Uh oh!
There was an error while loading. Please reload this page.
Simplify RID graph: keep floating and last versioned RIDs - #125076
Simplify RID graph: keep floating and last versioned RIDs#125076richlander wants to merge 1 commit into
Conversation
Softer approach to RID graph simplification. Instead of removing all versioned and distro-specific RIDs, this keeps: - All portable/base RIDs (any, unix, linux, win, etc.) - All floating (unversioned) OS RIDs (alpine, ubuntu, osx, etc.) - The last versioned RID for each OS family and its arch variants This reduces runtime.json from ~4600 lines (798 RIDs) to ~1500 lines (266 RIDs) while preserving a migration path for packages that reference the most recent versioned RIDs. Removed RIDs: all intermediate versioned RIDs (e.g. alpine.3.7 through alpine.3.17, ubuntu.14.04 through ubuntu.23.04, osx.10.10 through osx.12, etc.) Kept last versions include: alpine.3.18, android.32, centos.9, debian.12, fedora.39, freebsd.13, ios.15, osx.13, rhel.9, rocky.9, tizen.7.0.0, tvos.15, ubuntu.23.10, and others. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Tagging subscribers to this area: @agocke, @jeffschwMSFT, @elinor-fung |
ericstj
commented
Mar 2, 2026
What have you done to test this? Did you consider making this change in the VMR so that we can see if source build and phase2 product build are happy with the changes? |
richlander
commented
Mar 2, 2026
Great questions. I haven't done anything to test it. My belief/understanding is that the fallout is compat. I can make this change in VMR. That didn't occur to me. My thinking -- perhaps naive -- is that any break there is a bug that needs to be fixed downstream. I'll apply this change to VMR. |
Softer approach to RID graph simplification. Instead of removing all versioned and distro-specific RIDs, this keeps: - All portable/base RIDs (any, unix, linux, win, etc.) - All floating (unversioned) OS RIDs (alpine, ubuntu, osx, etc.) - The last versioned RID for each OS family and its arch variants This reduces runtime.json from ~4600 lines (798 RIDs) to ~1500 lines (266 RIDs) while preserving a migration path for packages that reference the most recent versioned RIDs. Backport of dotnet/runtime#125076 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
It should be, but we might want to get that signal now. I think the VMR can give at least a sampling of those it won't run the SDK tests which I think might be equally valuable, but we don't have a good way to do that in PR. I don't have any objections to the PR itself, it's more "what's the fallout". I trust its well formed, but I am not reviewing for that. We used to have a generator for this that made it a bit easier to review the changes to the generator inputs. |
richlander
commented
Mar 2, 2026
Understood. We're on the same page. |
tmds
commented
Mar 3, 2026
I'd consider removing these too. The "modern" rid handling for a non-portable SDK depends on: |
ericstj
commented
Mar 3, 2026
I shared some data offline to show what packages are using the RIDs being discussed here. That provides a view of the producer side - we can't know what people use when building apps (consumer side) but a review of that data might help understand the compat impact of removing RIDs that might be used during restore. |
richlander
commented
Mar 3, 2026
This proposal is on top of this break that we made 3 years ago: dotnet/docs#36466 That break has a workaround. You can see it in use: https://grep.app/search?q=%3CUseRidGraph%3Etrue%3C%2FUseRidGraph%3E The use of |
am11
commented
Mar 3, 2026
My understanding is; with Microsoft SDK:
Distro maintainers use 3rd option (graph gets generated during the source-bulid). If this is correct, then we don't need to change runtime.json for older/outdated SDK deployments to keep working. Otherwise, I misunderstood. 🙂 cc @elinor-fung |
richlander
commented
Mar 3, 2026
The intent is to remove cruft from the product. The RID file is basically a wikipedia of old distros. I'd like to eventually delete the file. I grasp, as you are suggesting, that its also enable compatibility for old assets on nuget.org, otherise they are basically unusable. Consumers can either use old SDKs, insert their own RID graph, or user a newer version. Some of the libraries may be abandoned. It's OK to have to take extra steps to use abandoned software. |
richlander
commented
Mar 3, 2026
Here is the data we collected, sorted by package count with a cutoff at 10 packages. The primary data I see is that we should consider bringing some Ubuntu RIDs back. On that note, I see that Ubuntu 23.10 is the latest/retained Ubuntu RID. That's not super useful. If we follow the data, we could replace 23.10 with 18.04 and that would solve most of the problem.
