LeakCanary: Detect all memory leaks!

java.lang.OutOfMemoryError
        at android.graphics.Bitmap.nativeCreate(Bitmap.java:-2)
        at android.graphics.Bitmap.createBitmap(Bitmap.java:689)
        at com.squareup.ui.SignView.createSignatureBitmap(SignView.java:121)

Nobody likes OutOfMemoryError crashes

In Square Register, we draw the customer’s signature on a bitmap cache. This bitmap is the size of the device’s screen, and we had a significant number of out of memory (OOM) crashes when creating it.

We tried a few approaches, none of which solved the issue:

• Use Bitmap.Config.ALPHA_8 (a signature doesn’t need color).

• Catch OutOfMemoryError, trigger the GC and retry a few times (inspired from GCUtils).

• We didn’t think of allocating bitmaps off the Java heap. Lucky for us, Frescodidn’t exist yet.

We were looking at it the wrong way

The bitmap size was not a problem. When the memory is almost full, an OOM can happen anywhere. It tends to happen more often in places where you create big objects, like bitmaps. The OOM is a symptom of a deeper problem: memory leaks.

What is a memory leak?

Some objects have a limited lifetime. When their job is done, they are expected to be garbage collected. If a chain of references holds an object in memory after the end of its expected lifetime, this creates a memory leak. When these leaks accumulate, the app runs out of memory.

For instance, after Activity.onDestroy() is called, the activity, its view hierarchy and their associated bitmaps should all be garbage collectable. If a thread running in the background holds a reference to the activity, then the corresponding memory cannot be reclaimed. This eventually leads to an OutOfMemoryError crash.

Hunting memory leaks

Hunting memory leaks is a manual process, well described in Raizlabs’ Wrangling Dalvik series.

Here are the key steps:

1. Learn about OutOfMemoryError crashes via Bugsnag, Crashlytics, or the Developer Console.

2. Attempt to reproduce the problem. You might need to buy, borrow, or steal the specific device that suffered the crash. (Not all devices will exhibit all leaks!) You also need to figure out what navigation sequence triggers the leak, possibly by brute force.

3. Dump the heap when the OOM occurs (here’s how).

4. Poke around the heap dump with MAT or YourKit and find an object that should have been garbage collected.

5. Compute the shortest strong reference path from that object to the GC roots.

6. Figure out which reference in the path should not exist, and fix the memory leak.

What if a library could do all this before you even get to an OOM, and let you focus on fixing the memory leak?

Introducing LeakCanary

LeakCanary is an Open Source Java library to detect memory leaks in your debug builds.

Let’s look at a cat example:

class Cat {
}
class Box {
  Cat hiddenCat;
}
class Docker {
  static Box container;
}
// ...
Box box = new Box();
Cat schrodingerCat = new Cat();
box.hiddenCat = schrodingerCat;
Docker.container = box;

You create a RefWatcher instance and give it an object to watch:

// We expect schrodingerCat to be gone soon (or not), let's watch it.
refWatcher.watch(schrodingerCat);

When the leak is detected, you automatically get a nice leak trace:

* GC ROOT static Docker.container
* references Box.hiddenCat
* leaks Cat instance

We know you’re busy writing features, so we made it very easy to setup. With just one line of code, LeakCanary will automatically detect activity leaks:

public class ExampleApplication extends Application {
  @Override public void onCreate() {
    super.onCreate();
    LeakCanary.install(this);
  }
}

You get a notification and a nice display out of the box:

Conclusion

After enabling LeakCanary, we discovered and fixed many memory leaks in our app. We even found a few leaks in the Android SDK.

The results are amazing. We now have 94% fewer crashes from OOM errors.

If you want to eliminate OOM crashes, install LeakCanary now! Pierre-Yves Ricau (@Piwai) | Twitter The latest Tweets from Pierre-Yves Ricau (@Piwai). Android baker @Square. Paris / San Franciscotwitter.com