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.
12 KiB
Fileserver configuration
Every key below lives under app.fileserver-platform (environment form
APP_FILESERVER_PLATFORM_*).
While app.fileserver-platform.enabled is false none of these keys is bound at all — the
auto-configuration that binds them is not processed — so a malformed value in a block nobody
enabled cannot fail a startup. Once enabled, binding is strict: an unknown key under the prefix is
refused rather than ignored. The defaults are the conservative ones: the
admin plane is off, background reclamation is on (leaving it off is what fills the volume), copies
and moves may only target the default namespace, and there is no permissive authorization fallback. Turning the capability on is a deliberate act, and so is every surface it
exposes.
Minimum to start
app:
fileserver-platform:
enabled: true
instance-id: ${HOSTNAME} # writer-lease owner; must be unique per node
storage:
root: /var/lib/backend/files # absolute, outside any webroot or config dir
security:
access-policy: role-based # or supply your own FileAccessPolicyPort bean
observability:
fingerprint-key: ${FILESERVER_FINGERPRINT_KEY}
Startup fails, rather than degrading, when any of these is missing or unsafe:
| Condition | Why it is fatal |
|---|---|
security.access-policy left at required with no FileAccessPolicyPort bean |
a file capability that authorizes by default is worse than one that refuses to start |
observability.fingerprint-key unset while metrics are on |
an unkeyed digest of an enumerable identifier is reversible |
| the storage root fails a mandatory capability probe | a volume that cannot create atomically, keep staging and content on one FileStore, or refuse symlinks is unsafe, not degraded |
storage.publish-mode: atomic-move-required on a volume where the probe could not prove an atomic move |
the configured guarantee cannot be delivered |
security.access-policy: unenforced under a prod profile |
a value that was convenient in development must not survive promotion |
Authorization — security
| Key | Default | Meaning |
|---|---|---|
access-policy |
required |
required (supply your own bean), role-based, or unenforced |
read-roles |
ROLE_FILE_READ |
grants metadata read and download |
write-roles |
ROLE_FILE_WRITE |
grants create, append, finalize, delete, copy, move |
admin-roles |
ROLE_FILE_ADMIN |
grants reverify and force-delete, and gates /internal/fileserver/** at the servlet chain |
These roles decide the tier inside each operation. In front of it, every operation's use case
carries a file:* permission that the web adapter's method security checks against
app.authz.role-permissions; security.yml grants file:read to ROLE_FILE_READ,
file:upload, file:delete, file:copy and file:move to ROLE_FILE_WRITE, and file:admin to
ROLE_FILE_ADMIN. Renaming a tier role here means mapping the new name there as well, or the
caller is refused before the tier is consulted (docs/fileserver/design-deviations.md §5).
There is no anonymous-read switch. Every Fileserver route is authenticated by the servlet chain before any application policy is consulted, so such a setting could only ever have described a permission the transport had already refused — a configuration that reads as if it grants access and does not.
The three tiers do not inherit. An admin role cannot delete through the data plane, and a write role cannot reach the management plane — a role model where "can delete" implied "can force-delete" would make the audited plane reachable through the unaudited one.
unenforced authorizes everything and exists so a developer can exercise upload and download before
deciding on a role model. It is refused under a production profile.
Storage — storage
| Key | Default | Meaning |
|---|---|---|
root |
/var/lib/backend/files |
absolute path; the only place a path exists |
publish-mode |
atomic-move-preferred |
atomic-move-required, atomic-move-preferred, metadata-pointer |
buffer-size |
128KB |
bounds every transfer allocation; memory never scales with file size |
forbidden-root-ancestors |
/app,/etc,/usr/share/nginx/html |
roots the storage root must not live under (webroot, config dirs) |
root must be absolute. A relative root resolves against the process working directory, which is
one path in a container and another in a test, so it is refused at binding time.
Three former keys are gone, pinned as constants instead: staging and content share one FileStore, symbolic links are never followed, and the object and its directory are synced before READY. Each is an invariant the atomic publish and the namespace boundary are built on — a deployment that could switch one off would be running a different capability under the same name and the same tests.
The storage provider has no selector either. There is exactly one implementation, and a type key
with one legal value is a promise of pluggability that nothing keeps.
Upload, download, transfer
| Key | Default | Meaning |
|---|---|---|
upload.max-file-size |
100MB |
hard ceiling; also drives spring.servlet.multipart.max-file-size |
upload.max-request-size |
110MB |
request envelope; must be at least max-file-size |
upload.initial-reservation |
8MB |
quota reserved when the length is unknown |
upload.ttl |
1h |
how long a resumable upload stays claimable |
upload.reservation-ttl |
24h |
how long an unsettled quota reservation survives |
upload.lease-duration |
30s |
writer lease; renewed at one third of this |
upload.max-parts |
16 |
multipart part ceiling |
upload.require-content-length |
false |
refuse chunked raw uploads |
download.cache-control |
private, no-store |
emitted on every content response |
download.inline-allowed |
false |
scriptable content is always an attachment regardless |
download.max-ranges |
1 |
multi-range responses are opt-in |
download.max-range-bytes |
100MB |
total bytes one ranged response may cover |
download.zero-copy-enabled |
true |
hand large plaintext responses to the kernel |
download.zero-copy-minimum-bytes |
16MB |
below this the syscall setup costs more than it saves |
transfer.core-size / max-size / queue-capacity |
8 / 32 / 64 |
blocking transfer pool bounds |
transfer.await-seconds |
300 |
how long a transfer may occupy a pool thread |
Zero copy changes no header and no status. When storage declines it — an unreadable region, an unsupported backend — the response is streamed instead and is byte-identical.
