More telemetry is not more control
Mobile teams either drown in unsampled logs nobody reads, or fly blind when crash-free sessions move. The vocabulary is muddled: action log, breadcrumb, metric, alert — four words for overlapping pipes. The fix is not another SDK. It is choosing signals that drive decisions and owning who pages when they move.
ANR forensics deserve their own deep dive — see ANR traces lie by omission. Client↔server correlation is covered in telemetry across client and server. This post is the stack map: what to collect, with which industry tools, without repeating those essays.
Related reading
- Internal: ANR / Crashlytics
- Internal: Telemetry across client and server
- Docs: Play Vitals, Firebase Crashlytics
Three pipes (do not merge them in your head)
| Pipe | Answers | Examples |
|---|---|---|
| Crashes / ANRs / non-fatals | What broke, how often | Stack + breadcrumbs + custom keys |
| Action / product logs | What the user tried | open thread, send, attach (privacy-reviewed) |
| Metrics | Is the system healthy | counters, timers, histograms (cold start, sync p95) |
Play Vitals and vendor dashboards will not match one-to-one. That is normal — collection windows differ. Design for decisions (page on crash-free drop, p95 sync, send success), not vanity event counts.
Tools in the wild
- Firebase Crashlytics — crashes, non-fatals, ANRs (with constraints called out in the ANR post); breadcrumbs when Analytics sharing is configured.
- Datadog — RUM, logs, metrics, APM correlation; strong when you already live in that ecosystem.
- Sentry / Bugsnag — error-centric; watch PII and sampling defaults.
- Perfetto / systrace — local truth for jank; not a production pager.
- Play Console vitals — store-facing rates that affect discoverability.
- Custom — Kafka → warehouse for product analytics at scale.
Shared concerns everywhere: sampling, PII redaction, offline batching, and correlation IDs from APK to backend (see the telemetry post).
A minimal ownership model
- Pick three SLIs the team will defend (e.g. crash-free users, p95 inbox open, send success).
- Map each SLI to one primary tool and one human owner.
- Alert on rate of change and absolute thresholds — not on every log line.
- Keep action logs privacy-reviewed; treat them as production data, not debug leftovers.
If everything is a “log,” nothing is an alert. If everything is an alert, nobody pages.
Sampling without lying to yourself
Log everything in debug builds; sample in production. Crash and ANR pipelines usually want high capture; verbose action logs need aggressive sampling or they become a storage and privacy problem. Metrics should be cheap enough to be continuous. The mistake is one global “log level” that either blinds you or bankrupts the pipeline — set policy per pipe.
Wrap-up
Instrument for decisions: separate crash, action, and metric pipes; pick tools that match how you already operate; sample and redact on purpose. Pair Crashlytics/ANR deep dives and cross-tier correlation posts with this map — do not paste every stack into a log pipeline and call it observability.