57/F2b: A settings layer no longer replaces what a provider serves (issue 696) #321
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: novox/mesh-controller#321
Reference in New Issue
Block a user
Work package F2b of hq design 57, issue 696, ADR 0348 clause 11. It finishes the part F2 (#316) left:
Settlewas unchanged there because influxdb on ace relied on it.Merge order: mesh-catalog#290 (influxdb serves
${setting:org}and${setting:bucket}) first, then this pull request. Merged before #290 is delivered, grafana's data source on ace would name the literal bucketdefaultin place of the bucket ace's layer sets.What changed
catalogue.Settleno longer lays a layer over served keys. It only fills${setting:<key>}in served values from the layers, with module defaults lowest (filledFromSettings, shared withoverridden). A key a layer sets that the provider serves as a literal changes nothing a consumer is told, and a setting still never adds a served key.overriddenkeeps its override for contributions; that is not in this package's scope.UnusedSettingsno longer counts a literal served key as a place where a setting lands. A served key that asks for${setting:k}still counts, throughsettingKeysUsedBy. A layer that sets a literal served key is now reported as reaching nothing, instead of being accepted silently.Which live layers relied on the override
Only influxdb on ace. I read the settings layer, by key, of every provider assigned on ace, novox, g14 and shanks. That covered the mesh-wide and per-machine layers of distribution, gitea, influxdb, keycloak, mailu, minio, mongodb, mosquitto, mssql, postgres, redis, route-proxy, step-ca and umami. I compared each layer's keys with what the module serves:
domain) and keycloak (issuer) serve${setting:<same key>}, so the value is the same with or without the override.ports,exposeorendpoints. Settle skippedportsalready, and the other two are not served keys.ZoneOn, which also calls Settle, is unaffected. Zones are${setting:zone}in the tests.In the tests, three fixtures encoded the old override. step-ca serving an empty
rootthat a layer replaced: three co-located tests now usestepCAWithItsRoot, which serves${setting:root}. A store whose layer replaced the derived bucket:TestASettingComposesWithADerivedValuenow serves${setting:bucket-prefix}${consumer:as:dns}.TestASettingOverridesAServedKeyAndAddsNoneasserted the override and is replaced byTestALayerNoLongerReplacesAServedKey. Each keeps its intent: an operator's value reaches what is served where the definition asks for it.How it is proven
TestGrafanaComposesTheSameOnceInfluxdbServesItsSettings: the seed is like ace. influxdb's layer holds org, bucket (made-up values) and an endpoints block, and grafana consumes influxdb-api with its data source file copied verbatim from the catalogue. Every grafana resource composed, including the data source file and the binding file, is compared byte for byte with golden files intestdata/influxdb-served-by-setting/. The golden files were composed by this branch's base (71362c1, Settle still overriding) from influxdb serving the literals. The test covers a machine with no layer as well. It passes before this commit (catalogue step alone) and after it (both steps).TestALayerNoLongerReplacesAServedKey: Settle leaves a literal served key alone and adds no key.UnusedSettingssays such a layer reaches nothing. influxdb as it was (literals) under an ace-like layer no longer passes the layer's bucket to grafana, which shows why the catalogue must go first. It fails before this commit.TestInfluxdbsConfigTakesItsOrgAndBucket: influxdb's config.json takes org and bucket from an ace-like layer, and stays{}without a layer, because defaults never reach the file. Note that config.json on ace drops theendpointskey that today's form wrote there; see mesh-catalog#290.Commands run:
gofmt -l cmd internal examples(empty),go vet ./...,go test ./....MESH_TEST_POSTGRES=postgres://…@127.0.0.1:<port>/postgres sh merge-check.shagainst a throwaway PostgreSQL exits 0, and every listed store test passes, both at the first base and rebased ondc36536(after V1, #317).Left for other packages
overriddenstill lets a layer replace a contribution's declared keys. That is not served, and was not in this package.${config:}. K5 (issue 664) moves ace's layer into mesh-config.Reviewer's checklist
UnusedSettingschange follows from it, because a setting that lands nowhere is said so.TestALayerNoLongerReplacesAServedKeyfails before this commit. The three changed fixtures are justified above.TestGrafanaComposesTheSameOnceInfluxdbServesItsSettingsagainst golden files from71362c1.What a provider serves is its definition's (ADR 0348 clause 11): a layer fills only the keys it asks for as ${setting:<key>}. Settle replaced any served key a layer named, so a setting for one of a provider's files could change what every consumer was told. influxdb on ace was the one live layer relying on it and now serves its settings (mesh-catalog, merged first), so grafana's data source composes byte for byte as before. UnusedSettings no longer counts a literal served key as a place a setting lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>Delivery
novox/mesh-controller@c5fe7d4795d7— delivered since 2026-10-14T05:16:18ZDelivery plan — builds nothing: the change touches no module of the mesh's graph
Group
pkg/57-f2b-served-keys, in order: novox/mesh-catalog@2dd6510ded → novox/mesh-controller@c5fe7d4795Transitions
The commit's note under
refs/notes/mesh-plankeeps every transition:git log --notes=mesh-plan.Verdict: APPROVE at
c5fe7d4795Independent review for issue 696 (a settings layer still overrides what a provider serves). Merge after mesh-catalog#290.
What I ran on
c5fe7d4:gofmt -l cmd internal examples: empty.go vet ./...: clean.MESH_TEST_POSTGRES=postgres://postgres@127.0.0.1:<port>/postgres sh merge-check.shagainst a throwaway PostgreSQL over TCP on 127.0.0.1 with mmap shared memory: exit 0, store tests included. I stopped the server afterwards.dc36536, the old Settle):TestALayerNoLongerReplacesAServedKeyfails there, on all three of its assertions.TestGrafanaComposesTheSameOnceInfluxdbServesItsSettingspasses there with both the old and the new influxdb. That means the golden files are what the old controller composes.TestInfluxdbsConfigTakesItsOrgAndBucketpasses there, because it tests the catalogue's change.What I checked:
ports,exposeorendpoints, or nothing. No catalogue module declares azone, soZoneOnis unaffected.${setting:}fill. That is the intended rule. The secret andprotectedrefusals insettlenever applied to a served map, so nothing is lost there.${setting:root}. The derived-value test still proves that the setting is filled before the consumer's half is derived. The deleted override test is replaced by its inverse.Findings. None of them blocks the merge, and all can follow:
overridden(settings.go:141) still lets a layer replace a contribution's declared keys. That is outside issue 696 and ADR 0348 clause 11, which are about served keys and the keys a file takes. Under ADR 0348 clauses 4 and 5, though, a value lands only in a template as${config:}or is read at runtime, so this override is a second instance of the same rule. I recommend its own issue, linked to K6 (issue 665), instead of leaving it as a line in this pull request's description.UnusedSettingsmessage (settings.go, "declares no %q in what it contributes") no longer says that a literal served key is not settable. A hint like "a served key is set only where the definition asks for ${setting:}" would explain the refusal to the person who sets it. This one is optional.