android-trace-sections

Android Trace Sections — Perfetto-Backed Code Probes

Why trace sections beat logging for performance

Log.d answers "did this run." Trace sections answer "did this run, on what thread, when, and how long" — with frame boundaries and other concurrent work in the same view. For any performance investigation that's the difference between guessing and knowing.

Question	Logging answers	Tracing answers
"Did this code run?"	Yes	Yes
"On what thread?"	Only via Thread.currentThread().name	Yes — visualized per-thread
"How long did it take?"	Only with manual timestamps	Microsecond-precise duration
"Did it block a frame?"	No	Yes — frame markers in the same view
"What ran concurrently?"	No	Yes — every thread visible at once

When to use

• "App janks during scroll" — wrap suspect adapter / Composable / RecyclerView paths
• "Cold start is slow" — wrap Application.onCreate, Activity.onCreate, first setContent
• "Is this expensive call on the main thread?" — wrap and read the lane it lands in
• "Does this finish within the frame budget?" — 16.6 ms (60 Hz) or 8.3 ms (120 Hz)
• "Why is this withContext(Dispatchers.IO) block holding things up?" — wrap and see if it's actually on IO

When NOT to use

• "Did this run at all?" — android-probe-logging is simpler
• The bug is a crash or exception — read logcat
• You need to inspect specific values — tracing supports key/value but logging is more flexible

Pre-flight: detect what your project supports

# 1. androidx.tracing on the classpath? (preferred for the trace { } lambda)
grep -r 'androidx.tracing' app/build.gradle* gradle/libs.versions.toml 2>/dev/null

# 2. minSdk — Trace.beginSection works on all APIs; Trace.beginAsyncSection
#    requires API 29+
grep -E 'minSdk' app/build.gradle* gradle/libs.versions.toml 2>/dev/null

# 3. Build is debuggable (Trace sections fire only in debuggable apps unless
#    the device is rooted)
grep -A2 'buildTypes' app/build.gradle* | grep -i 'debuggable'

No androidx.tracing? Two options:

• Add implementation("androidx.tracing:tracing-ktx:1.2.0") (or tracing for the Java-only API). Keep this even after the probe — it's tiny and useful long-term.
• Use the platform android.os.Trace directly: Trace.beginSection("...") / Trace.endSection() work on every API level without any dependency. You lose the trace { } lambda but the slices come through identically.

Groovy DSL. Use single-quoted strings:

implementation 'androidx.tracing:tracing-ktx:1.2.0'

Java codebase. Trace.beginSection("AGENT_TRACE_<id>....") / Trace.endSection() in a try/finally — same as Kotlin. The trace { } lambda is Kotlin-only; in Java, fall back to the manual begin/end form everywhere.

R8/ProGuard stripping Trace calls. Aggressive minification can elide Trace.beginSection calls in release. For the probe loop this isn't an issue (debug builds don't minify), but if you instrument a release build keep -keep class android.os.Trace { *; } in your ProGuard rules.

The pattern: instrument → capture → inspect → remove

1. Add the import and wrap the suspect block

import androidx.tracing.Trace      // androidx.tracing:tracing-ktx — preferred
// or: import android.os.Trace      // platform — works without dep but no ktx helpers

Wrap with traceSection (idiomatic) or manual begin/end:

import androidx.tracing.trace

trace("AGENT_TRACE_a4f9c2e1.fetchUser") {
    val user = repository.fetchUser(id)
    cache.put(id, user)
}

The label format AGENT_TRACE_<id>.<name> keeps the sentinel prefix for cleanup grep, with a human label for the trace UI.

For begin/end style (when wrapping isn't ergonomic, e.g. across coroutine boundaries):

Trace.beginSection("AGENT_TRACE_a4f9c2e1.fetchUser")
try {
    val user = repository.fetchUser(id)
    // ...
} finally {
    Trace.endSection()
}

Always pair begin with end in finally. A missing endSection corrupts the trace for the rest of the process.

