The advantage of our current java_heap value is that regressions are very visible. But I also like the idea of displaying just the allocated_objects_size because that's the thing we have control over.

I suspect the latter should be sufficient to notice regressions as well -- considering I'd expect Android implementation of VM not to change TOO often.

I guess the only case I can think of is: if we allocate and free large number of objects creating fragmentation more than we did before, that would be caught by the former but not the latter.

On the other hand, the step function growth can also mask more objects being allocated more efficiently. ¯\_(ツ)_/¯

> I guess the only case I can think of is: if we allocate and free large number of objects creating fragmentation more than we did before, that would be caught by the former but not the latter.

Yeah but IIRC ART has a moving GC, so at some point the situation should normalize.

The proble of the step function is that blames just *some* people who cross the step and hides most of other ones.

I like the idea of changing it to allocated_objects_size. We should definitely not stick to the step function. Also can we just turn on altering for the java_heap:allocated_objects_size ?