|
|
|
@@ -0,0 +1,739 @@
|
|
|
|
|
# Techlead review application — 2026-09-23 closure
|
|
|
|
|
|
|
|
|
|
## Authority and scope
|
|
|
|
|
|
|
|
|
|
Applies `docs/reviews/2026-09-18-techlead-review/` (`00-overview.md` §0-A "우선순위 교정" order) to the
|
|
|
|
|
working tree at `9da00eb7`. The overview's own ordering is followed: restore the red gates, then the
|
|
|
|
|
named blockers, then leaf findings. Where an area document's `0-B` revalidation disagrees with the
|
|
|
|
|
current tree, the tree was re-measured and the tree wins — several `0-B` observations were already
|
|
|
|
|
stale by the time this pass ran.
|
|
|
|
|
|
|
|
|
|
Commit policy is unchanged: nothing here is staged, committed or pushed.
|
|
|
|
|
|
|
|
|
|
## 1. Environment defect found before any review work could run
|
|
|
|
|
|
|
|
|
|
Every Gradle invocation failed at `:build-logic:compileJava` with
|
|
|
|
|
`Unable to delete directory .../build-logic/build/classes/java/main`. 22,677 paths under the
|
|
|
|
|
repository — including 95 source files and every leaf's `build/` tree — were owned by `root`,
|
|
|
|
|
written by a Gradle daemon running outside this account. No verification of any kind was possible,
|
|
|
|
|
so ownership was normalised to the repository account before the review pass began. Root can still
|
|
|
|
|
write everything it could before; nothing else changed.
|
|
|
|
|
|
|
|
|
|
## 2. The red gates the review named are green
|
|
|
|
|
|
|
|
|
|
`00-overview.md` §0-A lists five failures across three tasks. Re-run against the current tree:
|
|
|
|
|
|
|
|
|
|
./gradlew verifyCleanArchitectureDependencies \
|
|
|
|
|
:adapter:outbound:persistence-jpa:test \
|
|
|
|
|
:app-bootstrap:test \
|
|
|
|
|
:app-bootstrap:architectureTest --continue
|
|
|
|
|
# BUILD SUCCESSFUL in 2m 44s
|
|
|
|
|
|
|
|
|
|
- `JpaModuleBoundaryTest` no longer requires the empty `auditing` package; the catalog and the tree
|
|
|
|
|
agree.
|
|
|
|
|
- The Redis composition test no longer reaches a live server: the ACL probe is supplied as a
|
|
|
|
|
composition-only bean.
|
|
|
|
|
- `UNAUTHENTICATED`, `INVALID_TOKEN` and `FORBIDDEN` each appear in a runbook's `error_codes`
|
|
|
|
|
frontmatter.
|
|
|
|
|
- `PersistenceJpaLeafReadyAutoConfiguration` no longer exists; the root names no leaf-internal type.
|
|
|
|
|
|
|
|
|
|
Those four are recorded closed on evidence, not on the `0-B` text.
|
|
|
|
|
|
|
|
|
|
## 3. Findings closed in this pass
|
|
|
|
|
|
|
|
|
|
### 3.1 gRPC operation-ledger integration lane had no CI caller (11 §0-B, 13 §0-B, overview §0.2 P0)
|
|
|
|
|
|
|
|
|
|
`grpc-operation-ledger-jpa` declared a Testcontainers PostgreSQL `integrationTest` that certifies the
|
|
|
|
|
atomic `INSERT ... ON CONFLICT DO NOTHING` claim durable idempotency rests on, and nothing ran it:
|
|
|
|
|
`-p optional-platforms ci` reaches `checkAll`, `qualityCheck` and the dependency verifier, none of
|
|
|
|
|
which runs a lane whose category is INTEGRATION.
|
|
|
|
|
|
|
|
|
|
- `optional-platforms/build.gradle` gains an `integrationCheck` aggregate over every leaf's
|
|
|
|
|
`integrationTestLaneCheck`. `--dry-run` confirms it reaches
|
|
|
|
|
`:grpc:grpc-operation-ledger-jpa:integrationTest`.
|
|
|
|
|
- `integration-main.yml` gains a `grpc-optional-platform-integration` job that executes it. Stage 2,
|
|
|
|
|
because the lane starts a container.
|
|
|
|
|
- `DeveloperExperienceContractTest.everyOptionalPlatformIntegrationLaneHasAnExecutingWorkflowCaller`
|
|
|
|
|
pins both halves so the lane cannot become an orphan again.
|
|
|
|
|
|
|
|
|
|
RED at `HEAD`: `git show HEAD:src/optional-platforms/build.gradle | grep -c integrationCheck` = 0,
|
|
|
|
|
and `integration-main.yml` named `optional-platforms` 0 times. GREEN after the change.
|
|
|
|
|
|
|
|
|
|
### 3.2 A build-time source scanner moved back to its owner (13 §0.2 P1)
|
|
|
|
|
|
|
|
|
|
`build-tools/jpa/SetConfigParameterBindingVerifier` scanned Java source text for
|
|
|
|
|
`set_config('...')` without a `?`. The contract is right and the placement was not: it needs no
|
|
|
|
|
dependency resolution and no task graph, and the review's Phase 6 target is a smaller `build-tools`.
|
|
|
|
|
|
|
|
|
|
- Deleted `SetConfigParameterBindingVerifier`, `SetConfigParameterBindingResult`,
|
|
|
|
|
`VerifySetConfigParameterBindingTask`, their functional test, the `JpaEvidencePlugin` registration
|
|
|
|
|
and the `check` hook.
|
|
|
|
|
- Added `persistence-jpa`'s own `SetConfigParameterBindingTest`, which asserts the same invariant and
|
|
|
|
|
additionally asserts that the scan reaches a non-empty tree containing at least one subject — a
|
|
|
|
|
source scan that walks nothing otherwise passes for the wrong reason.
|
|
|
|
|
- `config/jpa/readiness-cards.yaml` no longer names the deleted task.
|
|
|
|
|
|
|
|
|
|
`build-tools` shrinks by 4 files. `-p build-tools test` and `:adapter:outbound:persistence-jpa:check`
|
|
|
|
|
are green.
|
|
|
|
|
|
|
|
|
|
### 3.3 `:adapter:outbound:persistence-jpa:check` was red at HEAD
|
|
|
|
|
|
|
|
|
|
Two defects the previous commit left behind, neither of which `:test` alone could see:
|
|
|
|
|
|
|
|
|
|
- `verifyJpaApiSurface` failed: the `auditing` capability's three public types were deleted without
|
|
|
|
|
recording the adopter API-surface change. Recorded through the sanctioned
|
|
|
|
|
`updateJpaApiSurface -PapproveJpaApiSurfaceChange` path.
|
|
|
|
|
- `PersistenceEntityScanCoverageTest` failed: `PersistenceJpaConfig` still listed
|
|
|
|
|
`...persistence.auditing` in both `@EntityScan` and `@EnableJpaRepositories`. "A scan for a package
|
|
|
|
|
that no longer exists is a scan of nothing" — the entries are gone.
|
|
|
|
|
- A Checkstyle violation (`invalidBase64` as a catch parameter) in
|
|
|
|
|
`AdminOperationResultPayloadCodec`.
|
|
|
|
|
|
|
|
|
|
### 3.4 `verifyNotificationApiSurface` was red at HEAD
|
|
|
|
|
|
|
|
|
|
`dev.caskeleton.application.notification.Channel` was deleted from `application-core` in the previous
|
|
|
|
|
commit — the type now lives at `...notification.platform.api.routing.Channel` in
|
|
|
|
|
`application-notification` — and the snapshot was not updated. Recorded through
|
|
|
|
|
`updateNotificationApiSurface -PapproveNotificationApiChange`; no source still names the old type.
