but you do not have to. Fixing the top few signatures usually removes the large majority of real-world failures, and prioritising by frequency means your limited hours always go to the bug that matters most right now.

Why the report you get is never the whole story

When a player does take the time to tell you something broke, the message is almost always thin: “it crashed,” maybe a screenshot, rarely a version number, and almost never the exact steps. You are left reconstructing the scene of an accident from a single blurry photo. The information you actually need to fix the bug — the stack trace, the device, the build, the state the game was in — is precisely what a human report leaves out.

That is why working from manual reports alone keeps you slow. Every ticket becomes a back-and-forth interrogation, and half the time the player has moved on before you get an answer. Automatic capture removes the interrogation entirely, because the context travels with the failure the instant it happens.

Capturing it and debugging on demand

The practical method is to capture the failure with the network state on each client to find where it diverges. With the failure captured — its stack trace, the device and OS, the build, and the breadcrumb trail of events just before it — a crash that only happens online stops being a ghost. The breadcrumbs record the path in, the trace points at the failing line, and the device and build narrow the conditions.

Collect a few occurrences and it gets easier still, because the conditions they share isolate exactly what matters. Once you can trigger it on demand, it is an ordinary bug: fix the root, tie failures to builds, and confirm the signature disappears in the next release.

This is where a tool like Bugnet earns its place. Its SDK captures every failure automatically with the full stack trace plus device, OS, memory, build, and game-state context, folds identical failures into one grouped issue with an occurrence count, and ties each to the build it happened on. The result is that the abstract idea above stops being theory and becomes a ranked list you work down — the worst problem first, verified fixed when its signature disappears from the next release.

Most of the failures hurting your game are silent. The first job is making them visible; the fixes get a lot easier after that.