dulwich_log performance regression in 3.10

[Fidget-Spinner]

BPO	45330
Nosy	@methane, @pablogsal, @brandtbucher, @iritkatriel, @Fidget-Spinner

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = <Date 2021-10-18.14:03:36.110>
created_at = <Date 2021-09-30.13:09:28.525>
labels = ['3.11', '3.10', 'performance']
title = 'dulwich_log performance regression in 3.10'
updated_at = <Date 2021-10-18.14:03:36.110>
user = 'https://github.com/Fidget-Spinner'

bugs.python.org fields:

activity = <Date 2021-10-18.14:03:36.110>
actor = 'kj'
assignee = 'none'
closed = True
closed_date = <Date 2021-10-18.14:03:36.110>
closer = 'kj'
components = []
creation = <Date 2021-09-30.13:09:28.525>
creator = 'kj'
dependencies = []
files = []
hgrepos = []
issue_num = 45330
keywords = []
message_count = 7.0
messages = ['402955', '402978', '403006', '403036', '403037', '403286', '404187']
nosy_count = 5.0
nosy_names = ['methane', 'pablogsal', 'brandtbucher', 'iritkatriel', 'kj']
pr_nums = []
priority = 'normal'
resolution = 'rejected'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'performance'
url = 'https://bugs.python.org/issue45330'
versions = ['Python 3.10', 'Python 3.11']

[Fidget-Spinner]

Somewhere between May02-May11, dulwich_log benchmark on pyperformance had a major performance regression on the 3.10 branch. https://speed.python.org/timeline/?exe=12&base=&ben=dulwich_log&env=1&revs=200&equid=off&quarts=on&extr=on

For a start, we could git bisect with pyperformance. FWIW, I couldn't reproduce the perf regression on Windows. So it might be a Linux-only thing.

[iritkatriel]

Looks like this is using a 3rd party module (dulwitch.repo). Could it be a new version of that?

[Fidget-Spinner]

Ooh thanks Irit, I hadn't considered that. I thought that versions were fixed, but it seems that we *do* change the 3rd party package versions once in a while.

In that time period, we bumped dulwich from 0.20.20 to 0.20.21 python/pyperformance@6484a40.

That dulwich release had the following commits: jelmer/dulwich@dulwich-0.20.21...master.

This may also explain why I couldn't see any perf regression in dulwich between 3.10a4 and 3.10rc2 on Windows.

Once I verify this, I'll close this issue and open one on pyperformance repo.

[vstinner]

The pyperformance version in pinned on speed.python.org. Moreover, the timeline tracks the main branch, not the 3.10 branch.

[pablogsal]

We use a pinned version of pyperformance so the code has not changed. I need to manually bump the version and when they happens I wipe all previous benchmarks