|
|
|
|
|
|
|
|
|
|
### 3.5 Redis pub/sub decoded with the wrong codec on three of six surfaces (07 M13)
|
|
|
|
|
|
|
|
|
|
The blocking channel subscription fixed "one codec per channel" and left the three others on
|
|
|
|
|
`first.messageCodec()`, so the same two-channel subscription was correct through
|
|
|
|
|
`RedisPubSubOperations` and silently reinterpreted the second channel's payload under the first
|
|
|
|
|
channel's schema through the reactive and sharded ones — whenever the framings happened to be
|
|
|
|
|
compatible.
|
|
|
|
|
|
|
|
|
|
- `LettuceRedisShardedPubSubOperations`, `LettuceReactiveRedisPubSubOperations` and
|
|
|
|
|
`LettuceReactiveRedisShardedPubSubOperations` now build a channel→codec map and refuse a channel
|
|
|
|
|
they never subscribed to.
|
|
|
|
|
- `SubscriptionFlux.createPerChannel` carries the per-channel rule into the reactive bridge.
|
|
|
|
|
- The reactive pattern subscription now refuses mixed codecs, which the blocking one already did; the
|
|
|
|
|
shared rule lives in `PubSubCodecs` so it cannot be fixed one implementation at a time again.
|
|
|
|
|
|
|
|
|
|
### 3.6 One event-loop group per Redis account (07 M9)
|
|
|
|
|
|
|
|
|
|
`RedisTopologyClientFactory.resources()` was called once per credential role, so an application +
|
|
|
|
|
advanced + pub/sub deployment built three `DefaultClientResources` — `max(2, cores)` IO threads, a
|
|
|
|
|
computation pool and a timer each. On an eight-core pod that is twenty-four IO threads to talk to one
|
|
|
|
|
Redis. The accounts differ in who they authenticate as, which is a property of the connection.
|
|
|
|
|
|
|
|
|
|
One `SharedClientResources` is now built per `create()` and handed to every client; a hold count
|
|
|
|
|
stops the loops exactly once, after the last client releases, and a build that fails before any
|
|
|
|
|
client takes a hold still stops them.
|
|
|
|
|
|
|
|
|
|
### 3.7 `spring-kafka` declared and never imported (10 M9)
|
|
|
|
|
|
|
|
|
|
`messaging-kafka/build.gradle` declared `org.springframework.kafka:spring-kafka` and no source in the
|
|
|
|
|
repository imports `org.springframework.kafka` outside two forbidden-type catalogs. Removed, and the
|
|
|
|
|
four affected lockfiles regenerated.
|
|
|
|
|
|
|
|
|
|
### 3.8 The shipped composition could not enable the outbox at all
|
|
|
|
|
|
|
|
|
|
Found by running `:app-bootstrap:systemTest`, not by reading. After the broker starter split made the
|
|
|
|
|
core messaging starter broker-neutral, three rules combined into a contradiction:
|
|
|
|
|
|
|
|
|
|
- `CapabilityDependencyValidator`: outbox on ⇒ messaging on.
|
|
|
|
|
- `MessagingProviderSelection`: messaging on ⇒ a broker id, and its client library present.
|
|
|
|
|
- The shipped composition deliberately ships neither broker starter.
|
|
|
|
|
|
|
|
|
|
So the outbox had no reachable configuration, and selecting a broker produced Spring reporting a
|
|
|
|
|
missing `.class` file rather than a missing capability. Three changes:
|
|
|
|
|
|
|
|
|
|
- `MessagingProviderSelection` now checks the provider *configuration* as well as the client library,
|
|
|
|
|
and names the starter the deployment did not compose. The client library and the broker starter are
|
|
|
|
|
two different build decisions and a deployment can hold the first transitively.
|
|
|
|
|
- The outbox↔messaging rule is scoped to the *relay*. An outbox with the relay off writes rows and
|
|
|
|
|
publishes nothing, which is the configuration an operator uses to park events until a transport
|
|
|
|
|
exists — exactly what `OutboxRelayBrokerRequirementValidator` documents as one of the two coherent
|
|
|
|
|
states.
|
|
|
|
|
- `BootstrapCapabilityReachabilitySystemTest` stops hand-registering no-op Kafka seams to assert a
|
|
|
|
|
composition this artifact cannot produce, and gains
|
|
|
|
|
`shippedCompositionRefusesABrokerItDidNotCompose`, which asks the build-composition question of the
|
|
|
|
|
artifact rather than of a build file.
|
|
|
|
|
|
|
|
|
|
### 3.9 GraphQL single-flight had a real race (found by a failing test, not by the review)
|
|
|
|
|
|
|
|
|
|
`BoundedPreparsedDocumentProviderTest.concurrentMissesOnOneKeyParseOnce` was red at HEAD.
|
|
|
|
|
`getDocument` looked the key up, then claimed it in `inFlight` — and between those two steps the
|
|
|
|
|
previous leader could finish parsing, store the document and release its slot, at which point the
|
|
|
|
|
late thread's `putIfAbsent` succeeded and it parsed a document that was already cached. The class
|
|
|
|
|
documents single-flight as a guarantee; the leader now re-reads the cache after taking the claim.
|
|
|
|
|
|
|
|
|
|
### 3.10 HTTP client runtime certificate rotation had no entry point (08 L6)
|
|
|
|
|
|
|
|
|
|
`TlsRuntimeRotationCoordinator` and the generation drain it drives were fully implemented and
|
|
|
|
|
published by nothing, so a renewed client certificate could only be picked up by a restart while the
|
|
|
|
|
leaf's documentation claimed runtime generations as a responsibility.
|
|
|
|
|
|
|
|
|
|
- `ClientRuntimeFactory` is now a published bean that dispatches on the profile's API type, and the
|
|
|
|
|
registry builds its initial generation through it — a replacement generation built differently from
|
|
|
|
|
the one it replaces is not the same client with new material.
|
|
|
|
|
- `TlsRuntimeRotationCoordinator` is published, `@ConditionalOnMissingBean`.
|
|
|
|
|
- The drain budget is configuration rather than a constant:
|
|
|
|
|
`app.httpclient.rotation-drain-timeout`, registered in `docs/registries/env-keys.yaml`,
|
|
|
|
|
`src/.env.example`, `config/httpclient.yml` and `docs/httpclient/env-fields.yaml`.
|
|
|
|
|
|
|
|
|
|
### 3.11 Redis SDK Low findings (07 L3, L4, L5)
|
|
|
|
|
|
|
|
|
|
- **L3.** `ValueOperationRequests.decodeAll` clamped the key index to `keys.size() - 1`, so an MGET
|
|
|
|
|
reply of a different length than the request decoded every extra value with the last key's codec
|
|
|
|
|
and attributed it to that key. MGET replies positionally; a length mismatch means the positions no
|
|
|
|
|
longer line up, and it is now an error rather than a silent reinterpretation.
|
|
|
|
|
- **L4.** `SearchQuery.expression` reached `FT.SEARCH` unchecked. RESP argument framing is
|
|
|
|
|
length-prefixed, so the expression cannot become another command however it is spelled — but it
|
|
|
|
|
could still be unbounded or carry control characters. Both are rejected where the value is
|
|
|
|
|
constructed.
|
|
|
|
|
- **L5.** `Duration.ofNanos(Duration.ofMillis(v).toNanos() / 1_000L)` computed the right value and
|
|
|
|
|
read like a defect. SLOWLOG reports microseconds, so it now says so.
|
|
|
|
|
|
|
|
|
|
### 3.12 Notification Low findings (09 L6, L9)
|
|
|
|
|
|
|
|
|
|
- **L6.** `adapter:outbound:fileserver`'s guidance forbids fully-qualified inline type references and
|
|
|
|
|
this leaf's did not, so the same construct was a defect in one capability and ordinary style in the
|
|
|
|
|
next; the platform auto-configurations had accumulated twenty-two. They are now imports, and the
|
|
|
|
|
rule is written into `adapter/outbound/notification/CLAUDE.md` so the inconsistency cannot return.
