chore(review): 1106 캐시·업로드 조치 브랜치를 합친다

This commit is contained in:
DongHyeonka
2026-11-06 07:02:34 +09:00
11 changed files with 156 additions and 8 deletions
@@ -1016,6 +1016,16 @@ rollback 목록 없이 덮어쓴다. pointer 읽기의 일시 오류는 activati
cleanup이 무엇을 지울지 볼 수 있다. 일시 오류에서는 inspect도 실패한다
(2026-11-03 02-05).
pointer entry는 있는데 저장된 본문이 다시 쓰기 전까지 계속 읽히지 않으면(blob
손실 등) 일시 오류와 구별되지 않는다. 조회, activation, cleanup, inspect가 모두
같은 실패(`UNAVAILABLE`/`RETRY` 등)를 계속 돌려준다. pointer를 쓰는 곳은
activation 하나라서 기다려도 돌아오지 않는다. 이때 운영자는
`activateRelease(releaseRegistryId, manifestDigestHex, { replaceUnreadablePointer: true })`로
검증된 후보를 활성화한다. 손상 pointer와 같이 rollback 목록 없이 덮어쓴다. 그 뒤
cleanup은 새 활성 release 밖의 owned release cache를 모두 지운다. 옵션이 없으면
지금처럼 멈춘다. 일시 오류 한 번으로 rollback 목록을 잃지 않기 위해서다
(1106/02-01).
현재 `cleanupOwned()`와 `inspect()`는 위 ownership/integrity 경계를 지키지만
cache count에 대한 cursor, per-invocation max count와 deadline은 제공하지 않는다.
따라서 cache 수에 비례한 unbounded maintenance를 production boot path에
+1 -1
View File
@@ -325,7 +325,7 @@ context는 URL 앞부분 하나(라우트 루트)를 가진다. 그 아래 기
- 멱등성 응답(`IDEMPOTENT_*`, `IDEMPOTENCY_*`)과 V3 표에 있는 web 공통 응답은 V3 실행기와 같은 표(`backend-envelope.ts`, 4절)로 판정한다. 400 `IDEMPOTENCY_KEY_INVALID`·422 `IDEMPOTENT_REQUEST_MISMATCH`는 `NOT_APPLIED`, 409 `IDEMPOTENT_IN_FLIGHT`·503 `IDEMPOTENCY_UNAVAILABLE`·503 `IDEMPOTENCY_COMPLETION_UNKNOWN`·500 `IDEMPOTENCY_RECOVERY_REQUIRED`는 `MAYBE_APPLIED`다.
- `APPLIED_CONFIRMED`: 서버는 만들었다고 답했지만 응답을 읽지 못했다. 다시 올리지 않는다.
- `code`는 백엔드 code 그대로다(`FILE_TOO_LARGE`, `UNSUPPORTED_MEDIA_TYPE`, `MALWARE_DETECTED` …). 백엔드가 code를 주지 않는 거절은 client가 이름을 붙인다(`FILE_REJECTED`, `FILE_EMPTY`, `FILE_NOT_READABLE`). 화면 문구는 기능이 code로 고른다.
- 실패 Envelope의 `meta.traceId`는 실패의 `traceId`로 온다. V3 실행기와 같다. 지원 담당이 서버 로그에서 그 요청을 찾을 때 쓴다.
- 실패 Envelope의 `meta.traceId`는 실패의 `traceId`로 온다. V3 실행기와 같다. 지원 담당이 서버 로그에서 그 요청을 찾을 때 쓴다. 202 `REJECTED`의 `FILE_REJECTED`도 성공 Envelope의 `meta.traceId`를 싣는다. 이 응답에는 거절 이유가 없다. 짝 백엔드는 검증기 code를 파일 레코드에만 적는다. 그래서 화면은 이유별 문구를 고를 수 없다.
- client는 스스로 다시 보내지 않는다. `retryable: true`이고 `retryAfterMs`가 있으면 그만큼 기다린 뒤 기능이 다시 부른다. 서버가 상한(`maxRetryAfterMs`, 기본 30초)보다 길게 기다리라고 하면 `retryable: false`다.
