Detecting Memory Leaks in Android A/B Tests: A Production-Focused Approach by Pavlo Stavytskyi
0 likes178 views
The document discusses memory leaks in Android applications, focusing on a production-oriented approach to detect these leaks during A/B testing. It evaluates existing tools, introduces memory metrics such as PSS, USS, and RSS, and emphasizes the need for efficient tracking methods to monitor memory usage in real-time. The presentation aims to improve A/B testing by providing insights into memory performance and identifying regressions or leaks effectively.
![How to retrieve RSS?
Refer to a system file located at /proc/[pid]/statm, where [pid] is the ID of
an application process.
■ android.os.Process.myPid() can be used to get [pid]
programmatically.](https://image.slidesharecdn.com/pavlostavytskyip99conf2024slides-241017201941-a51c044b/85/Detecting-Memory-Leaks-in-Android-A-B-Tests-A-Production-Focused-Approach-by-Pavlo-Stavytskyi-23-320.jpg)
Detecting Memory Leaks in Android A/B Tests: A Production-Focused Approach by Pavlo Stavytskyi
- 1. A ScyllaDB Community Detecting Memory Leaks in Android A/B Tests: A Production-Focused Approach Pavlo Stavytskyi Google Developer Expert for Android, Kotlin Experts
- 2. Pavlo Stavytskyi (he/him) Google Developer Expert - Android, Kotlin ■ Software Engineer at Meta ■ Previously Software Engineer at Lyft
- 3. Memory leaks
- 4. What are memory leaks? A memory leak occurs when objects in the app's memory that are no longer needed are still being referenced and therefore cannot be garbage collected. This leads to a gradual buildup of unused memory causing: ■ Performance slow down ■ Unresponsiveness ■ App crashes
- 5. Existing tools Tools for detecting memory leaks locally: ■ LeakCanary ■ Perfetto ■ Android Studio Memory Profiler
- 6. The problem with existing tools ■ Aimed at local profiling ● Debug or profilable release builds ● Manual testing ■ Hard to predict all possible circumstances ● Permutations of feature flags ● Specific Android devices ● Other conditions
- 7. Goals of the talk ■ Discover a memory monitoring approach for Android app that allows to: ● Understand the memory footprint of the app in production ● Detect regressions in A/B tests ■ Look into a tricky P99 memory leak detected at Lyft
- 8. Metrics
- 9. Metrics ■ There is no ultimate memory usage metric. ■ Where to retrieve metrics? ● Use Linux system file structure. ● Use Android SDK.
- 10. Android Studio Memory Profiler example
- 11. PSS Proportional set size (PSS) — the amount of private and shared memory used by the app where the amount of shared memory is proportional to the number of processes it is shared with. ■ If 3 processes are sharing 3MB, each process gets 1MB in PSS. ■ Used by Android Studio Memory Profiler
- 12. Full PSS breakdown import android.os.Debug import android.os.Debug.MemoryInfo val memoryInfo = MemoryInfo() Debug.getMemoryInfo(memoryInfo) val summary: Map<String, String> = memoryInfo.getMemoryStats()
- 13. Full PSS breakdown code: 12128 kB stack: 496 kB graphics: 996 kB java-heap: 8160 kB native-heap: 4516 kB private-other: 2720 kB system: 4955 kB // Includes all shared memory total-pss: 33971 kB // A sum of everything except 'total-swap' total-swap: 17520 kB
- 14. Just PSS import android.os.Debug val pssKb: Long = Debug.getPss()
- 15. Android Studio Memory Profiler example
- 16. USS Unique set size (USS) — the amount of private memory used by the app excluding shared memory. ■ Could be derived from PSS metrics on Android
- 17. USS import android.os.Debug.MemoryInfo val memoryInfo = MemoryInfo() Debug.getMemoryInfo(memoryInfo) val ussKb = with(memoryInfo) { getTotalPrivateClean() + getTotalPrivateDirty() }
- 18. Clean and dirty memory ■ Dirty memory — pages that have been modified by the app at runtime, so they must stay committed to RAM. ■ Clean memory — unmodified copy of a file in memory, so it can be cleared if not used. A clean page becomes a dirty page when it no longer contains an exact copy of the file (for example, from the result of an application operation).
- 19. What’s wrong with PSS and USS? Calling Debug.getMemoryInfo or Debug.getPss can take hundreds of milliseconds.
- 20. Different API for PSS import android.os.Debug.MemoryInfo val activityManager = context.getSystemService(Context.ACTIVITY_SERVICE) as ActivityManager val pid = intArrayOf(android.os.Process.myPid()) // Sample rate is limited to once per 5 minutes. val memoryInfo: MemoryInfo = activityManager.getProcessMemoryInfo(pid).first()
- 22. RSS Resident set size (RSS) — the amount of private and shared memory used by the app where all shared memory is included. ■ If three processes are sharing 3MB, each process gets 3MB in RSS. ■ Pessimistic metric. ■ Minimal performance overhead.
- 23. How to retrieve RSS? Refer to a system file located at /proc/[pid]/statm, where [pid] is the ID of an application process. ■ android.os.Process.myPid() can be used to get [pid] programmatically.
- 24. How to retrieve RSS? Reading statm file: ■ 3693120 27503 18904 1 0 319129 0 ■ (2) resident — resident set size, represented in pages. ■ rssKb = resident * 4 ■ The default page size on Linux is 4 kB
- 25. USS < PSS < RSS
- 27. Reporting metrics ■ Report a snapshot on every UI screen ■ Report a snapshot with metrics periodically with a given interval
- 28. Improving A/B testing A/B experiments allow comparing reported metrics between two app variants: ■ Treatment — the group of users that use the app with the new feature enabled. ■ Control — the group of users that use the app in a normal state with the new feature disabled.
- 29. Example 1 — no regression
- 30. Example 2 — regression
- 32. Memory allocated in heap
- 33. JVM heap val totalMemoryKb = Runtime.getRuntime().totalMemory() / 1024 val freeMemoryKb = Runtime.getRuntime().freeMemory() / 1024 val jvmHeapAllocatedKb = totalMemoryKb - freeMemoryKb
- 34. Native heap import android.os.Debug val nativeHeapAllocatedKb = Debug.getNativeHeapAllocatedSize() / 1024
- 35. Why is this approach efficient?
- 36. Example 3 — memory leak at 99th percentile
- 37. Thank you! Let’s connect. Pavlo Stavytskyi pavlo.stavytskyi@gmail.com @morfly_io morfly.medium.com