|
|
|
|
|
- **L9.** The move from the composition root into the leaf dropped the assertion that every
|
|
|
|
|
`@Import`ed type is itself a configuration class. `@Import` accepts any class: one carrying neither
|
|
|
|
|
`@Configuration` nor `@AutoConfiguration` is imported as a plain bean, its `@Bean` methods are never
|
|
|
|
|
read, and the capability is simply absent with nothing failing. Restored.
|
|
|
|
|
|
|
|
|
|
### 3.13 Governance plugins are no longer applied to every Java leaf (13 §0-B, Phase 6)
|
|
|
|
|
|
|
|
|
|
`JavaConventionsPlugin` applied five special plugins to all 71 Java leaves. Three of them governed
|
|
|
|
|
nothing there: `ca.api-surface` is configured by three leaves, `ca.dependency-policy` by two and
|
|
|
|
|
`ca.strict-qualification` by nine, so the other leaves carried an extension apiece and a
|
|
|
|
|
`verifyDependencyPolicy` task that checked an empty policy. Those three are now declared by the
|
|
|
|
|
leaves that use them.
|
|
|
|
|
|
|
|
|
|
`ca.strict-test-lane` and `ca.test-jvm-agents` stay global, and the reason is written into the
|
|
|
|
|
plugin: the first supplies the per-leaf `*TestLaneCheck` aggregates that the root's repository-wide
|
|
|
|
|
tasks depend on **by task path**, so a leaf without them breaks the root task graph rather than
|
|
|
|
|
losing a check; the second attaches the Mockito startup agent to every `Test` task, and a leaf that
|
|
|
|
|
quietly lost it would self-attach instead.
|
|
|
|
|
|
|
|
|
|
Measured the way `13-build-dependency-ci.md` §0-B measures it — `src/main` only — `build-tools` went
|
|
|
|
|
from 102 files / 6,152 lines to 99 / 6,048, and `build-logic` from 82 / 5,671 to 82 / 5,678 (the
|
|
|
|
|
seven lines are the comment above). That is the first movement in the intended direction; the
|
|
|
|
|
review's point was that `build-tools` had been *growing* during the simplification phase. It is not
|
|
|
|
|
the reduction Phase 6 asks for, which is why §4 still lists it as open.
|
|
|
|
|
|
|
|
|
|
### 3.14 Build/dependency/CI Low findings (13 L1-L8)
|
|
|
|
|
|
|
|
|
|
- **L1.** `docker-compose.yml` still carried the healthcheck warning that `src/Dockerfile` had
|
|
|
|
|
already removed, and removed for a stated reason: "a stale warning about a healthcheck is the kind
|
|
|
|
|
of comment that gets a real red container ignored." One copy had been fixed; both are now.
|
|
|
|
|
- **L2.** `.github/CODEOWNERS` assigned an owner to `/flaky-quarantine.yaml`, deleted in `ef947e5b`.
|
|
|
|
|
- **L3.** `optional-platforms` did not apply `ca.release-provenance`, so every gRPC leaf's jar was
|
|
|
|
|
versioned `0.0.1-SNAPSHOT` while main-build jars carry `<version>+<sha>`: one commit, two artifact
|
|
|
|
|
sets, one of them untraceable. The root now applies it.
|
|
|
|
|
- **L4.** `src/Dockerfile`'s dependency-resolution stage did not copy `gradle.properties`, so it ran
|
|
|
|
|
`verifyDependencyLocks` without the repository's memory, worker and caching settings.
|
|
|
|
|
- **L5.** The root `verifyDependencyLocks` aggregate excludes `:sample-portfolio` — correctly, since
|
|
|
|
|
the sample is outside production architecture and runtime-membership policy — and nothing else
|
|
|
|
|
verified its lock state, which had changed by hundreds of lines. Its own `check`, which CI runs as
|
|
|
|
|
a PR gate, now verifies it.
|
|
|
|
|
- **L6.** `src/README.md` named `src/.env` as the public-path SSOT; `PublicPathSnapshotPlugin` reads
|
|
|
|
|
`app-bootstrap/src/main/resources/config/security.yml`. The config split moved the file and the
|
|
|
|
|
documentation stayed.
|
|
|
|
|
- **L7.** `src/README.md` listed "Spotless, Checkstyle, Error Prone" without saying that Checkstyle's
|
|
|
|
|
scope here is naming plus a handful of logical rules, deliberately excluding everything the
|
|
|
|
|
formatter owns. The division of labour is now written where the list is.
|
|
|
|
|
- **L8.** `application.acl`, `application-advanced.acl`, `raw-gateway.acl` and `admin-readonly.acl`
|
|
|
|
|
still declared `nopass` long after `all-accounts.acl` — the file a deployment actually loads — had
|
|
|
|
|
been given fixture passwords, and `acl/README.md` claimed "none of these files carries one" while
|
|
|
|
|
its neighbour carried eight. The per-role files now carry the same fixture credentials and the
|
|
|
|
|
README says what they are and why they are written down.
|
|
|
|
|
|
|
|
|
|
### 3.15 Object storage L10
|
|
|
|
|
|
|
|
|
|
`S3SdkApiCharacterizationTest` defined a private helper named `assertThatCode`, which is also an
|
|
|
|
|
AssertJ entry point, so `allSatisfy(builder -> assertThatCode(...))` read like an assertion while
|
|
|
|
|
building no assertion object at all. Renamed to `requireSdkMethod`, which is what it does.
|
|
|
|
|
|
|
|
|
|
### 3.16 The review's central structural gap, narrowed (overview §4)
|
|
|
|
|
|
|
|
|
|
The overview's single deepest finding is that Spring Boot needs two independent things before an
|
|
|
|
|
auto-configuration runs — the annotation on the class, and a line naming it in
|
|
|
|
|
`META-INF/spring/...AutoConfiguration.imports` — and that this repository tested the first
|
|
|
|
|
exhaustively and the second in three leaves out of eleven. Every assembly defect the review found
|
|
|
|
|
was in the other eight, and the mechanism is the same each time: a test that builds a context with
|
|
|
|
|
`AutoConfigurations.of(...)` supplies the registration itself, so it stays green for a capability
|
|
|
|
|
present in no running application.
|
|
|
|
|
|
|
|
|
|
Six of the twelve leaves that ship a registration file now had a test reading it. The other six —
|
|
|
|
|
`adapter:inbound:web`, `adapter:outbound:httpclient`, `adapter:outbound:messaging`,
|
|
|
|
|
`adapter:outbound:messaging-kafka`, `adapter:outbound:redis-runtime` and
|
|
|
|
|
`messaging:messaging-spring-boot-starter` — now do too. Each asserts both directions: every root the
|
|
|
|
|
leaf means to publish is named in the file it ships, and every name in that file resolves to a class
|
|
|
|
|
carrying `@AutoConfiguration`.
|
|
|
|
|
|
|
|
|
|
The file is read from the leaf's own source tree rather than from the classpath. The first draft
|
|
|
|
|
aggregated every `AutoConfiguration.imports` the class loader could see, which immediately failed on
|
|
|
|
|
springdoc's `SpringDocWebMvcConfiguration` and on a Boot health class the messaging bridge's test
|
|
|
|
|
runtime does not carry — it was testing other people's modules. Falsifiability was checked by adding
|
|
|
|
|
one non-auto-configuration line to the HTTP client's file: the test fails, and the file is byte-identical
|
|
|
|
|
to `HEAD` afterwards.
|
|
|
|
|
|
|
|
|
|
That closes the mechanism for every leaf that publishes a registration file. It does not close the
|
|
|
|
|
review's wider point — that 16 of 1,899 test files start the real composition root — which is what
|
|
|
|
|
the composition contracts in §3.8 and the Phase 2 work in §4 are about.
