Java performance audit

Java Android Runtime Performance Audit

Instructions

Performance work is measurement-first. Do not optimize without a trace.

1. Establish a Baseline

Measure these before changing any code:

• Startup time — adb shell am start -W ... → TotalTime, WaitTime.
• Frame rate — Profiler → CPU → System Trace, or Macrobenchmark library.
• Memory — Profiler → Memory → record allocations.
• Network — Profiler → Network timeline.

Save traces in docs/perf/ with a date; every future change compares against the baseline.

2. StrictMode — catch main-thread IO immediately

Enable StrictMode in debug builds only:

public class App extends Application {
    @Override
    public void onCreate() {
        super.onCreate();
        if (BuildConfig.DEBUG) {
            StrictMode.setThreadPolicy(new StrictMode.ThreadPolicy.Builder()
                    .detectDiskReads()
                    .detectDiskWrites()
                    .detectNetwork()
                    .penaltyLog()
                    .penaltyFlashScreen()
                    .build());
            StrictMode.setVmPolicy(new StrictMode.VmPolicy.Builder()
                    .detectLeakedClosableObjects()
                    .detectLeakedSqlLiteObjects()
                    .detectActivityLeaks()
                    .penaltyLog()
                    .build());
        }
    }
}

Treat every StrictMode log line as a bug to file.

3. Layout Inspector & Hierarchy

• Aim for depth < 8 from the Activity's root. Anything deeper pays for a measure pass per level.
• Flatten nested LinearLayouts into ConstraintLayout (see android-views-xml skill).
• RelativeLayout does two measure passes — prefer ConstraintLayout.
• Remove android:background set to the same color as the parent; it causes overdraw.

Enable Debug GPU Overdraw (developer options) — any red regions on a scrollable list are the main cost to fix.

4. RecyclerView

• Use DiffUtil / ListAdapter; never call notifyDataSetChanged() for content changes.
• Set setHasStableIds(true) + getItemId(position) when IDs are stable.
• Use view pools across RecyclerViews showing the same cell type.
• Set fixed sizes on ImageViews so Glide can downsample (see glide-images skill).
• Binding should do zero allocations on the hot path — move formatters to fields, pre-compute contentDescription.

5. Startup

• Move work out of Application.onCreate — use Jetpack App Startup (androidx.startup) to lazy-initialize libraries.
• Use Baseline Profiles (androidx.benchmark:benchmark-macro-junit4) to generate a profile for R8 to pre-JIT critical paths.
• Defer WorkManager enqueues behind DelayableInitializer.

public class CrashReporterInitializer implements Initializer<Unit> {
    @NonNull
    @Override public Unit create(@NonNull Context context) {
        CrashReporter.init(context);
        return Unit.INSTANCE;
    }
    @NonNull
    @Override public List<Class<? extends Initializer<?>>> dependencies() {
        return List.of();
    }
}

6. Memory

• Profiler → Memory → "Dump Java heap" → filter by Activity. If an Activity instance persists after finish(), you have a leak.
• Use LeakCanary (com.squareup.leakcanary:leakcanary-android) in debugImplementation.
• Common leaks:
  • Inner classes holding implicit outer reference (Fragment listeners, Handlers).
  • Singletons that cache Activity or View.
  • Forgotten BroadcastReceiver / Observer registrations.

7. ANRs

An ANR is any unresponsive main-thread window > 5s. Diagnose with:

• adb bugreport → traces.txt.
• Firebase Crashlytics ANR collection.
• Google Play vitals.

Root causes map 1:1 to what StrictMode catches: disk IO, network, large layouts, broadcast processing on main thread.

8. Background Work

• Use JobScheduler/WorkManager instead of BOOT_COMPLETED receivers.
• Coalesce network calls; prefer OkHttp connection pool reuse (one OkHttpClient per app).
• Batch Room writes in a transaction: db.runInTransaction(() -> { ... }).

9. Profiling Cadence

• Run Systrace on release builds (AOT-compiled) — debug builds with R8 off are not representative.
• Test on a mid-tier device (Pixel 6a, Samsung A-series), not a flagship.
• Record a 30-second trace of the primary user journey each release.

Checklist

• StrictMode is enabled in debug and produces zero violations during the smoke journey.
• Profile GPU Rendering shows no frames above the 16ms line on the main list screen.
• View hierarchy depth is under 8 for every primary screen.
• notifyDataSetChanged() does not appear in production code.
• LeakCanary is integrated in debug and the primary journey produces no leaks.
• Cold startup TotalTime is tracked release-over-release and trending flat or down.
• App Startup library is used; heavy libraries initialize lazily.