fix(compile): stop dual-writing the compile-state satellite nobody reads (#5389) by rbuergi · Pull Request #5571 · Systemorph/MeshWeaver

rbuergi · 2026-09-23T18:16:27Z

What this removes

On every NodeType hub activation, MeshDataSource installed NodeTypeCompileStateMirror, which wrote the fixed-id satellite {type}/_Activity/compile-state. That cost:

one CreateOrUpdateNodeRequest through the node-operation hub on every activation;
one extra grain activation per NodeType whenever the state moved. A framework roll re-stamps every type, which is about 371 on memex-cloud.

Phase 2 of #748 (moving readers onto the satellite) never happened. Nothing reads the satellite in core src/, in MeshWeaver.Plugins origin/main or in the samples. Every gate and view reads the NodeType node itself, including Plugins' Store/Publishing ReadCompileState.

Evidence (memex-cloud, 2026-09-23, images that contain #5327/#5375)

instrument	reading
`Admin/_LogIncident/f7d8f9982b20cce5` — placement timeouts (#5334)	9 of 10 retained samples at 14:22Z are `messagehub/*/_Activity/compile-state`
`Admin/_LogIncident/df1ef8b39a2d8601` — `QueryProviderStalledException` at activation (#5531)	6 of 7 retained samples are compile-state satellites
back-pressure crossings filed after 10:00Z (51, parsed from issue bodies)	the oldest-leg family is compile-state in 8 of 51, second only to `cache/*` (15)

Plugins' gate had already measured the same writes as the per-package nodeops tax (#4141, folded into #2543 without an answer).

Change

MeshDataSource no longer installs the mirror.
NodeTypeCompileStateMirror.Install is [Obsolete], not deleted. In-mesh callers are invisible to CI, so a deletion could break them at runtime compile.
StatePath, StateNode and Parse stay. Satellites already on disk are now inert records.
Doc: Doc/Architecture/CompileStateSatelliteRetired, added to the Architecture topic map.

Verification

NodeTypeCompileStateMirrorTest.NodeTypeActivationAndStateChange_WriteNoCompileStateSatellite runs on a real mesh with two positive controls:
- the state change does land on the NodeType node;
- the same reader does see a satellite that was written directly.
The test passes with the fix.
Negative control: with the Install call restored in MeshDataSource, the test fails on its assertion ("a NodeType activation must not write MirrorSpace/Widget/_Activity/compile-state").
dotnet build -c Release -warnaserror, run one project at a time on MeshWeaver.Compiler.Pipeline, MeshWeaver.Graph, MeshWeaver.Graph.Test and MeshWeaver.Documentation.Test: 0 errors on each.
ArchitectureTopicMapTest and DocumentationLinkIntegrityTest pass (5/5).

Not established

How much of the back-pressure this removes. The largest family, cache/*, and the per-user _Install/_Access/_Entitlements legs are untouched.
Whether roll-time placement timeouts also need a departing silo's directory churn.

The post-roll read that settles #5389, #5037 and #5334: the next boots on an image carrying this change should show no */_Activity/compile-state destinations in placement timeouts, activation faults or back-pressure oldest legs.

No recycle is needed; this takes effect as hubs activate on the new image.

Refs #5389 #5037 #5334 #5531 #4141 #748

🤖 Generated with Claude Code

…ads (#5389) Every NodeType hub activation installed NodeTypeCompileStateMirror, which wrote {type}/_Activity/compile-state: one CreateOrUpdateNodeRequest per activation through the node-operation hub, and one extra grain activation per NodeType whenever the state moved (every framework roll re-stamps every type). Phase 2 of #748 (moving readers onto the satellite) never happened, so nothing in core, Plugins or the samples reads it. On memex-cloud, running images that contain #5327/#5375, those satellites were the destinations of the boot-time placement timeouts (#5334, 9 of 10 samples), the stalled-resolution activation faults (#5531, 6 of 7) and 8 of 51 back-pressure oldest legs. MeshDataSource no longer installs the mirror. Install stays as an [Obsolete] member so in-mesh callers keep compiling. Test pinned with positive controls on both sides; a negative control with the install restored fails the assertion. Refs #5389 #5037 #5334 #5531 #4141 Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Copilot

Copilot review overview

🟡 Changes recommended

Address the documentation inconsistency and revise the negative test's missing-node check.

Get a fresh assessment by requesting another Copilot review.

Review effort: Lite
Findings: 1 · 1

Open (2)

Negative test point-read triggers missing-node storm breaker · New
Other architecture pages still document retired satellite as live · New

What changed in this PR

Retires the unused compile-state satellite mirror to reduce NodeType activation and routing overhead.

Changes:

Stops installing the mirror from MeshDataSource.
Retains compatibility helpers and marks installation obsolete.
Adds regression coverage and retirement documentation.

File	Summary	Review notes
`test/MeshWeaver.Graph.Test/NodeTypeCompileStateMirrorTest.cs`	Verifies no satellite is written.	Moderate: avoid repeated point-reads for the negative check.
`src/MeshWeaver.Graph/MeshDataSource.cs`	Removes mirror installation.	No review comment.
`src/MeshWeaver.Documentation/Data/Architecture/CompileStateSatelliteRetired.md`	Documents retirement and evidence.	Nit: update other architecture pages that still describe the satellite as active.
`src/MeshWeaver.Documentation/Data/Architecture.md`	Adds the architecture topic entry.	No review comment.
`src/MeshWeaver.Compiler.Pipeline/NodeTypeCompileState.cs`	Retains helpers and obsoletes installation.	Moderate: clarify that the write-discipline documentation is historical.

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

rbuergi · 2026-09-23T18:24:55Z

+        var satellite = await Observable.Interval(TimeSpan.FromMilliseconds(250)).StartWith(0L)
+            .Take(20)
+            .SelectMany(_ => ReadNode(statePath))
+            .Where(n => n is not null)
+            .FirstOrDefaultAsync()


Agreed, fixed in 6c4ff20. Both the negative check and the instrument's positive control now read existence off a path:{type}/_Activity scope:children listing (the Initial snapshot). A point read is now used only for the NodeType node, which is known to exist.

The positive control proves the listing sees a satellite that was written directly. I re-ran the negative control with the listing instrument: with Install restored in MeshDataSource, the test fails on the assertion. With the fix it passes.

rbuergi · 2026-09-23T18:24:56Z

+Batching would have kept a write that has no reader. So `MeshDataSource` no longer installs the
+mirror, and `NodeTypeCompileStateMirror.Install` is `[Obsolete]` rather than deleted. A caller in
+in-mesh source is invisible to every CI build, so deleting the method would break that caller at
+runtime compile. The value shape (`StatePath`, `StateNode`, `Parse`) stays, so satellites that
+already exist can still be read. They are now inert records.


Agreed, fixed in 6c4ff20. These pages now say the satellite is retired, link CompileStateSatelliteRetired, and point readers at the NodeType node:

ExecuteTimeInterlock

NodeTypeCompilation

MissingDeclaredSources

ReleaseFailureClasses

RebakeWaves (now in the past tense)

AParkedTypeNamesTheImport

Mentions that are historical (incident timelines, DisposalStallVerdicts, BakeSealNodeOpsSaturation) are left as they are. ArchitectureTopicMapTest and DocumentationLinkIntegrityTest pass (5/5).

github-actions · 2026-09-23T18:21:48Z