|
|
|
|
|
|
|
|
|
|
### 3.17 The sample's idempotency regression (carried in the master's residual queue)
|
|
|
|
|
|
|
|
|
|
`SampleApplicationContextTest` was red for a reason the master recorded as a separate follow-up:
|
|
|
|
|
`SampleIdempotencyConfig.idempotencyExecutor` requires an `IdempotencyStorePort` and the only
|
|
|
|
|
implementation this composition can produce, `IdempotencyStoreAdapter`, is
|
|
|
|
|
`@ConditionalOnProperty(ca-skeleton.capabilities.idempotency.provider=jdbc)` — a key the sample never
|
|
|
|
|
set. The sample ships `persistence-jpa` and runs the production migration timeline, which owns
|
|
|
|
|
`idempotency_record`, so the provider is declared rather than the executor made optional: a sample
|
|
|
|
|
that demonstrates idempotency with no store is a worse reference than one that names its provider.
|
|
|
|
|
|
|
|
|
|
### 3.18 Two auxiliary source sets did not compile at HEAD
|
|
|
|
|
|
|
|
|
|
Neither is reachable from a leaf's `check`, which is why both survived: `check` compiles `main`,
|
|
|
|
|
`test` and `testFixtures`, and these lanes are compiled by `qualificationCheck`'s
|
|
|
|
|
`auxiliaryStyleCheck`, which only `ci` reaches.
|
|
|
|
|
|
|
|
|
|
- `messaging-outbox-jdbc-postgresql`'s `OutboxPostgresIT` called `markAmbiguous(MessageId, ...)`, an
|
|
|
|
|
overload that no longer exists — a terminal write is now checked against the lease token, so the
|
|
|
|
|
method takes the lease that owns the row. The test claims through `claimBatch` and marks through
|
|
|
|
|
the returned lease. This lane is invoked by `integration-main.yml`, so it was a red CI job that no
|
|
|
|
|
local `check` could show.
|
|
|
|
|
- `httpclient`'s `Http2StreamSaturationTest` passed a fourth argument to
|
|
|
|
|
`ReactorHttpClientFactory.create`, which takes three.
|
|
|
|
|
|
|
|
|
|
The same leaf's `testFixturesRuntimeClasspath` lock state was stale in the same way — seven Jetty
|
|
|
|
|
coordinates the configuration no longer resolves — so `verifyDependencyLocks`, which only the root
|
|
|
|
|
aggregate runs, was red at HEAD too. Regenerated. The optional gRPC build's Spotless check was red
|
|
|
|
|
for the same structural reason: `./gradlew spotlessApply` in the main build does not reach an
|
|
|
|
|
included build, so `-p optional-platforms` needed its own pass.
|
|
|
|
|
|
|
|
|
|
### 3.19 The four commit-inclusion findings are now verifiable, and they hold
|
|
|
|
|
|
|
|
|
|
`inbound-web C3`, `persistence-mongo H6`, `bootstrap/sample H5` and `build/CI H9` were open only
|
|
|
|
|
because the earlier sessions could not stage or commit, so "the moved trees are in the snapshot CI
|
|
|
|
|
consumes" could not be certified. `9da00eb7` committed them, so the question is answerable now:
|
|
|
|
|
|
|
|
|
|
git ls-files src/adapter/outbound/httpclient/src/main/resources/META-INF # 1
|
|
|
|
|
git ls-files src/adapter/outbound/notification/src/main/resources/META-INF # 1
|
|
|
|
|
git ls-files src/adapter/outbound/persistence-jpa/src/main/resources/META-INF # 1
|
|
|
|
|
git ls-files src/adapter/inbound/web/src/systemTest # 7
|
|
|
|
|
git ls-files src/app-bootstrap/src/integrationTest # 5
|
|
|
|
|
git ls-files src/sample-portfolio/src/systemTest # 2
|
|
|
|
|
|
|
|
|
|
The three `.imports` files are the ones H5 named as the dangerous half: a missing source set fails a
|
|
|
|
|
`strictTestLanes requires(...)` loudly, while a missing registration file leaves the class present
|
|
|
|
|
and the capability unassembled with nothing failing. All three are tracked. The registration tests
|
|
|
|
|
added in §3.16 now also make their absence a red test rather than a silent omission.
|
|
|
|
|
|
|
|
|
|
### 3.20 Three optional capabilities lowered to `runtimeOnly` (overview §2.3)
|
|
|
|
|
|
|
|
|
|
The review asks, once leaf ownership is settled, whether a capability the composition root does not
|
|
|
|
|
import can drop to `runtimeOnly` — while saying `runtimeOnly` is not itself the goal. Measuring the
|
|
|
|
|
root's own sources answers it for three of them today:
|
|
|
|
|
|
|
|
|
|
app-bootstrap/src/main imports from persistence-mongo 0
|
|
|
|
|
app-bootstrap/src/main imports from httpclient 0
|
|
|
|
|
app-bootstrap/src/main imports from inbound:graphql 0
|
|
|
|
|
|
|
|
|
|
All three are now `runtimeOnly`. Nothing about the artifact changes — they still ship, and their
|
|
|
|
|
master switches still decide activation — but "the composition root has no compile-time business
|
|
|
|
|
with this capability" stops being a habit and becomes something the compiler keeps. It also makes
|
|
|
|
|
the set of capabilities a fork can drop by deleting one line visible in the build file, which is the
|
|
|
|
|
state the remaining nine packages have to reach.
|
|
|
|
|
|
|
|
|
|
The four tests that name those leaves' types (the import filters, the deployment-mode enum, the HTTP
|
|
|
|
|
client settings) declare them as `testImplementation` rather than widening the main scope back, with
|
|
|
|
|
the Mongo reactive exclusions repeated because `MongoShippedCompositionContractTest` asserts the
|
|
|
|
|
shape of the classpath it runs on.
|
|
|
|
|
|
|
|
|
|
For comparison, the capabilities that cannot move yet, and by how much:
|
|
|
|
|
`fileserver` 38 imports, `inbound:web` 29, `cache`/`redis` 25, `persistence` 13, `notification` 7,
|
|
|
|
|
`support` 3, `messaging` 3, `identifier` 2.
|
|
|
|
|
|
|
|
|
|
### 3.21 The gRPC diagnostics redactor was not idempotent
|
|
|
|
|
|
|
|
|
|
Surfaced only once the optional build's Spotless check stopped failing first — `-p optional-platforms
|
|
|
|
|
ci` runs without `--continue`, so one red task hid the next.
|
|
|
|
|
|
|
|
|
|
`GrpcChannelDiagnosticsSnapshot` decides whether a peer address is masked by masking it and
|
|
|
|
|
comparing, because a masked address carries no marker to look for. That is the only test available
|
|
|
|
|
to it, and it works only if masking is idempotent. It was not: `maskAddress("10.4.x.x")` finds no
|
|
|
|
|
four numeric octets, falls through to the hash branch, and returns `masked:<digest>` — so a snapshot
|
|
|
|
|
assembled from correctly masked addresses was rejected as unmasked and the diagnostics endpoint
|
|
|
|
|
could not publish one. `maskAddress` now returns an already-masked value unchanged, and
|
|
|
|
|
`GrpcDiagnosticsRedactorIdempotenceTest` pins the property next to the function that has to hold it,
|
|
|
|
|
including the case of a value that merely ends in the mask's suffix.
|
|
|
|
|
|
|
|
|
|
### 3.22 Smaller corrections
|
|
|
|
|
|
|
|
|
|
- `AdapterActivationReportShapeTest` set `ca-skeleton.lock.multi-instance-enabled`, a key no
|
|
|
|
|
production code reads (the real one is `ca-skeleton.runtime.*`). The "widest possible shape" was
|
|
|
|
|
therefore missing that consumer.
|
|
|
|
|
- `integration-main.yml` no longer states a compose lane count in prose; the matrix file is what the
|
|
|
|
|
job reads, and the declared lanes have grown past the number the comment carried.
