We definitely need to filter the verticals list before we merge, since we have some VMR legs we don't intend to publish. We also seem to have some cases where the wrong RID is being used, which is a bit scary, and we're going for a first-come-first-served approach on downloads, so:

  Taking Package 'Microsoft.AspNetCore.App.Runtime.linux-x64.9.0.0-rc.1.24414.4.nupkg' from 'AzureLinux_x64_Cross_Alpine_Mono_x64'

That's scary. a) the RID is wrong, it should be linux-musl-x64 (not your fault, that's a VMR bug) b) actually we don't want to use any of the _Mono_ verticals for final join, as far as I'm aware (and since this package name was pulled from this vertical, the correct package wasn't pulled from AzureLinux_x64_Cross_x64)

We definitely need to filter the verticals list before we merge, since we have some VMR legs we don't intend to publish. We also seem to have some cases where the wrong RID is being used, which is a bit scary, and we're going for a first-come-first-served approach on downloads, so:

  Taking Package 'Microsoft.AspNetCore.App.Runtime.linux-x64.9.0.0-rc.1.24414.4.nupkg' from 'AzureLinux_x64_Cross_Alpine_Mono_x64'

That's scary. a) the RID is wrong, it should be linux-musl-x64 (not your fault, that's a VMR bug) b) actually we don't want to use any of the _Mono_ verticals for final join, as far as I'm aware (and since this package name was pulled from this vertical, the correct package wasn't pulled from AzureLinux_x64_Cross_x64)

I filed dotnet/source-build#4628 for the issue in (a)

We definitely need to filter the verticals list before we merge, since we have some VMR legs we don't intend to publish. We also seem to have some cases where the wrong RID is being used, which is a bit scary, and we're going for a first-come-first-served approach on downloads, so:

  Taking Package 'Microsoft.AspNetCore.App.Runtime.linux-x64.9.0.0-rc.1.24414.4.nupkg' from 'AzureLinux_x64_Cross_Alpine_Mono_x64'

That's scary. a) the RID is wrong, it should be linux-musl-x64 (not your fault, that's a VMR bug) b) actually we don't want to use any of the _Mono_ verticals for final join, as far as I'm aware (and since this package name was pulled from this vertical, the correct package wasn't pulled from AzureLinux_x64_Cross_x64)

would it make sense to list the verticals to join implicitly as dependencies of the final join job? We were thinking about that option but I was under the impression that all verticals need to be joined and that maintaining the list would be difficult

We definitely need to filter the verticals list before we merge, since we have some VMR legs we don't intend to publish. We also seem to have some cases where the wrong RID is being used, which is a bit scary, and we're going for a first-come-first-served approach on downloads, so:

  Taking Package 'Microsoft.AspNetCore.App.Runtime.linux-x64.9.0.0-rc.1.24414.4.nupkg' from 'AzureLinux_x64_Cross_Alpine_Mono_x64'

That's scary. a) the RID is wrong, it should be linux-musl-x64 (not your fault, that's a VMR bug) b) actually we don't want to use any of the _Mono_ verticals for final join, as far as I'm aware (and since this package name was pulled from this vertical, the correct package wasn't pulled from AzureLinux_x64_Cross_x64)

would it make sense to list the verticals to join implicitly as dependencies of the final join job? We were thinking about that option but I was under the impression that all verticals need to be joined and that maintaining the list would be difficult

Perhaps an exclusion works better. We don't want certain verticals, but we want all by default?

@mmitche are we going to use JoinVerticals task only for the final build pass? I think the downloads for 2nd build pass are implemented with #43288

@mmitche are we going to use JoinVerticals task only for the final build pass? I think the downloads for 2nd build pass are implemented with #43288

We may need to use it for both. the join 2nd pass joins will probably have some overlapping artifacts.