- `MAYBE_APPLIED`는 언제나 `retryable: false`다. 백엔드가 Envelope에 `retryable: true`와 `Retry-After`를 붙여도(504 `TRANSFER_TIMEOUT`) 기능이 자동으로 다시 부르지 않는다. 아래 주의대로 먼저 확인한다.
@@ -27,6 +27,12 @@ session/account Query lifecycle, strict query policy와 Web Storage v2 lifecycle
않는다. 두 operation은 Cache Storage와 mutation lock만 요구하고 실패 시
recovery는 `RETRY`다. `stageRelease`만 fetcher를 요구하며 recovery는
`ONLINE_ONLY`다.
- active pointer 본문을 계속 읽지 못하면 조회, `activateRelease`,
`cleanupOwned`, `inspect`가 모두 같은 실패를 계속 돌려준다. 다시 시도해도
돌아오지 않는다. pointer를 쓰는 곳은 activation 하나다. 이때
`activateRelease(id, digest, { replaceUnreadablePointer: true })`로 검증된
release를 활성화한다. rollback 목록은 잃고, 다음 cleanup이 다른 owned release
cache를 지운다. 옵션 없는 activation은 일시 오류에서 멈춘다(1106/02-01).
- variant를 사용하는 policy(`allowedVaryHeaderNames` 비어 있지 않음)는 반드시
`allowedResponseHeaderNames`에 `vary`를 포함해야 한다. 아니면 composition이
`TypeError`로 즉시 실패한다. 저장된 variant가 같은 key로 충돌하는 상태를 만들지
@@ -719,9 +719,12 @@ export function createPublicResponseCacheAdapter(
// under this policy) is replaced by this verified candidate with
// an empty rollback list, rather than blocking every later
// deployment. A transient read failure still stops activation.
// 1106/02-01. Unless the caller asks to replace it: a body that
// stays unreadable fails the same way, and only this writes it.
if (
!previousPointer.ok &&
previousPointer.error.code !== "CORRUPT_DATA"
previousPointer.error.code !== "CORRUPT_DATA" &&
options.replaceUnreadablePointer !== true
) {
return rebaseFailure(
previousPointer.error,
+13 -2
View File
@@ -93,13 +93,23 @@ const ENVELOPE_INVALID: EnvelopeRead<never> = Object.freeze({ ok: false });
/**
* The success `data`. A missing key is the null the backend dropped.
* 1103/03-02. A page list's `meta.page` comes with it.
* 1106/01-01. And the backend's `meta.traceId`, when it sent one, as
* `readFailureEnvelope` gives it, for a client that turns a well-formed
* answer into a failure (Fileserver's 202 `REJECTED`).
*/
export function readSuccessEnvelope(
value: unknown,
): Readonly<{ ok: true; value: unknown; page?: EnvelopePage }> | Readonly<{ ok: false }> {
):
| Readonly<{
ok: true;
value: unknown;
page?: EnvelopePage;
traceId?: string;
}>
| Readonly<{ ok: false }> {
const parsed = successEnvelopeSchema.safeParse(value);
if (!parsed.success) return ENVELOPE_INVALID;
const page = parsed.data.meta.page;
const { page, traceId } = parsed.data.meta;
return Object.freeze({
ok: true as const,
value: parsed.data.data,
@@ -113,6 +123,7 @@ export function readSuccessEnvelope(
...(typeof page.sort === "string" ? { sort: page.sort } : {}),
}),
}),
...(traceId === undefined ? {} : { traceId }),
});
}
+7
View File
@@ -682,9 +682,15 @@ async function readAnswer(
// 1105/01-01. The file is the user's to replace, as with a 413, 415 or
// 422 refusal. The backend keeps the record unpublished and schedules
// its bytes for removal, so nothing usable was stored.