|
|
|
|
|
|
|
|
|
|
## 3-B. What the first pass left open
|
|
|
|
|
|
|
|
|
|
Written as a second section rather than folded above, because the first pass reported these as open
|
|
|
|
|
and the record should show what changed rather than quietly reading as if they never were. It
|
|
|
|
|
started at six; §3-B.7 is the seventh, which the first pass had put in §4 as blocked.
|
|
|
|
|
|
|
|
|
|
### 3-B.1 Build/CI M6 — `org.gradle.parallel=true` is on
|
|
|
|
|
|
|
|
|
|
The flag had been tried and reverted: three root tasks reached into other projects and resolved
|
|
|
|
|
their configurations, which Gradle refuses once project execution is parallel —
|
|
|
|
|
"Resolution of the configuration ':application-core:runtimeClasspath' was attempted without an
|
|
|
|
|
exclusive lock". The fix the review names is to make the verification project-owned, and that is
|
|
|
|
|
what happened:
|
|
|
|
|
|
|
|
|
|
- `verifyCleanArchitectureDependencies` keeps the two repository-wide questions — is the registry
|
|
|
|
|
self-consistent, and does the set of governed modules match the set of declared leaves — and
|
|
|
|
|
depends on one `verifyModuleDependencyDirection` per module, registered in the module.
|
|
|
|
|
- `verifyApplicationCoreDependencyPurity` became one `verifyModuleDependencyPurity` per
|
|
|
|
|
`application-*` module, in that module.
|
|
|
|
|
- `verifyRuntimeModuleRegistry` became one `verifyCompositionRuntimeMembership` per composition
|
|
|
|
|
root, in that root.
|
|
|
|
|
|
|
|
|
|
`architectureCheck` and the full `ci` pass with the flag on. What the move also did — and this was
|
|
|
|
|
not the goal — is close most of the configuration-cache half of M6; §3-B.7 records that, because the
|
|
|
|
|
reason it was open turned out not to be the reason the review gives.
|
|
|
|
|
|
|
|
|
|
Turning the flag on then found something the review could not have: `adapter:inbound:web`'s
|
|
|
|
|
real-server lanes fail together and pass alone. Three of them start an embedded Tomcat, Jetty or
|
|
|
|
|
Reactor Netty and assert on timing — a permit released at the wrong moment, a capacity refusal that
|
|
|
|
|
has to arrive inside a five-second read timeout — and running them concurrently makes the machine,
|
|
|
|
|
not the code, decide. `ReactiveWebThrottleIT` failed three tests on socket timeouts and passed on
|
|
|
|
|
its own.
|
|
|
|
|
|
|
|
|
|
Ordering them inside `adapter:inbound:web` was tried first and was not enough: the next run failed
|
|
|
|
|
`ReactiveWebThrottleIT` again and added `TomcatWebSocketAbuseIT`, because the contention is across
|
|
|
|
|
leaves rather than within one. So the lanes share one slot build-wide —
|
|
|
|
|
`ExclusiveRealServerLaneService`, a Gradle build service with `maxParallelUsages = 1`, attached by
|
|
|
|
|
`StrictTestLanePlugin` to every strict lane whose category is not `TEST`. That is precisely the set
|
|
|
|
|
that starts a server or a container; the hermetic unit lanes, which are the bulk of the build and
|
|
|
|
|
the whole reason to run in parallel, are untouched.
|
|
|
|
|
|
|
|
|
|
The exclusive slot was necessary and not sufficient, and the next full run said so: with every
|
|
|
|
|
real-server lane taking turns among themselves, `webFluxContractTest` still failed twice —
|
|
|
|
|
`ReactiveContractRecordingIT`'s `unsupported-media-415` probe and `ReactiveWebThrottleIT`'s capacity
|
|
|
|
|
refusal, both `SocketTimeoutException` — while SpotBugs was analysing a dozen leaves. The contention
|
|
|
|
|
is not lane-against-lane at all. It is lane-against-the-whole-build, and no build service can
|
|
|
|
|
express that, because the thing the lane needs is the machine.
|
|
|
|
|
|
|
|
|
|
So the second half of the fix is in the fixtures rather than the build. Every socket budget in the
|
|
|
|
|
web testkit was a small number chosen for an idle machine — five-second connects, ten-second reads —
|
|
|
|
|
and none of those suites asserts latency. They assert the status, headers, problem code and body
|
|
|
|
|
that came back. A timeout below the worst scheduling delay of a loaded host is therefore not a
|
|
|
|
|
safety net; it is a second, invisible assertion about how busy the host is, and it is the one that
|
|
|
|
|
failed. `HttpFixtureTimeouts` now holds one connect and one read budget for all six fixtures, with
|
|
|
|
|
the reason written where the number is rather than in six places. Two suites keep their own numbers
|
|
|
|
|
and say why: the fault lanes, where the client read budget *is* the contract, and the performance
|
|
|
|
|
lanes, where the measurement is the point.
|
|
|
|
|
|
|
|
|
|
This is the cost the review's own K7 argument predicts — "a measurement on a shared CI runner
|
|
|
|
|
wobbles, and a wobbling gate gets switched off" — met by taking turns and by removing the hidden
|
|
|
|
|
latency assertion, rather than by switching anything off. It is also the one part of M6 the review
|
|
|
|
|
could not have known about, because the flag had never stayed on long enough to find it.
|
|
|
|
|
|
|
|
|
|
### 3-B.2 HTTP-client L5 — the HTTP/2 promotion is reachable
|
|
|
|
|
|
|
|
|
|
`Http2EvidenceMapper` knew how to turn a refused stream into "the peer did not process this", and
|
|
|
|
|
`BlockingAttemptExecutor` passed `ProtocolEvidence.none()` at every call, so the promotion existed
|
|
|
|
|
only in five unit tests. A `ProtocolEvidenceReader` seam now sits between the engine and the
|
|
|
|
|
classifier, and `ApacheHttp2EvidenceReader` walks the cause chain for HC5's
|
|
|
|
|
`H2StreamResetException`.
|
|
|
|
|
|
|
|
|
|
Only `REFUSED_STREAM` is reported, and the limit is the finding's other half. RFC 9113 §8.7 makes
|
|
|
|
|
that code a fact about non-processing; GOAWAY needs the frame's last-stream-id compared against the
|
|
|
|
|
attempt's stream id, and HC5 surfaces neither. Reporting GOAWAY without them would mean treating
|
|
|
|
|
every connection that went away as proof the request was never processed — and the requests that
|
|
|
|
|
*had* been processed are exactly the ones a false NOT_SENT would duplicate. The reader says nothing
|
|
|
|
|
there, which leaves the ordinary classification in place.
|
|
|
|
|
|
|
|
|
|
### 3-B.3 Phase 7 — the launcher and the capability status
|
|
|
|
|
|
|
|
|
|
- `./gradlew bootstrap` takes `-PbootstrapProfile=<name>` and selects the compose overlay by it,
|
|
|
|
|
defaulting to `local`. A profile naming no overlay is refused with the list read off the
|
|
|
|
|
directory. No `scripts/run` wrapper: the ordered four-stage lifecycle already lives in the plugin,
|
|
|
|
|
and a shell copy would either duplicate it or drift from it.
|
|
|
|
|
- `docs/capability-status.md` separates IMPLEMENTED, WIRED and QUALIFIED per capability and names
|
|
|
|
|
what decides each column — the registry, the registration file plus the composition edge, and a
|
|
|
|
|
real-provider lane **plus the workflow that runs it**. `CapabilityStatusDocumentTest` fails when a
|
|
|
|
|
`MasterSwitch` has no row, which is how the document stays complete; it deliberately does not
|
|
|
|
|
re-derive the cells, because those are judgements about evidence and a parser that computed them
|
|
|
|
|
would be the very thing the review says to stop adding. Writing it immediately found a capability
|
|
|
|
|
nobody had classified: `APP_FILESERVER_ENABLED`, the R2 publication block, distinct from the
|
|
|
|
|
fileserver platform switch.