/cc @directhex

@mmitche are we going to use JoinVerticals task only for the final build pass? I think the downloads for 2nd build pass are implemented with #43288

We may need to use it for both. the join 2nd pass joins will probably have some overlapping artifacts.

/cc @directhex

The main thing #43288 did was allow a given job to depend on multiple other jobs' nupkg output. For example, workload creation requires all component nupkgs to exist on disk, so we need the results of the existing Windows_x64 vertical, but also the new vertical dotnet/source-build#3891 would add. The implementation is done in AzDO tasks, sequentially downloading every nupkg from the named jobs into the join point job's main packages folder.

It's not entirely obvious to me how that behaviour would be expected to intersect with this one (e.g. if a vertical creates runtime packs and uploads them, a second vertical creates aot compilers and uploads them, and a third vertical downloads the previous two uploads, adds in more packages, then uploads the full set of both the new packages and the ingested packages from earlier verticals, then the expected behaviour of the final join point isn't clear to me)

Some subset of the functionality is necessary. If you need Windows x86 and Window x64 as inputs to product a given vertical in pass 2, there is a very good chance that there will be some overlap of artifacts between the input verticals. We need a selection algorithm for that. The output vertical manifest is not the interesting output. The selection of inputs is.

@mmitche I'm working on checking for duplicate assets in non-primary verticals and it turns out there a lot of them.

I ran the task locally on artifacts from https://dev.azure.com/dnceng/internal/_build/results?buildId=2544654&view=results and there is a lot of overlap between Ubuntu2404_x64 and AzureLinux_x64_Cross_x64 also between all the *AzureLinux* verticals

So far I'm excluding everything containing "Mono", "Pgo", "Shortstack" and "DevVersions", let me know if that's correct

@mmitche I'm working on checking for duplicate assets in non-primary verticals and it turns out there a lot of them.

I ran the task locally on artifacts from https://dev.azure.com/dnceng/internal/_build/results?buildId=2544654&view=results and there is a lot of overlap between Ubuntu2404_x64 and AzureLinux_x64_Cross_x64 also between all the *AzureLinux* verticals

So far I'm excluding everything containing "Mono", "Pgo", "Shortstack" and "DevVersions", let me know if that's correct

Okay super interesting. I'll take a look at the list in a few. I think for now let's print this list as info, pick the first, and open a work item to drive it down.

I'm not sure how valid it will be that we run the join job in the normal PR job. Or at least, we may want to exclude some legs entirely from the join since some portion of the legs are for validation purposes and may have expected overlap. They may not represent the official build. For instance, the DevVersions and normal ubuntu legs would be expected to produce the same assets, just at different versions.

@mmitche I'm working on checking for duplicate assets in non-primary verticals and it turns out there a lot of them.
I ran the task locally on artifacts from https://dev.azure.com/dnceng/internal/_build/results?buildId=2544654&view=results and there is a lot of overlap between Ubuntu2404_x64 and AzureLinux_x64_Cross_x64 also between all the *AzureLinux* verticals
So far I'm excluding everything containing "Mono", "Pgo", "Shortstack" and "DevVersions", let me know if that's correct

Okay super interesting. I'll take a look at the list in a few. I think for now let's print this list as info, pick the first, and open a work item to drive it down.

I'm not sure how valid it will be that we run the join job in the normal PR job. Or at least, we may want to exclude some legs entirely from the join since some portion of the legs are for validation purposes and may have expected overlap. They may not represent the official build. For instance, the DevVersions and normal ubuntu legs would be expected to produce the same assets, just at different versions.

This test run prints a list https://dev.azure.com/dnceng/internal/_build/results?buildId=2547353&view=logs&j=012d8f49-bbe9-5cff-0c2d-eb37469742f3&t=a088a11a-052c-5a48-307b-2470eb67bf36&l=3378

I'd like to get #44153 merged in first and then react to that in this PR (should just be 1-2 conflicts). Mainly because testing the join point is very difficult and time consuming. Hope that's OK with everyone.