杨杰, yangjie

the profiling default recorder max events is 16384. so the alloc_buffer_size / 0.5 * 60s = 16384. the alloc_buffer_size would be safely set to 128.

i also noticed that all event in same allocation array, even they have different stacktrace. they have the same count, this would make the finally result count x of the...

使用 tracemalloc 获取的实际内存分配次数以及 bytes： ``` python3 benchmakr.py iteration 10000 [ Top 10 ] /root/miniconda3/lib/python3.11/site-packages/PIL/JpegImagePlugin.py:273: size=12.3 MiB, count=49959, average=259 B :88: size=10.0 MiB, count=108131, average=97 B /home/centos/pillow-bench/benchmakr.py:20: size=7241 KiB, count=108132, average=69...

> 感谢您的报告，@zdyj3170101136。您认为您的应用程序可能需要比默认值更大的缓冲区吗？也许我们可以更改一些文档来澄清这一点？ > > 抄送@sanchda sure, maybe the event_buffer dynamic set from reporter buffer and flush interval.

@felixge could you check it?

> https://github.com/mosn/holmes dd-trace-go on my own applications. no, just a proposal. now, per minute push of memory profiler is not enough for oom, especially for pod. the k8s pod memory...

> @zdyj3170101136感谢您的报告。 > > 你能详细说明你的二进制文件吗？二进制文件是否包含 DWARF？（帮助我们了解使用了哪种符号化策略） > > 如果可能的话，能否请您附上捕获的个人资料？ > what i see about parca agent profile is in [demo.parca.dev.](https://demo.parca.dev/) you can find that the function runtime.goexit's have wrong file...

@benfred

here is flamegraph of perf record -g py-spy: 

when i use --gil, the issue disapppear.