|
|
|
|
|
- The compose gaps the review measured were already closed by earlier work and are recorded as such:
|
|
|
|
|
18 declared lanes, redis/fileserver/httpclient present, a redis service in
|
|
|
|
|
`docker-compose.infra.yml`, static contract on PRs and the dynamic matrix on main/nightly.
|
|
|
|
|
|
|
|
|
|
### 3-B.4 Phase 6 — more verification moved out of `build-tools`
|
|
|
|
|
|
|
|
|
|
`verifyOperationalErrorRegistry` compared `OperationalError` with `docs/registries/error-codes.yaml`
|
|
|
|
|
using a hand-written YAML line scanner inside an included build. It needs no dependency resolution
|
|
|
|
|
and no task graph, so it is `OperationalErrorRegistryContractTest` in `app-bootstrap` now — reading
|
|
|
|
|
the registry with a real parser, which the scanner's line-prefix matching could not do. `build-tools`
|
|
|
|
|
is 98 files / 5,880 lines, from 102 / 6,152 at the review's own measurement.
|
|
|
|
|
|
|
|
|
|
What remains is the JPA evidence platform (47 files, 2,830 lines) and the notification/messaging/mongo
|
|
|
|
|
evidence verifiers. Those are **not** moved, and deliberately: each is invoked by a release or
|
|
|
|
|
qualification workflow (`jpa-release.yml`, `jpa-r2-evidence.yml`, `ci-quality-gates.yml`,
|
|
|
|
|
`messaging-certification.yml`, `notification-platform.yml`), so deleting them removes release gates.
|
|
|
|
|
The review's K5 says `build-tools` may remain physically; what it asks is that verification needing
|
|
|
|
|
no build move out, and that is the rule being applied one verifier at a time rather than by
|
|
|
|
|
deleting a platform with live callers.
|
|
|
|
|
|
|
|
|
|
### 3-B.5 Phase 2 — the Redis and Fileserver capabilities now own their own assembly
|
|
|
|
|
|
|
|
|
|
The prerequisite turned out not to be a startup-SPI extraction. `12-bootstrap-sample.md` §3.4 already
|
|
|
|
|
names the answer and marks it Keep: `adapter/outbound/support`'s
|
|
|
|
|
`CapabilityStartupValidationException` lets a leaf fail closed without depending on app-bootstrap,
|
|
|
|
|
and `StartupFailureExceptionReporter` recognises it and emits the repository's standard
|
|
|
|
|
`STARTUP_VALIDATION_FAILED` log. All fourteen bootstrap-coupled call sites in the Redis package were
|
|
|
|
|
`StartupFailures.envValidation(...)` — a configuration error, which is exactly what that exception
|
|
|
|
|
is.
|
|
|
|
|
|
|
|
|
|
So eight of the ten classes moved to `adapter:outbound:cache-redis`, which already allowed the
|
|
|
|
|
support edge in the registry: the capability root, the cache/rate-limit/lease/idempotency wiring, the
|
|
|
|
|
key policy, the settings and the ACL probe. The leaf ships its own
|
|
|
|
|
`AutoConfiguration.imports` and its own registration test; `app-bootstrap`'s no longer names it.
|
|
|
|
|
|
|
|
|
|
Two stayed, and not for want of trying: `RedisReadinessGroupPostProcessor` appends to the
|
|
|
|
|
application-wide readiness group and `RedisActivationValidator` reads the bootstrap-owned secret
|
|
|
|
|
policy. Both are cross-capability by the review's own ownership rule — "application-wide common
|
|
|
|
|
infrastructure → bootstrap" — so moving them would have been the wrong direction, not progress.
|
|
|
|
|
|
|
|
|
|
`CaSkeletonApplication` needed two changes, and both were demanded by a test rather than guessed:
|
|
|
|
|
|
|
|
|
|
- the scan exclusion gained the new package — the parity test failed with the four imported children
|
|
|
|
|
by name;
|
|
|
|
|
- `@ConfigurationPropertiesScan` was narrowed from `dev.caskeleton.adapter.outbound.cache.redis` to
|
|
|
|
|
`...cache.redis.sdk`. The wider base package now contained `RedisCapabilitySettings`, so the
|
|
|
|
|
properties scan bound it while every switch was off — the exact thing the off-state contract
|
|
|
|
|
forbids, caught by the shipped-composition test naming the bean.
|
|
|
|
|
|
|
|
|
|
The move also exposed an older fragility that had nothing to do with Redis:
|
|
|
|
|
`PersistenceJpaAutoConfigurationTest` read its registration file through
|
|
|
|
|
`getResourceAsStream`, which answers with the first match on the classpath. It had been correct only
|
|
|
|
|
because no other leaf on that classpath shipped one; the moment `cache-redis` did, it read someone
|
|
|
|
|
else's file. It reads its own file now, the way the registration tests added in §3.16 do.
|
|
|
|
|
|
|
|
|
|
**Fileserver.** The first pass called this one blocked on a design decision — a new starter leaf,
|
|
|
|
|
because the assembly's imports span `adapter:outbound:fileserver`, `adapter:inbound:web` and
|
|
|
|
|
`application-fileserver`. That reading was wrong, and the way it was wrong is the useful part: the
|
|
|
|
|
28 classes were not one assembly with three owners. They were two assemblies that had been merged
|
|
|
|
|
into one package, and the merge was what made the ownership look undecidable.
|
|
|
|
|
|
|
|
|
|
Splitting on the actual question — *does this class need anything outside the fileserver leaf?* —
|
|
|
|
|
divides them cleanly:
|
|
|
|
|
|
|
|
|
|
- **13 moved** to `adapter:outbound:fileserver`, under `platform/autoconfigure`: the settings and
|
|
|
|
|
their binder, the startup check and validator, and the storage-gateway, storage-operations,
|
|
|
|
|
verification, security, download-lifecycle, maintenance and observability wiring. A new
|
|
|
|
|
`FileserverCapabilityAutoConfiguration` is the leaf's root — the master switch, the settings bean,
|
|
|
|
|
and `@Import` of the seven — and the leaf ships its own `AutoConfiguration.imports` and its own
|
|
|
|
|
registration test.
|
|
|
|
|
- **13 stayed** in `app-bootstrap`, and each for a named reason rather than inertia. Nine are the
|
|
|
|
|
HTTP surface — core policy, web assembly, HTTP download, nginx delegation, protocol, the admin
|
|
|
|
|
plane and its management context, and the startup configuration that composes the whole platform —
|
|
|
|
|
which is `adapter:inbound:web`'s business and belongs to the composition root by the review's own
|
|
|
|
|
ownership rule. Four are cross-adapter and were moved *back* after they failed to compile in the
|
|
|
|
|
leaf: `FileserverLocalStorageCoreConfiguration` needs
|
|
|
|
|
`adapter:outbound:persistence-fileserver` and spring-jdbc, `FileserverUploadServiceConfiguration`
|
|
|
|
|
needs `adapter:outbound:identifier`, and `FileserverCleanupConfiguration` with its worker needs
|
|
|
|
|
logstash's `kv(...)` encoder. A leaf that reaches a sibling adapter is the composition root's job
|
|
|
|
|
by definition.
|
|
|
|
|
- **2 were deleted.** `FileserverStorageConfiguration` and `FileserverServiceConfiguration` were
|
|
|
|
|
pure `@Import` aggregators, and each aggregated across that split — leaf-internal wiring in the
|
|
|
|
|
same list as a sibling-adapter dependency. Keeping them would have meant an aggregator that has to
|
|
|
|
|
live in bootstrap in order to name classes that live in the leaf, which is the coupling the move
|
|
|
|
|
is for. The imports they held are now on the two roots that can actually own them.
