Ignore backport PRs

[gundalow]

https://stats.eng.ansible.com/~/admin/community.general.prs.html#large-subsections

Ignore any PR created by patchback Bot (example ansible-collections/community.general#1407)

OR maybe only look at PRs against the Git default branch (generally, though not always main)

[GregSutcliffe]

So I looked into this.

1. Ignoring the label does not work. The label is applied to the real PR so that patchback knows what to copy. The backported PR is not labelled, so we can't use this
2. Ignoring an author is of course easy and fine. I have added the patchback and modular-magician bots to an author filter in my internal reports, and will add an issue for doing so in the crawler. I'm also happy to be told about other names to exclude.
3. As for default branches, this is complex:

• When we index PRs, we store which branch it was made against (.e.g. "master") which comes from GitHub's API. This is static, and does not change over the life of the PR.
• However, to learn the default branch we must consult the /repos/ API, which returns the default branch today.

If that has changed (e.g. to "main") then any PRs which were valid before the change (e.g. "master" == "master") no longer match the filter (now it's "master" == "main", which is false). Worse, we can't even use tricks like storing the default branch when indexing PRs, because we overwrite the PR object in Mongo (we need to, because its state will change, say from OPEN to CLOSED). So when a default branch change occurs, we would end up overwriting it.

As a result, filtering on default branch currently results in removing more than 7000 PRs from the 9,500 PR collections dataset.

I can only think of two resolutions, neither ideal:

1. We create a new, append-only, table of repo, pr-number, is-it-default that is populated when we first see a PR appear, and never changed thereafter. This isn't too much work, but is extra complexity in our tooling.
2. We add requirements about default branch names to collections, so we can filter for them. Much easier to implement, but I highly doubt collection maintainers would agree to it.

In either case, these fixes would only work for PRs created after implementation. We still have no solution to filter the past year's dataset. It's a tricky one.

/cc @gundalow @dmsimard @Andersson007

[gundalow]

Only a few repos use PatchBot. Those that do currently use stable branches. So we could just ignore any branch that stays with stable

[GregSutcliffe]

"starts with stable" should that be?

That's an option, I'll look into it, thanks.

[Andersson007]

1. Ignoring the label does not work. The label is applied to the real PR so that patchback knows what to copy. The backported PR is not labelled, so we can't use this

They usually have some patterns in their short descriptions like [stable-1] or something. Couldn't we use them to identify such PRs?

[GregSutcliffe]

If it's consistent, then yes, I can string-match on the titles. The issue is generally that we change things over time, and I'm using 1 year of data, so if it works for things from a year ago, then great :)

[Andersson007]

It's not consistent:) but we could filter out at least most of backport PRs if there's no other option.
A merged PR is a commit to a branch. If we are able to get the commit with its branch, we could filter.