[PROF-9469] Skip heap samples with age 0

[AlexJF]

What does this PR do?

Skip updating and serialization of heap samples for objects that were allocated in the current GC epoch/generation. The fact that these objects are still alive is usually just a happy accident from the fact that GC hasn't run yet and thus just add noise when you want to understand long-term heap usage and, more specifically, memory leaks.

In fact, allocation profiles alone can give you a rather accurate description of what the population of age == 0 should look like at any-one time. If no GCs ever ran, allocation profiles would be === to heap profiles as all objects would be technically alive.

Motivation:

• Objects with age == 0 tend to be mostly noise when investigating heap usage.
• Unless GCs occur at intervals <1 minute or you have big memory leaks, objects with age == 0 will naturally be much more common than objects with age > 0 and so end up being major contributors to sampling and serialization overhead. Skipping them can result in significant savings.

Additional Notes:

Some micro-benchmark results:

Comparison:
sample+serialize min_age_1 retain_every=10 heap_samples=true heap_size=true heap_sample_every=10 skip_end_gc=true:        5.7 i/s
sample+serialize min_age_0 retain_every=10 heap_samples=true heap_size=true heap_sample_every=10 skip_end_gc=true:        5.5 i/s - 1.02x  slower
sample+serialize min_age_1 retain_every=10 heap_samples=true heap_size=true heap_sample_every=1 skip_end_gc=true:        4.3 i/s - 1.33x  slower
sample+serialize min_age_0 retain_every=10 heap_samples=true heap_size=true heap_sample_every=1 skip_end_gc=true:        3.1 i/s - 1.85x  slower

How to test the change?

Unsure? Have a question? Request a review!

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 98.27%. Comparing base (f6b234f) to head (f7989e4).
Report is 19 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #3573      +/-   ##
==========================================
- Coverage   98.27%   98.27%   -0.01%     
==========================================
  Files        1274     1274              
  Lines       75233    75247      +14     
  Branches     3545     3545              
==========================================
+ Hits        73934    73947      +13     
- Misses       1299     1300       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[ivoanjo]

Looks good! I left my usual smattering of minor suggestions, one bigger suggestion and one maybe-a-bug note ;)

[ivoanjo]

👍 Looks great!

I'm slightly curious if the change moved the micro-benchmark results needle, but feel free to ignore my curiosity haha

[AlexJF]

I'm slightly curious if the change moved the micro-benchmark results needle, but feel free to ignore my curiosity haha

Did a local run and saw

Comparison:
sample+serialize min_age_1 retain_every=10 heap_samples=true heap_size=true heap_sample_every=10 skip_end_gc=true:        5.7 i/s
sample+serialize min_age_0 retain_every=10 heap_samples=true heap_size=true heap_sample_every=10 skip_end_gc=true:        5.2 i/s - 1.08x  slowe
r
sample+serialize min_age_1 retain_every=10 heap_samples=true heap_size=true heap_sample_every=1 skip_end_gc=true:        4.0 i/s - 1.40x  slower
sample+serialize min_age_0 retain_every=10 heap_samples=true heap_size=true heap_sample_every=1 skip_end_gc=true:        3.0 i/s - 1.91x  slower

which looks similar enough to the original in the PR description

[ivoanjo]

Ah! Somewhat sad, I was hoping this tweak would buy is a bit more headroom. I guess it might still on real-world usecases, but would've been cool to see in the microbenchmark as well :)