2. Capture a Perfetto trace and analyze it

Capturing the trace and querying it are separate concerns — covered by dedicated skills:

• android-perfetto-capture — strategies (one-shot host script, on-device long-running, in-app start/stop), data-source selection, capture verification
• android-perfetto-analyze — trace_processor SQL recipes, including the slice-by-thread query you'll want for AGENT_TRACE_* markers

The minimum viable loop, with the sentinel-aware slice query baked in:

# 1. Capture (see android-perfetto-capture for strategy detail)
/tmp/record_android_trace -o /tmp/trace.perfetto-trace -t 10s -b 32mb \
    -a com.example.app sched freq am wm gfx view binder_driver dalvik &
CAPTURE_PID=$!
sleep 2
adb shell input tap 540 1200          # drive the suspect flow
sleep 6
wait $CAPTURE_PID

# 2. Query for your sentinel slices (see android-perfetto-analyze Recipe 1)
/tmp/trace_processor /tmp/trace.perfetto-trace -q - <<'SQL' > /tmp/trace-results.txt
SELECT
  s.name,
  s.ts,
  s.dur / 1e6 AS dur_ms,
  COALESCE(t.name, 'process/async') AS thread_name
FROM slice s
LEFT JOIN thread_track tt ON s.track_id = tt.id
LEFT JOIN thread t ON tt.utid = t.utid
WHERE s.name LIKE 'AGENT_TRACE_%'
ORDER BY s.ts;
SQL

LEFT JOIN is required: slices wrapped across coroutine withContext boundaries land on async tracks, not thread tracks, and an inner join would silently drop them.

3. Delegate the verdict

Hand the (small) result file to a Sonnet sub-agent:

Read /tmp/trace-results.txt. For each AGENT_TRACE slice, return <label>: <dur_ms>ms on <thread_name>. Flag any slice on main longer than 16ms. Under 60 words. model: "sonnet".

For frame-budget questions, jank attribution, or main-thread breakdowns, see android-perfetto-analyze Recipes 2–4.

4. Cleanup gate (BLOCKING)

rg 'AGENT_TRACE_'

Must return zero. Remove every wrapping trace { } block, every Trace.beginSection/endSection pair, and any imports that were added only for tracing.

rm -f /tmp/trace.perfetto-trace /tmp/trace-results.txt

If the project already had production trace sections (legitimate, named with the team's convention), don't touch those — only the AGENT_TRACE_ ones.

Common patterns

Suspect Composable recomposition cost:

@Composable
fun ExpensiveItem(data: Item) {
    androidx.tracing.trace("AGENT_TRACE_a4f9c2e1.ExpensiveItem") {
        // ... composable body ...
    }
}

Coroutine block suspected of running on main:

viewModelScope.launch {
    Trace.beginSection("AGENT_TRACE_a4f9c2e1.fetchAndCache.before-withContext")
    // ...
    Trace.endSection()

    withContext(Dispatchers.IO) {
        Trace.beginSection("AGENT_TRACE_a4f9c2e1.fetchAndCache.io-block")
        try {
            // ...
        } finally {
            Trace.endSection()
        }
    }
}

If the io-block slice lands on the main lane, your dispatcher is misconfigured.

Common mistakes

Mistake	Fix
Skipping the cleanup gate	rg 'AGENT_TRACE_' must return zero before commit
Missing endSection in try/finally	Pairs must always close — use trace { } lambda form when possible
Generic label like "work"	Use AGENT_TRACE_<id>.<name> so it's unique in the UI and greppable for cleanup
Confusing slice duration with thread time	Default dur is wall time; for CPU time use tts (thread timestamp) columns
Wrapping the wrong block	Sections need to bracket the work — wrap the call, not just its declaration
Inner JOIN on thread_track when querying	Async slices have no thread; use LEFT JOIN (see android-perfetto-analyze)

For the capture-side equivalents (no -a filter, buffer too small, didn't drive the flow during the window) see android-perfetto-capture.