P-Y
@p-y.wtf
https://p-y.wtf | Android Eng @Square | hashnode.com/@py | instagram.com/py.ricau | LeakCanary, Radiography, square/logcat, square/PAPA. He/Him
TIL that you can easily slice a pizza or pie in 5 pieces: make a paper pentagon, place it at the center, use corners as slice starting points.
Paper pentagon? Tie a simple overhand knot with a long strip of paper, keeping it flat, cut off the ends.
Wow. Life changing trick 😊
Paper pentagon? Tie a simple overhand knot with a long strip of paper, keeping it flat, cut off the ends.
Wow. Life changing trick 😊
November 8, 2025 at 2:51 PM
TIL that you can easily slice a pizza or pie in 5 pieces: make a paper pentagon, place it at the center, use corners as slice starting points.
Paper pentagon? Tie a simple overhand knot with a long strip of paper, keeping it flat, cut off the ends.
Wow. Life changing trick 😊
Paper pentagon? Tie a simple overhand knot with a long strip of paper, keeping it flat, cut off the ends.
Wow. Life changing trick 😊
In my experience, in a perf context (benchmarks or prod data), if you manage to isolate dimensions, you often end up with overlapping normal distributions (e.g. in the previous example 3 gaussians)
I need to look into this clustering method called Gaussian Mixture Models
I need to look into this clustering method called Gaussian Mixture Models
October 24, 2025 at 1:13 PM
In my experience, in a perf context (benchmarks or prod data), if you manage to isolate dimensions, you often end up with overlapping normal distributions (e.g. in the previous example 3 gaussians)
I need to look into this clustering method called Gaussian Mixture Models
I need to look into this clustering method called Gaussian Mixture Models
Even with a test that does ONLY scrolling frames, the distribution different peaks/modes (e.g. fast, less fast, slow)
Why? Some frames just move things already drawn by a few px, others have to redraw a whole new item.
Saying "the P50 shifted from 6 ms to 4.6 ms" hides these clusters.
Why? Some frames just move things already drawn by a few px, others have to redraw a whole new item.
Saying "the P50 shifted from 6 ms to 4.6 ms" hides these clusters.
October 24, 2025 at 1:13 PM
Even with a test that does ONLY scrolling frames, the distribution different peaks/modes (e.g. fast, less fast, slow)
Why? Some frames just move things already drawn by a few px, others have to redraw a whole new item.
Saying "the P50 shifted from 6 ms to 4.6 ms" hides these clusters.
Why? Some frames just move things already drawn by a few px, others have to redraw a whole new item.
Saying "the P50 shifted from 6 ms to 4.6 ms" hides these clusters.
Frames durations are ANYTHING BUT normally distributed.
First, maybe your test sucks.
Here, both "before" & "after" show a slow 1st (starting activity) & 40th frame (load images)
That test isn't correctly focused on scrolling frames, it includes initial loading frames.
First, maybe your test sucks.
Here, both "before" & "after" show a slow 1st (starting activity) & 40th frame (load images)
That test isn't correctly focused on scrolling frames, it includes initial loading frames.
October 24, 2025 at 1:13 PM
Frames durations are ANYTHING BUT normally distributed.
First, maybe your test sucks.
Here, both "before" & "after" show a slow 1st (starting activity) & 40th frame (load images)
That test isn't correctly focused on scrolling frames, it includes initial loading frames.
First, maybe your test sucks.
Here, both "before" & "after" show a slow 1st (starting activity) & 40th frame (load images)
That test isn't correctly focused on scrolling frames, it includes initial loading frames.
Anyway, thanks ChatGPT, it does look like the Hodges–Lehmann estimator is what I was looking for to compute the P50 with CI shift in frame duration between a before & after benchmark of 100 iterations each (where each iteration has N frame and computes a P50 frame time)
cc @rahulrav.com
cc @rahulrav.com
October 24, 2025 at 12:02 PM
Anyway, thanks ChatGPT, it does look like the Hodges–Lehmann estimator is what I was looking for to compute the P50 with CI shift in frame duration between a before & after benchmark of 100 iterations each (where each iteration has N frame and computes a P50 frame time)
cc @rahulrav.com
cc @rahulrav.com
Sometimes it's also helpful to visualize the values in order instead of as a distribution. For example, in this blog post I'm writing, you can see how the first frame is super slow, and then some other frame later is slow, in both before & after benchmark runs.
October 22, 2025 at 8:52 PM
Sometimes it's also helpful to visualize the values in order instead of as a distribution. For example, in this blog post I'm writing, you can see how the first frame is super slow, and then some other frame later is slow, in both before & after benchmark runs.