|
tmds
commented
Mar 4, 2026
Ah, yes. I think users should use APIs like It doesn't seem a mechanism that is relied on for splitting nuget packages either. Searching for latest Ubuntu LTS shows only 7 packages (by 5 users): https://www.nuget.org/packages?q=ubuntu.24.04-x64. |
richlander
commented
Mar 4, 2026
That's a good point. As I hinted at, my hope would be to make this change for 11 and the delete most of the infra for 12. |
This fallback code (`get_current_os_rid_platform`) will probably go away in the future once #125076 has landed, but as it stands, we need this for parity. C++ libraries version-less symlinks are needed for cmake/clang to avoid falling back to non-fPIC libc++.a (because libc++.so.x.y is versioned). With this patch, `./build.sh host.native --cross --os openbsd --arch x64` succeeds. Upstream PR for eng/common change: dotnet/arcade#16564
Workflow state for the Holistic Review Orchestrator. {
"version": 5,
"last_dispatched_commit": "1226615d9d617de8344fecfa35c9e91c65417538",
"last_dispatched_base_ref": "main",
"last_dispatched_base_sha": "b0bc48eea7e28be9670d4ac499a5778d0ac99909",
"last_reviewed_commit": "1226615d9d617de8344fecfa35c9e91c65417538",
"last_reviewed_base_ref": "main",
"last_reviewed_base_sha": "b0bc48eea7e28be9670d4ac499a5778d0ac99909",
"last_recorded_worker_run_id": "29673725116",
"review_attempt_commit": "",
"review_attempt_base_ref": "",
"review_attempt_count": 0,
"max_review_attempts": 5,
"review_history_format": "holistic-review-disclosure-v1",
"review_history": [
{
"commit": "1226615d9d617de8344fecfa35c9e91c65417538",
"review_id": 4730113525
}
]
} |
There was a problem hiding this comment.
Holistic Review
Motivation: The PR trims src/libraries/Microsoft.NETCore.Platforms/src/runtime.json from 798 RIDs (~4600 lines) to 266 RIDs (~1500 lines). The goal is to stop the unbounded growth of the non-portable RID graph by removing every intermediate versioned/distro-specific RID while preserving base/portable RIDs, all floating (unversioned) OS RIDs, and the single most-recent versioned RID for each OS family plus its arch variants. It is a softer successor to #123161 (which removed all versioned/distro RIDs) intended to leave a migration path for packages that reference the newest versioned RIDs.
Approach: This is a pure data-file change to the transport package that is redistributed into the dotnet/sdk RID graph; no code, tooling (UpdateRuntimeIdentifierGraph.cs), or PortableRuntimeIdentifierGraph.json was touched. I verified the resulting graph independently of the description:
- JSON is well-formed (
jqparses;runtimeshas 266 entries). - No dangling imports: every RID referenced in a
#importarray (198 distinct) is still defined as a top-level key. So removing ~721 keys did not orphan any surviving RID. - Kept versioned RIDs were correctly re-pointed: previously each last-version RID chained through the removed intermediate (e.g.
osx.13->osx.12,alpine.3.18->alpine.3.17,osx.13-x64->osx.12-x64). In the new graph they import their floating base directly (osx.13->osx,osx.13-x64->osx.13,osx-x64), which is the correct rewrite to keep the graph connected after the intermediates are gone. - No repo-internal fallout: I grepped
eng/andsrc/for representative removed RIDs (e.g.osx.10.12,ubuntu.18.04,alpine.3.17,rhel.7) and found no build/test/pipeline references that depend on them.
The change is internally consistent and the mechanical execution looks correct.
Summary: LGTM on correctness of the graph edit itself. The substantive consideration is not a code defect but a compatibility/policy judgment that belongs to the maintainers and the SDK team: removing intermediate versioned RIDs is an intentional behavior change to NuGet RID-based asset selection. Any existing package that ships assets exclusively under a now-removed RID (e.g. runtimes/ubuntu.16.04-x64/...) will no longer have that RID in the graph, so a project resolving for such a RID would fall back to the floating/base asset (or fail to see the RID at all) rather than the version-specific one. The PR description acknowledges this as the intended tradeoff with a documented migration story, and readme.md already states the graph should no longer gain new version-specific RIDs. Because this is the whole point of the PR and there is no single changed line to attach it to, I raise it here rather than as an inline finding. I'd recommend confirming (a) the SDK-side consumption and any published compat/breaking-change guidance are lined up, and (b) that the "last versioned RID per family" selection is genuinely the newest for each family (spot checks such as alpine.3.18, osx.13, ubuntu.23.10, debian.12, fedora.39 looked right). No blocking correctness issues found in the diff.
Note
This review was generated by this repository's Holistic Review agentic workflow to complement the built-in Copilot review.
Generated by Holistic Review · 130.8 AIC · ⌖ 10.7 AIC · ⊞ 10K
elinor-fung
commented
Jul 30, 2026
I assume we'd want to breaking change doc this? The change seems generally reasonable to me as a step towards a future larger breaking change of removing the non-portable RID graph compat mode entirely. |
Replaces: #123161
Softer approach to RID graph simplification. Instead of removing all versioned and distro-specific RIDs, this keeps:
This reduces runtime.json from ~4600 lines (798 RIDs) to ~1500 lines (266 RIDs) while preserving a migration path for packages that reference the most recent versioned RIDs.
We can do more cleanup later. This approach is a much safer starting point with easier migration guidance. It's also more defensible.
Removed RIDs: all intermediate versioned RIDs (e.g. alpine.3.7 through alpine.3.17, ubuntu.14.04 through ubuntu.23.04, osx.10.10 through osx.12, etc.)
Kept last versions include: alpine.3.18, android.32, centos.9, debian.12, fedora.39, freebsd.13, ios.15, osx.13, rhel.9, rocky.9, tizen.7.0.0, tvos.15, ubuntu.23.10, and others.