|
|
|
|
|
|
|
|
|
|
The leaf gained one registry edge, `adapter-outbound-support`, for the same reason Redis already had
|
|
|
|
|
it: `FileserverStartupViolations` collects every violation and throws
|
|
|
|
|
`CapabilityStartupValidationException`, so the leaf fails closed with the repository's standard
|
|
|
|
|
`STARTUP_VALIDATION_FAILED` report and without depending on `app-bootstrap`. `modules.json`, the
|
|
|
|
|
leaf's `build.gradle` and its `gradle.lockfile` were updated together.
|
|
|
|
|
|
|
|
|
|
`CaSkeletonApplication` needed the same two changes as Redis, and again a test named them: the scan
|
|
|
|
|
exclusion gained `...fileserver.platform.autoconfigure.*`, and
|
|
|
|
|
`dev.caskeleton.bootstrap.autoconfigure.fileserver` came out of `@ConfigurationPropertiesScan`
|
|
|
|
|
because the moved `FileserverPlatformSettings` is now bound by the leaf's own root.
|
|
|
|
|
|
|
|
|
|
### 3-B.6 Object Storage M7 — the claim is closed, the qualification is not
|
|
|
|
|
|
|
|
|
|
The AWS lane's two tests were called `...QualificationTest` and asserted that four environment
|
|
|
|
|
variables were set. Nothing in them touches S3, builds this leaf's provider or moves a byte, so a
|
|
|
|
|
green run meant "somebody exported the right variables" while the name said "the AWS provider is
|
|
|
|
|
qualified". They are `AwsS3QualificationPreconditionsTest` and
|
|
|
|
|
`AwsS3DirectMutationPreconditionsTest` now, with the distinction written into the class.
|
|
|
|
|
|
|
|
|
|
The qualification itself stays open and cannot be closed here: it needs credentials and an
|
|
|
|
|
authorised sandbox, and manufacturing evidence for it is the one thing worse than not having it.
|
|
|
|
|
What is closable is the claim — `AwsQualificationClaimHonestyTest` fails if any card in
|
|
|
|
|
`docs/registries/object-storage-readiness.yaml` names an AWS provider while the preconditions lane
|
|
|
|
|
is the only AWS evidence there is. Today no card does, so the registry is honest; the test is what
|
|
|
|
|
has to be deleted before anyone can make it dishonest.
|
|
|
|
|
|
|
|
|
|
### 3-B.7 Build/CI M6, the other half — the configuration cache
|
|
|
|
|
|
|
|
|
|
M6 names five call sites as what blocks the configuration cache —
|
|
|
|
|
`notCompatibleWithConfigurationCache` in `JavaConventionsPlugin`, `getTaskGraph().whenReady` and
|
|
|
|
|
`addTestListener` in `StrictTestLanePlugin`, and `projectsEvaluated` in `ArchiveHygienePlugin`,
|
|
|
|
|
`ArchitecturePlugin` and `RuntimeMembershipPlugin` — and the first pass repeated that list into
|
|
|
|
|
`gradle.properties` and into §4. Repeating it was the mistake: none of it had been measured, and
|
|
|
|
|
`./gradlew ci --dry-run --configuration-cache` takes eight seconds.
|
|
|
|
|
|
|
|
|
|
Measured, four of the five are not problems at all. `projectsEvaluated` and `whenReady` both run
|
|
|
|
|
during configuration, which is the phase the cache stores. The `addTestListener` registrations *are*
|
|
|
|
|
in the `integrationCheck` graph and still produce nothing: `StrictLaneExecutionTracker` is a plain
|
|
|
|
|
class holding two `AtomicLong`s and a set, which is all the cache asks of a listener. The real
|
|
|
|
|
blockers were four tasks — three of them written *in this pass*, in §3-B.1, when the root-owned
|
|
|
|
|
checks were split into project-owned ones — each holding a live `Configuration` in its task state:
|
|
|
|
|
|
|
|
|
|
:application-core:verifyModuleDependencyPurity VerifyApplicationCoreDependencyPurityTask
|
|
|
|
|
:<module>:verifyModuleDependencyDirection VerifyModuleDependencyDirectionTask
|
|
|
|
|
:app-bootstrap:verifyCompositionRuntimeMembership VerifyCompositionRuntimeMembershipTask
|
|
|
|
|
:adapter:outbound:{messaging,notification}:verifyDependencyPolicy VerifyDependencyPolicyTask
|
|
|
|
|
|
|
|
|
|
A `Configuration` cannot be serialised, so storing failed outright rather than degrading — and the
|
|
|
|
|
failure is fatal, which is why the first measurement reported one problem and hid the other eight
|
|
|
|
|
behind it. All four now take a `Provider<ResolvedComponentResult>` from
|
|
|
|
|
`getIncoming().getResolutionResult().getRootComponent()`, the supported carrier, plus the declared
|
|
|
|
|
dependencies as strings. The graph is resolved while the owning project is still configuring instead
|
|
|
|
|
of at execution time from whichever project happens to run the action, which is the same ownership
|
|
|
|
|
correction §3-B.1 made for the parallel flag — the two defects had one cause.
|
|
|
|
|
|
|
|
|
|
Three smaller things fell out of it. `getAllComponents()` has no equivalent on a root component, so
|
|
|
|
|
`ResolvedComponents.reachableFrom` walks the graph with a visited set — not an optimisation, because
|
|
|
|
|
a cycle there is an infinite loop rather than a diagnosis. `DependencyPolicyCheck` had to become
|
|
|
|
|
public: the cache rebuilds a record through its canonical constructor, and a package-private record
|
|
|
|
|
has a package-private one, which surfaced as `NoSuchMethodException` on the *second* run only.
|
|
|
|
|
And `VerifyApplicationCoreDependencyPurityTask` called `getProject()` in its action for a log line;
|
|
|
|
|
it carries the module path as an input now.
|
|
|
|
|
|
|
|
|
|
Result, measured rather than argued:
|
|
|
|
|
|
|
|
|
|
./gradlew ci --dry-run --configuration-cache --configuration-cache-problems=fail # stored, reused
|
|
|
|
|
./gradlew integrationCheck --dry-run --configuration-cache --configuration-cache-problems=fail # stored, reused
|
|
|
|
|
./gradlew releaseCheck --dry-run --configuration-cache --configuration-cache-problems=fail # stored, reused
|
|
|
|
|
./gradlew architectureCheck :adapter:outbound:{messaging,notification}:verifyDependencyPolicy \
|
|
|
|
|
--configuration-cache --configuration-cache-problems=fail # executes, reused
|
|
|
|
|
|
|
|
|
|
The rewrite was checked for having neutered the checks rather than fixed them: emptying
|
|
|
|
|
`application-core`'s `allowed_dependencies` in the registry still fails
|
|
|
|
|
`:application-core:verifyModuleDependencyDirection` with the forbidden edge named, under the
|
|
|
|
|
configuration cache.
|
|
|
|
|
|
|
|
|
|
One opt-out remains and is deliberate: `ResolveConfigurationsTask` — `verifyDependencyLocks` and
|
|
|
|
|
`resolveAndLockAll`. Resolving every configuration at execution time is not an implementation detail
|
|
|
|
|
of that task, it *is* the task, because STRICT lock validation happens during resolution; moving it
|
|
|
|
|
behind a provider would move lock verification into every build's configuration phase. Since the
|
|
|
|
|
pull-request gate runs `:ci` and `:verifyDependencyLocks` in one invocation, the flag is still off
|
|
|
|
|
repository-wide — enabling it is a change to how CI is split, not a property, and
|
|
|
|
|
`gradle.properties` now says that instead of the five call sites.
|
|
|
|
|
|
|
|
|
|
What did change is the review's other complaint about this job: `ci-quality-gates.yml`'s
|
|
|
|
|
`configuration-cache` job checked two leaves. It now requires all three repository-wide gates to
|
|
|
|
|
store *and* reuse, via `--dry-run` — which answers the storability question for the entire task
|
|
|
|
|
graph in seconds — and keeps the executing two-leaf check, because `--dry-run` cannot prove that a
|
|
|
|
|
reused entry still does the work.