- This would be most useful if you could load TWO json files, the before & after, and visually compare the distributions / see shifts
- I'm guess you picked 100 bins as default, and (max-min)/ bin_count as bucket width. Not always good. Try other things, e.g. en.wikipedia.org/wiki/Freedma... ?
- I'm guess you picked 100 bins as default, and (max-min)/ bin_count as bucket width. Not always good. Try other things, e.g. en.wikipedia.org/wiki/Freedma... ?
October 22, 2025 at 8:49 PM
- This would be most useful if you could load TWO json files, the before & after, and visually compare the distributions / see shifts
- I'm guess you picked 100 bins as default, and (max-min)/ bin_count as bucket width. Not always good. Try other things, e.g. en.wikipedia.org/wiki/Freedma... ?
- I'm guess you picked 100 bins as default, and (max-min)/ bin_count as bucket width. Not always good. Try other things, e.g. en.wikipedia.org/wiki/Freedma... ?
Neat! I gave it a quick try, some feedback
- The Drag&Drop box should be the entire screen (if you miss the rectangle, Chrome just opens your json file)
- I loaded a file, it showed a nice chart, but then bottom seems to have something else trying to load (I see a loader) and nothing is loading.
- The Drag&Drop box should be the entire screen (if you miss the rectangle, Chrome just opens your json file)
- I loaded a file, it showed a nice chart, but then bottom seems to have something else trying to load (I see a loader) and nothing is loading.
October 22, 2025 at 8:45 PM
Neat! I gave it a quick try, some feedback
- The Drag&Drop box should be the entire screen (if you miss the rectangle, Chrome just opens your json file)
- I loaded a file, it showed a nice chart, but then bottom seems to have something else trying to load (I see a loader) and nothing is loading.
- The Drag&Drop box should be the entire screen (if you miss the rectangle, Chrome just opens your json file)
- I loaded a file, it showed a nice chart, but then bottom seems to have something else trying to load (I see a loader) and nothing is loading.
Great crowd with the Paris Android User Group at Radio France
Showing how frame metrocs & percentiles are computed in Macrobenchmark, then I went on to explain how to draw a distribution, compute confidence intervals, and do bootstrapping. That was a little ambitious, but tons of fun
#AndroidFun
Showing how frame metrocs & percentiles are computed in Macrobenchmark, then I went on to explain how to draw a distribution, compute confidence intervals, and do bootstrapping. That was a little ambitious, but tons of fun
#AndroidFun
October 21, 2025 at 10:51 PM
Great crowd with the Paris Android User Group at Radio France
Showing how frame metrocs & percentiles are computed in Macrobenchmark, then I went on to explain how to draw a distribution, compute confidence intervals, and do bootstrapping. That was a little ambitious, but tons of fun
#AndroidFun
Showing how frame metrocs & percentiles are computed in Macrobenchmark, then I went on to explain how to draw a distribution, compute confidence intervals, and do bootstrapping. That was a little ambitious, but tons of fun
#AndroidFun
Neat! I like to make 2mm thick cookies (using a special rolling pin) and cook them in between 2 sylpain mats for half the bake, keeps them flat + adds a neat detailed pattern
October 19, 2025 at 4:16 PM
Neat! I like to make 2mm thick cookies (using a special rolling pin) and cook them in between 2 sylpain mats for half the bake, keeps them flat + adds a neat detailed pattern
This is in less than an hour!
I'll be talking about Benchmarking with confidence.. wish me luck, live demos incoming
www.meetup.com/new-york-and...
#AndroidDev
I'll be talking about Benchmarking with confidence.. wish me luck, live demos incoming
www.meetup.com/new-york-and...
#AndroidDev
October 16, 2025 at 8:09 PM
This is in less than an hour!
I'll be talking about Benchmarking with confidence.. wish me luck, live demos incoming
www.meetup.com/new-york-and...
#AndroidDev
I'll be talking about Benchmarking with confidence.. wish me luck, live demos incoming
www.meetup.com/new-york-and...
#AndroidDev
[French]
Mardi 21 Octobre (dans 2 semaines!) à 19h, le Paris Android User Group organise un meetup chez Radio France. Il paraît que le lieu est magnifique!
Je serai là en 2e partie pour vous parler de benchmarks dignes de confiance.
www.meetup.com/android-pari...
#AndroidDev #paris
Mardi 21 Octobre (dans 2 semaines!) à 19h, le Paris Android User Group organise un meetup chez Radio France. Il paraît que le lieu est magnifique!