// 1106/01-01. The server answered, so the failure carries its
// `meta.traceId` as an Envelope failure does (1105/01-03). The answer
// names no reason: the backend keeps the verifier's code on its record.
return failure("VALIDATION_REJECTED", "FILE_REJECTED", {
httpStatus: status,
effect: "NOT_APPLIED",
...(envelope.traceId === undefined
? {}
: { traceId: envelope.traceId }),
});
}
return Object.freeze({
@@ -790,6 +796,7 @@ function failure(
details: Readonly<{
effect: FailureEffectCertainty;
httpStatus?: number;
traceId?: string;
}>,
): Result<never, AppFailure> {
return Object.freeze({
@@ -45,7 +45,9 @@ export type PublicCacheInspection = Readonly<{
* no longer accepts: lookup and cleanup stop on it and the next verified
* activation replaces it with an empty rollback list. The candidates are
* still listed so the operator can see what that activation and the
* cleanup after it will keep. A transient pointer read is a failure.
* cleanup after it will keep. A transient pointer read is a failure. One
* that keeps failing is the case for `replaceUnreadablePointer` on
* `activateRelease` (1106/02-01).
*/
activePointer: "NONE" | "READABLE" | "UNREADABLE";
/** `null` unless `activePointer` is `READABLE`. */
@@ -94,10 +96,23 @@ export interface PublicResponseCacheAdminPort {
manifest: PublicCacheReleaseManifest,
options?: Readonly<{ signal?: AbortSignal }>,
): Promise<BrowserDataResult<PublicCacheReleaseSummary>>;
/**
* 1106/02-01. A pointer read that fails other than as damage stops the
* activation, so one bad read does not lose the rollback list. A pointer
* body that stays unreadable fails the same way every time, and lookup,
* cleanup and inspect stop on it too; only an activation writes the
* pointer. `replaceUnreadablePointer: true` is the caller's way back: this
* activation then replaces a pointer it cannot read, as it always replaces
* a damaged one, with an empty rollback list, and the next cleanup deletes
* every other release cache.
*/
activateRelease(
releaseRegistryId: string,
manifestDigestHex: string,
options?: Readonly<{ signal?: AbortSignal }>,
options?: Readonly<{
signal?: AbortSignal;
replaceUnreadablePointer?: boolean;
}>,
): Promise<BrowserDataResult<PublicCacheReleaseSummary>>;
cleanupOwned(
request?: Readonly<{
@@ -214,6 +214,8 @@ describe("Fileserver multipart upload — success", () => {
it("reads 202 REJECTED as a refused file, not as a broken answer", async () => {
// 1105/01-01. A verifier refused the file for a reason other than malware.
// 1106/01-01. The server answered, so its `meta.traceId` comes with the
// failure for support to find the upload in the server logs.
answer({
status: 202,
body: successEnvelope(
@@ -233,6 +235,7 @@ describe("Fileserver multipart upload — success", () => {
retryable: false,
action: "none",
operationId: "FILESERVER_UPLOAD",
traceId: "trace-1",
}),
});
});
+2 -1
View File
@@ -19,13 +19,14 @@ describe("paired backend Envelope boundary", () => {
});
it("reads data, and a dropped data key as null", () => {
// 1106/01-01. With `meta.traceId`, as the failure reader gives it.
expect(
readSuccessEnvelope({
success: true,
data: [],
meta: { requestId: "request-1", traceId: "trace-1" },
}),
).toEqual({ ok: true, value: [] });
).toEqual({ ok: true, value: [], traceId: "trace-1" });
expect(readSuccessEnvelope({ success: true, meta: {} })).toEqual({
ok: true,
value: null,
+6 -1
View File
@@ -50,7 +50,12 @@ describe("accessibility e2e API answers", () => {
expect(answer.kind).toBe("empty-list");
if (answer.kind !== "empty-list") return;
// What the executor reads out of the frame is what the operation accepts.
expect(readSuccessEnvelope(answer.body)).toEqual({ ok: true, value: [] });
// 1106/01-01. The reader also gives the scenario meta's `traceId`.
