Sorry, the table got mangled in my previous mail. Here it is again: Non-committer (n=1766): p50=1, p75=1, p90=3, p95=4, p99=10, max=23 Committer (n=70): p50=6, p75=10, p90=18, p95=20, p99=27, max=28
Andrew Chang <[email protected]> 於 2026年7月21日週二 下午4:50寫道： > Hi everyone, > > +1 on saving maintainer time. I collected ~22000 PRs opened since July > 2024 (the AI era) and calculated each author’s peak number of open PRs > (hopefully it would help) > First, I calculated the peak for both non-committers and committers (write > access). “Peak” means the maximum number of PRs an author had open at the > same time during this period. > > n p50 p75 p90 p95 p99 max Non-committer 1766 1 1 3 4 10 23 Committer 70 6 > 10 18 20 27 28 > > For non-committers, 95% had no more than 4 PRs open at once. A cap of 10 > is around p99 and would affect about 15 people (0.8%). A cap of 8 would > have affected about 28 people (1.6%). > > > having some aggregated statistics on how many open PRs contributors and > first-time contributors had in the past could help us determine that. > > Hi Przemek, > I took a closer look at non-committers and also tried to separate > first-time contributors here. I could not identify historical first-time > contributors exactly, so I used authors with only author_association = > NONE as a rough proxy. > > The interesting fact I found here is that newcomers usually had lower > peaks: 95% never had more than 2 open at once, and only 4 of 569 had a peak > above 10. There was still a very small tail in both groups, with a maximum > of 23. > > There is another thing I am concerning: I found that non-committers with > high peaks had "merge rates" from 0% to 100%, with a median around 71%. > This makes me think that the number of same-time-opened-PRs alone might not > tell us much about contribution quality. > > > However, there are some limitations in my statistics: > > (1) In the script, I found it hard to identify the "committer" since there > is no clear label in github API. So I consider a person a committer if > any of following conditions is true: > a. Any of its PR's author_association column is member/owner/collaborator > b. It has merged others' PR at least once > c. Its GitHub account is on the official Apache Airflow committer roster > (from ASF LDAP) > > (2) Historical draft status is not available, the two-year window may miss > PRs opened before it, and identifying committers was more difficult than I > expected, especially with private GitHub organization memberships. I > cross-checked the results with another implementation, but I may still have > missed something. > > I am happy to share the public script and aggregate report if anyone would > like to check them. > > Thanks, > Andrew > >