Je serai là en 2e partie pour vous parler de benchmarks dignes de confiance.
www.meetup.com/android-pari...
#AndroidDev #paris
October 7, 2025 at 8:59 AM
[French]
Mardi 21 Octobre (dans 2 semaines!) à 19h, le Paris Android User Group organise un meetup chez Radio France. Il paraît que le lieu est magnifique!
Je serai là en 2e partie pour vous parler de benchmarks dignes de confiance.
www.meetup.com/android-pari...
#AndroidDev #paris
Mardi 21 Octobre (dans 2 semaines!) à 19h, le Paris Android User Group organise un meetup chez Radio France. Il paraît que le lieu est magnifique!
Je serai là en 2e partie pour vous parler de benchmarks dignes de confiance.
www.meetup.com/android-pari...
#AndroidDev #paris
Packed schedule:
- Last week, Block offsite in California with 8000 employees... Sleeping from midnight (party!) to 4am (jetlag!) every night.. + 2 talk engs talks to give
- then Octoberfest in Munich this weekend
- then Droidcon Berlin next week, with 5mn in keynote + 40mm talk
#AndroidDev
- Last week, Block offsite in California with 8000 employees... Sleeping from midnight (party!) to 4am (jetlag!) every night.. + 2 talk engs talks to give
- then Octoberfest in Munich this weekend
- then Droidcon Berlin next week, with 5mn in keynote + 40mm talk
#AndroidDev
September 21, 2025 at 2:13 PM
Packed schedule:
- Last week, Block offsite in California with 8000 employees... Sleeping from midnight (party!) to 4am (jetlag!) every night.. + 2 talk engs talks to give
- then Octoberfest in Munich this weekend
- then Droidcon Berlin next week, with 5mn in keynote + 40mm talk
#AndroidDev
- Last week, Block offsite in California with 8000 employees... Sleeping from midnight (party!) to 4am (jetlag!) every night.. + 2 talk engs talks to give
- then Octoberfest in Munich this weekend
- then Droidcon Berlin next week, with 5mn in keynote + 40mm talk
#AndroidDev
September 20, 2025 at 2:42 PM
This isn't a valid way to present improvements.
Let's say we run a benchmark again with no change and get a slight different aggregate. E.g. for "before" P50 -5 instead of -4.2. Suddenly the improvement vs "after" looks less good!
What now? Which one do you pick?
This is why we need stats.
Let's say we run a benchmark again with no change and get a slight different aggregate. E.g. for "before" P50 -5 instead of -4.2. Suddenly the improvement vs "after" looks less good!
What now? Which one do you pick?
This is why we need stats.
August 31, 2025 at 7:19 AM
This isn't a valid way to present improvements.
Let's say we run a benchmark again with no change and get a slight different aggregate. E.g. for "before" P50 -5 instead of -4.2. Suddenly the improvement vs "after" looks less good!
What now? Which one do you pick?
This is why we need stats.
Let's say we run a benchmark again with no change and get a slight different aggregate. E.g. for "before" P50 -5 instead of -4.2. Suddenly the improvement vs "after" looks less good!
What now? Which one do you pick?
This is why we need stats.
Android devices have all sorts of screen sizes, so you can't really predict how many items will be on / off screen. Feels like something Compose could help with, dynamically switching Lazy off in those cases? idk
August 31, 2025 at 6:28 AM
Android devices have all sorts of screen sizes, so you can't really predict how many items will be on / off screen. Feels like something Compose could help with, dynamically switching Lazy off in those cases? idk
Ahem. This can't be serious advice, how could one audit a large codebase for this, when it's not even documented which APIs trigger binder calls?
AOSP should document APIs that make binder calls + have custom trace sections for all binder calls.
AOSP should document APIs that make binder calls + have custom trace sections for all binder calls.
August 31, 2025 at 6:22 AM
Ahem. This can't be serious advice, how could one audit a large codebase for this, when it's not even documented which APIs trigger binder calls?
AOSP should document APIs that make binder calls + have custom trace sections for all binder calls.
AOSP should document APIs that make binder calls + have custom trace sections for all binder calls.
I love this simplified example, but the colors are too vibrant, Perfetto could never produce this 😉. Trivia: span color in Perfetto is actually a hash of the trace section string
August 31, 2025 at 6:17 AM
I love this simplified example, but the colors are too vibrant, Perfetto could never produce this 😉. Trivia: span color in Perfetto is actually a hash of the trace section string