As a dev working on a repo upstream from the SDK, it would be desirable to be able to build a full SDK that incorporates my changes to see if the SDK's behavior is changed as expected. It is also convenient to test changes in a full end-to-end environment.
This has additional uses in the context of source-build. In order to build from source, we need to be able to bootstrap: build a .NET Core SDK, then build the entire product again using that .NET Core SDK. Without speculatively building an SDK, bootstrapping can only be tested in the Core-SDK repository. In some cases, this is too late to actually fix any problems that are encountered. We need to push validation for this scenario as far upstream as we can.
Speculative builds are complex. They have been considered before for the Microsoft build, but haven't been done. It may be significantly easier to do in source-build, because the entire build can happen on a single machine, rather than every supported platform (requiring transporting artifacts across machines).
The red circle is the current build. For example, dotnet/aspnetcore on master.
The black circles are the builds that were used for intermediates.
First, we need to choose a speculative root build. We choose Core-SDK because it produces the SDK, and the latest version because it's most likely what we want to test against. This results in the blue graph. Note that because we have to make a choice here, there are many possible speculative builds we could perform, and they might not all work!
Then, we replace all the nodes in the speculative root graph with our local build graph nodes, based on repository identity. This forms the third graph in the diagram.
Finally, we rebuild the remaining nodes in the speculative graph, using our local build outputs. This gives us a speculative SDK as an output.
(The big arrow means from and/or to many nodes.)
We can use the output SDK and intermediates from our bootstrap build to try to build the product again, using the same set of repositories/commits. If this is successful, we have validated the bootstrap build scenario.
(The big arrow means from and/or to many nodes.)
If we instead use the output SDK and intermediates from an N-1 build, instead of building the local speculative graph, we validate the N-1 scenario.