The release-gate dashboard in Production Crash Analytics replaced a weekly argument about crash-free sessions. We now walk into launch review with three named clusters, an owner, and a hold-or-ship line. Our PM still dislikes how long symbolication takes, but at least we stopped treating the delay as “the app got worse overnight.”
Reviews
What changed after the labs, not how people felt in week one.
We ask alumni for a specific artefact or decision, not a star average. A few voices below still disagree with parts of the method. That is expected; crash work is argumentative by nature.
ANR triage helped, with a caveat: the programme assumes a crash vendor is already emitting main-thread traces. Ours was not. I lost the first week to wiring, and the pairing labs only made sense after that. Once the export existed, the freeze-vs-death distinction finally stuck.
We sent two product managers to Stability SLOs. They came back able to ignore half the charts we used to paste into Slack during store-review weeks. The session on traffic-mix bias was the one that mattered; festival traffic in Thailand had been masquerading as a quality drop for two years.
★★★★☆
iOS symbolication workshop, June 2025 cohort. dSYM checklist was dry and exactly what we needed. Clustering heuristic felt opinionated; we still merge more aggressively than Arthit prefers. Would sit the flagship next, not another iOS-only day.
Short note: the Observatory critique night is better than the public cohort if you already know the vocabulary. We brought one nasty Kotlin crash that looked like three groups. Nalinee split it on a user-journey axis we had never used. No slides.
Case study 01
Holding a staged rollout without freezing the whole train
A Bangkok e-commerce app entered a payday weekend with a new checkout SDK. Crash-free sessions dipped 0.9 points on Android 13 devices only. The team’s instinct was a full halt. After a Release Room desk, they isolated a single cluster: a null currency formatter when the SDK returned an empty region code. They paused 13+ devices in staged rollout, left other tracks open, and shipped a hotfix in 11 hours. Store vitals recovered before Monday traffic.
The lasting change was not the hotfix. It was the one-page brief: cluster name, affected journey, hold condition, reopen condition. That template now sits in their launch checklist.
Case study 02
When symbolication lag invented a “new” crash
An iOS team in Singapore thought a 2.4 release had introduced a never-before-seen crash in a map SDK. Unsymbolicated frames made it look novel. After the symbolication programme, they rebuilt dSYMs for the previous three versions and found the same stack, mis-grouped, dating back four months. Priority dropped from “page now” to “fix in the next map SDK bump.” They also added a rule: no P0 on unsymbolicated-only groups during the first 36 hours after a store push.
They still argue about sampling rates. The case did not make them love their vendor. It made them stop paging on ghosts.