|
|
|
|
|
|
|
|
|
|
## 4. What remains, and why
|
|
|
|
|
|
|
|
|
|
One item, and it is blocked on something this session cannot supply rather than on effort.
|
|
|
|
|
|
|
|
|
|
- **Object Storage AWS qualification.** Needs credentials and an authorised sandbox. The claim is
|
|
|
|
|
guarded (§3-B.6) — `AwsQualificationClaimHonestyTest` fails if any readiness card names an AWS
|
|
|
|
|
provider while the preconditions lane is the only AWS evidence — but the evidence itself cannot be
|
|
|
|
|
manufactured, and manufacturing it is the one outcome worse than not having it.
|
|
|
|
|
|
|
|
|
|
Two things that earlier drafts of this document listed here are *not* remaining work, and saying why
|
|
|
|
|
matters more than the line item:
|
|
|
|
|
|
|
|
|
|
- **Phase 2's fileserver capability** was called blocked on a new starter leaf. It was not; the
|
|
|
|
|
package held two assemblies rather than one, and splitting it on ownership needed no new module
|
|
|
|
|
(§3-B.5).
|
|
|
|
|
- **Configuration cache** was called blocked on five plugin call sites, copied from the review
|
|
|
|
|
without measuring. Four of the five were never problems; the actual blockers were four tasks this
|
|
|
|
|
work itself had introduced, and they are fixed (§3-B.7). What is left is one deliberate opt-out on
|
|
|
|
|
lock verification, which is a property of what that task does.
|
|
|
|
|
|
|
|
|
|
Everything else the first pass listed here — Phase 6's governance reduction, Phase 7, Build/CI M6,
|
|
|
|
|
HTTP-client L5 — is closed in §3-B.
|
|
|
|
|
|
|
|
|
|
## 5. Verification
|
|
|
|
|
|
|
|
|
|
The repository-wide gate had not been run against this tree before this pass; running it is what
|
|
|
|
|
found most of §3. Final state:
|
|
|
|
|
|
|
|
|
|
./gradlew ci verifyDependencyLocks verifyPublicPathSnapshot --continue # exit 0
|
|
|
|
|
./gradlew -p optional-platforms ci --continue # exit 0
|
|
|
|
|
./gradlew -p build-logic test # exit 0
|
|
|
|
|
./gradlew -p build-tools test # exit 0
|
|
|
|
|
|
|
|
|
|
Zero failing tasks. Run with `org.gradle.parallel=true`, which is itself part of what is being
|
|
|
|
|
verified: the same four commands failed eight tasks before the verification moved into the owning
|
|
|
|
|
projects, two more after the real-server lanes got their exclusive slot, and none once the web
|
|
|
|
|
testkit's socket budgets stopped asserting machine load (§3-B.1).
|
|
|
|
|
|
|
|
|
|
The configuration cache is verified separately, because it is a property of the task graph rather
|
|
|
|
|
than of a run:
|
|
|
|
|
|
|
|
|
|
./gradlew {ci,integrationCheck,releaseCheck} --dry-run \
|
|
|
|
|
--configuration-cache --configuration-cache-problems=fail # each: stored, then reused
|
|
|
|
|
|
|
|
|
|
Both invocations of each gate pass with problems set to `fail`, so this is not a count of warnings —
|
|
|
|
|
it is the absence of any. `verifyDependencyLocks` is outside that set by design (§3-B.7).
|
|
|
|
|
|
|
|
|
|
`ci` is leaf checks + `architectureCheck` + `qualityCheck` + `configContractCheck` +
|
|
|
|
|
`qualificationCheck` + `sampleWebContractCheck`, so it covers SpotBugs/FindSecBugs, the environment
|
|
|
|
|
contract, the auxiliary source sets that no leaf `check` compiles, and the app-bootstrap
|
|
|
|
|
`systemTest` and `architectureTest` lanes.
|
|
|
|
|
|
|
|
|
|
Container-backed lanes are not included: `integrationCheck` and the compose runtime matrix are
|
|
|
|
|
stage-2 CI jobs by design, and the review's Keep list says so. The one such lane this pass touched —
|
|
|
|
|
`messaging-outbox-jdbc-postgresql:integrationTest` — was verified to compile, which is what was
|
|
|
|
|
broken; its execution belongs to `integration-main.yml`.
|
|
|
|
|
|
|
|
|
|
One observation was first recorded here as a flake and turned out not to be one, which is the reason
|
|
|
|
|
it is worth keeping in the record. `adapter:inbound:web:webFluxContractTest` failed with a
|
|
|
|
|
`SocketTimeoutException` in one full run, passed in the next, and then failed again in a third — two
|
|
|
|
|
tests that time, in two different suites. Re-running would have "cleared" it twice. What it actually
|
|
|
|
|
was is described in §3-B.1: the web testkit's socket budgets were small enough to expire on a healthy
|
|
|
|
|
server under a parallel build, so every one of those suites carried an unstated assertion about
|
|
|
|
|
machine load. The three runs are the evidence that a lane starting real servers under contention is
|
|
|
|
|
exactly the kind of gate that gets turned off after being mistaken for a defect — and that calling
|
|
|
|
|
the first failure a flake was the beginning of doing that.
|
|
|
|
|
|
|
|
|
|
Falsifiability was checked where a new test could pass vacuously: the registration contract fails
|
|
|
|
|
when one non-auto-configuration line is added to a leaf's `.imports` file (and the file is
|
|
|
|
|
byte-identical to `HEAD` afterwards), and the gRPC operation-ledger aggregate was confirmed by
|
|
|
|
|
`--dry-run` to reach `:grpc:grpc-operation-ledger-jpa:integrationTest`.
|
|
|
|
|
|
|
|
|
|
Nothing is committed or pushed, and nothing is staged.
|
|
|
|
|
|
|
|
|
|
That last clause needed correcting rather than stating. The capability moves were made with
|
|
|
|
|
`git mv`, which stages, so 31 renames sat in the index — contrary to the repository's `human-only`
|
|
|
|
|
commit policy, and put there by this work rather than by anyone. The index is restored to `HEAD`.
|
|
|
|
|
Only the index was touched: the working tree's checksum is byte-identical before and after, and no
|
|
|
|
|
file was rewritten to achieve it.
|
|
|
|
|
|
|
|
|
|
Restoring it also makes the review's H5 warning sharper rather than softer, which is the reason to
|
|
|
|
|
say so precisely. The tree now stands at 252 modified files, **36 deleted** and **54 untracked**,
|
|
|
|
|
and the deletions and the additions are two halves of the same moves:
|
|
|
|
|
|
|
|
|
|
| Move | Old paths, now deleted | New paths, now untracked |
|
|
|
|
|
| --- | --- | --- |
|
|
|
|
|
| Redis capability → `cache-redis` | 8 main + 2 test | the same 10, plus a leaf `.imports` and a registration test |
|
|
|
|
|
| Fileserver capability → `fileserver` | 13 main + 1 test | 13 main + 2 new roots, plus a leaf `.imports` and 2 tests |
|
|
|
|
|
| Object storage qualification rename | 2 | 2 |
|
|
|
|
|
| `build-logic`/`build-tools` verification moves | 8 | 4, plus the tests that replaced them |
|
|
|
|
|
|
|
|
|
|
So `git commit -a` would record every deletion and none of the additions, and the result would
|
|
|
|
|
compile against nothing: `RedisCapabilityConfig` would be gone from `app-bootstrap` and absent from
|
|
|
|
|
`cache-redis` in the same commit. **`git add -A` is required**, and a `git status` that shows zero
|
|
|
|
|
untracked paths afterwards is the check that it worked.
|