Verification — verification
| Key | Default | Meaning |
|---|---|---|
timeout |
5s |
per-verifier ceiling |
require-media-type-verdict |
false |
refuse a file whose type could not be determined |
inline-safe-profile |
false |
accept scriptable content instead of quarantining it |
Set inline-safe-profile: true only when downloads are never served inline from a trusted origin.
Quota and admission — quota
| Key | Default | Meaning |
|---|---|---|
instance-upload-permits |
16 |
concurrent uploads this node admits |
scope-upload-permits |
4 |
concurrent uploads one caller (principal) admits; over it is a retryable 429 |
direct-download-permits |
64 |
concurrent non-delegated downloads |
soft-high-water |
0.70 |
storage fraction at which pressure is reported |
hard-high-water |
0.85 |
storage fraction at which uploads and copies are refused |
namespace-limit |
unset (unlimited) | most bytes one namespace may hold, live reservations plus committed usage; a create, copy or move that would pass it is refused with QUOTA_EXCEEDED (413) before any record, byte or namespace changes |
namespace-limit is checked against the usage the quota ledger reports and then reserved, in one
transaction but without a per-namespace lock: reservations admitted at the same moment can overshoot
the limit by at most their own size, and the next one is refused. A service that needs an exact cap
adds that lock in its persistence adapter.
A move to another namespace carries the file's bytes in the ledger: in the transaction that relocates the record it checks the target's limit, books the bytes as committed usage there and reclaims them from the source. A refused move leaves the file, and both namespaces' usage, as they were.
When the storage fraction cannot be read, admission treats it as unknown and does not apply the
high-water rule — a synthetic 0 would silently disable the guard, and a synthetic 1 would take
the capability down over a failed syscall.
Background reclamation — cleanup
| Key | Default | Meaning |
|---|---|---|
enabled |
true |
run the cleanup worker on this node; false logs a startup warning |
interval |
60s |
fixed delay between batches, not fixed rate |
max-items |
100 |
items one batch may claim |
max-bytes |
1GB |
bytes one batch may reclaim |
retry-backoff |
5m |
how long a failed item waits before it is due again |
The worker is on by default: it deletes only what the cleanup queue names, and without it cancelled
and expired staging, rejected and deleted content stay on the volume until the hard high-water mark
stops every upload. A node may turn it off — for instance when another node runs it — and startup
then logs a WARN naming app.fileserver-platform.cleanup.enabled=false. A node without it still
queues cleanup items; another node or an operator reclaims them. (A required "reason" key was
considered instead of the warning and rejected: it would be a setting whose only effect is to
silence startup, carried by every non-worker node.) An item that fails
eight times is abandoned rather than retried forever — it stays visible to an operator, parked
rather than discarded.
Each pass also expires uploads whose expiresAt has passed (the record becomes EXPIRED and its
staged bytes are queued), purges orphans the admin plane retired to quarantine, and runs
reconciliation over stale UPLOADED/VERIFYING/FAILED records, so an ambiguous publish settles
without an operator.
Copy and move targets
app.fileserver-platform.target-namespaces (APP_FILESERVER_PLATFORM_TARGET_NAMESPACES, a comma
list, empty by default) names the namespaces copy and move may write into besides
default-namespace, which is always allowed. The target namespace of a copy or a move comes from the request body, and a namespace is also the
quota grouping. Any other target is refused with ACCESS_DENIED before a byte moves or a
reservation is taken.
Management plane — admin
| Key | Default | Meaning |
|---|---|---|
enabled |
false |
expose /internal/fileserver/** |
orphan-minimum-age |
1h |
how long an unreferenced object must exist before a scan may name it |
Publishing content and committing its record are two steps. Anything younger than
orphan-minimum-age is assumed to be mid-commit rather than abandoned; shortening this makes
concurrent uploads look like orphans.
Front-proxy delegation — nginx
| Key | Default | Meaning |
|---|---|---|
enabled |
false |
emit X-Accel-Redirect instead of a body |
internal-prefix |
/__files/ |
must be an internal location resolving to the content root |
object-suffix |
.bin |
layout suffix the proxy appends |
minimum-size |
16MB |
below this the application serves the transfer itself |
Delegation is decided only after authorization and the READY gate, so an internal redirect can only ever name content the caller was already allowed to read.
Protocols — tus
| Key | Default | Meaning |
|---|---|---|
enabled |
false |
expose the tus 1.0 endpoints |
The HTTPbis resumable-upload draft-12 surface is experimental and documented in support-matrix.md.