expect(readSuccessEnvelope(answer.body)).toEqual({
ok: true,
value: [],
traceId: "0af7651916cd43dd8448eb211c80319c",
});
expect(LIST!.contract.outputValidator.safeParse([]).success).toBe(true);
});
@@ -970,4 +970,91 @@ describe("public cache pointer after a policy change or damage (0929/02-01)", ()
await adapter.responses.matchActiveExact({ absoluteUrl: assetUrl("r1") }),
).toMatchObject({ ok: true, value: expect.anything() });
});
// 1106/02-01 (review experiment 02-R1 X3). The pointer entry is there, but
// its stored body fails to read until something writes it again (a lost
// blob). Without the option that looks like the transient failure above and
// everything stops; only an activation writes the pointer.
it("replaces a pointer whose body stays unreadable when the activation asks to (1106/02-01)", async () => {
const cacheStorage = new MemoryCacheStorage();
const policy = policyRetaining(1);
const adapter = adapterFor(cacheStorage, policy);
expect(await deploy(adapter, "r1", policy)).toMatchObject({ ok: true });
const control = cacheStorage.caches.get(controlCacheName(policy))!;
const pointerUrl = activePointerRequest(policy).url;
const pointer = { unreadable: true, rewrites: 0 };
const realMatch = control.match.bind(control);
const realPut = control.put.bind(control);
const urlOf = (request: RequestInfo | URL) =>
request instanceof Request ? request.url : new URL(String(request)).href;
control.put = async (request, response) => {
if (urlOf(request) === pointerUrl) {
pointer.unreadable = false;
pointer.rewrites += 1;
}
await realPut(request, response);
};
control.match = async (request) => {
const stored = await realMatch(request);
if (!stored || urlOf(request) !== pointerUrl || !pointer.unreadable) {
return stored;
}
return new Response(
new ReadableStream<Uint8Array>({
pull(controller) {
controller.error(new TypeError("blob missing"));
},
}),
{ headers: stored.headers },
);
};
const next = await manifestOf("r2", policy);
const activateNext = (
options?: Readonly<{ replaceUnreadablePointer?: boolean }>,
) =>
adapter.admin.activateRelease(
next.releaseRegistryId,
next.manifestDigestHex,
options,
);
const unavailable = (operation: string) => ({
ok: false,
error: { code: "UNAVAILABLE", operation, recovery: "RETRY" },
});
// Staging does not read the pointer; everything else stops on it.
expect(await adapter.admin.stageRelease(next)).toMatchObject({ ok: true });
expect(
await adapter.responses.matchActiveExact({ absoluteUrl: assetUrl("r1") }),
).toMatchObject(unavailable("CACHE_LOOKUP"));
expect(await activateNext()).toMatchObject(unavailable("CACHE_ACTIVATE"));
expect(await adapter.admin.cleanupOwned()).toMatchObject(
unavailable("CACHE_DELETE"),
);
expect(await adapter.admin.inspect()).toMatchObject(
unavailable("CACHE_LOOKUP"),
);
expect(await activateNext()).toMatchObject(unavailable("CACHE_ACTIVATE"));
expect(pointer).toEqual({ unreadable: true, rewrites: 0 });
// The way back: this activation replaces the pointer it cannot read.
expect(
await activateNext({ replaceUnreadablePointer: true }),
).toMatchObject({ ok: true, value: { releaseRegistryId: "r2" } });
expect(pointer).toEqual({ unreadable: false, rewrites: 1 });
expect(
await adapter.responses.matchActiveExact({ absoluteUrl: assetUrl("r2") }),
).toMatchObject({ ok: true, value: expect.anything() });
expect(await adapter.admin.inspect()).toMatchObject({
ok: true,
value: { activePointer: "READABLE", activeReleaseRegistryId: "r2" },
});
// The rollback list went with the old pointer, so cleanup drops r1.
expect(await adapter.admin.cleanupOwned()).toMatchObject({
ok: true,
value: { deletedOwnedCaches: 1, retainedOwnedCaches: 2 },
});
// A later activation without the option works as before.
expect(await deploy(adapter, "r3", policy)).toMatchObject({ ok: true });
});
});