fix!: 재리뷰가 찾은 결함과 색인에서 빠졌던 발견 5건을 고친다

문제: 성숙도 리뷰 수정을 다시 읽은 재리뷰(Critical 0, 회귀 0)가
Important 27건·Minor 41건을 찾았다. 수정이 새로 만든 결함이 많았다.
@SystemInvoked가 권한 규칙의 탈출구가 됐고, 데모 계정 거부가 정확히
prod일 때만 걸렸고, outbox 순서 번호가 업그레이드 때 뒤로 갔고,
GraphQL union 아래 복잡도가 0이었고, 서블릿 오류 응답이 success:true로
나갔고, 호출자 deadline이 breaker를 열었다. 원래 리뷰 색인이 06 영역
발견 5건(06-10~14)을 빠뜨려 그동안 처리되지 않았다.

방법: 영역별 7묶음이 test-driven-development 절차로 실패 테스트를
먼저 쓰고 고쳤다. 결정은 재리뷰 README 4절에 적었다: 순서 번호는 V21이
기록한 전역 floor부터, 데모 계정은 local·test·demo 허용 목록,
@SystemInvoked bean은 기동 실패, complete 단계 실패는
IDEMPOTENCY_COMPLETION_UNKNOWN(결과 불명), 캐시 조건부 쓰기는 별도
SCRIPT lane, strict lane 정책은 build cache 입력, 끝나지 않는 재시도
행은 원인 코드·지표·PARKED 절차. 레지스트리가 선언만 하고 아무 코드도
내보내지 않던 redis.capability.* 지표 6개는 빼고 runbook을 실제
지표(redis.lane.*) 기준으로 고쳤다.

위험·한계: V21을 V22 없이 직접 고쳤다(어느 배포에도 아직 없음). 이
브랜치의 V21을 적용한 개발 DB는 다시 만들어야 한다. 서블릿 오류
디스패치를 보안 체인이 허용한다. 발행·저장소 포트 시그니처가 바뀌어
파생 서비스 코드가 깨진다. prod에서 fixture-로 시작하는 설정값은
기동을 막는다. httpclient failure-injection lane의 main 단계 중복
실행은 이전 결정대로 남겼다. 실제 CI runner 실행은 확인하지 못했다.

검증: ci, mainCheck, :integrationCheck, -p optional-platforms ci
integrationCheck performanceCheck, :performanceCheck, :releaseCheck
(-PgitRevision), verify-compose-profile-contracts(18 lane) 통과.
첫 실행의 실패 2건(FindSecBugs 1, sample 동시 생성 테스트의 v4 event
id)은 고친 뒤 ci·integrationCheck·mainCheck를 다시 돌려 통과했다.

Refs: docs/reviews/2026-11-03-rereview/README.md 3·4·6·7절
BREAKING CHANGE: outbox 발행 포트와 저장소 포트 시그니처, 순서 번호
첫 값(V21 floor 다음), 샘플 데모 계정 프로필 허용 목록, 새 오류 코드
IDEMPOTENCY_COMPLETION_UNKNOWN, prod의 fixture- 값 거부, 기동 rule은
static·의존 없음이어야 한다.
This commit is contained in:
DongHyeonka
2026-11-03 23:29:35 +09:00
parent 7ef8d0b4cb
commit bbe944bd3c
406 changed files with 13493 additions and 1656 deletions
@@ -69,6 +69,36 @@ method can run longer than any fixed TTL; revisit when one exists. A sweeper job
rows — it needs a scheduler this family does not have, and deleting a claim makes the next retry a
fresh claim with no record that an attempt was abandoned.
**A taken-over owner is fenced by a claim token** (re-review 2026-11-03, RR-02-04). The owner a
takeover presumes dead may only be slow. Without knowing which attempt held the claim, its late
`markFailed` turned the new attempt's claim `FAILED_TERMINAL`, the next retry re-claimed it, and two
attempts ran the mutation at once; its late `markCommitted` committed beside the new attempt. Every
claim now carries a `GrpcClaimToken` (`claim_attempt`, migration V003): 1 for the first claim, one more
for each `reclaimFailed` or `reclaimStale`. `decide` returns it on `PROCEED`, and `recordCommit` /
`recordFailure` hand it back. A commit or a failure is one conditional
`UPDATE … WHERE state = 'IN_PROGRESS' AND claim_attempt = :attempt`: a commit by an attempt that lost
the claim is refused with an exception, so its business transaction rolls back; a failure by one is
ignored (`markFailed` returns false). The same conditional update closes the JPA commit race —
load-check-save let two commits that both read `IN_PROGRESS` both write; now exactly one matches.
A claim whose conflicting row is purged between insert and read retries the insert once instead of
failing the call.
Chosen over the claim moment as the token (`WHERE claimed_at = :claimedAt`), which needs no column:
PostgreSQL stores microseconds while `Instant` carries nanoseconds, so the token a caller holds and
the one stored can differ by rounding, and two processes' clocks can give two attempts the same
moment. The counter has neither problem. The heartbeat lease above stays rejected for the same reason
as before; the token fences without one.
**What `operation-ledger-enabled=true` promises (03-11 (b), recorded 2026-11-03, RR-02-07).** The
2026-11-01 review offered (a) delete the flag and judge by whether a ledger bean exists, or (b) keep the
flag and have the starter assemble the decision function from it. (b) is kept: the assembly and the
TTL validation live in one place instead of in every adopter. (a) was rejected because it removes the
false promise only by removing the assembly with it. The limit stays and is stated where an operator
meets it: the starter builds no server, so the flag assembles the decision function only, and keys are
honoured once the adopter's `IDEMPOTENCY` stage interceptor or service adapter calls `decide` and
`recordCommit`. `GrpcPlatformStartupValidator`'s refusal for a keyed method says exactly that; a running
process cannot see whether the stage exists, so the validator does not claim to check it.
**Finished rows are purged after a retention; in-progress rows never are** (03-10 follow-up).
`GrpcOperationLedger.purgeCompletedBefore(cutoff, limit)` deletes at most `limit` `COMMITTED` or
`FAILED_TERMINAL` rows that completed before the cutoff, oldest first — on PostgreSQL one bounded