문제: 3차 수정의 확인 리뷰가 새 Important 없이 Minor 19건을 남겼다. 동작에 닿는 것이 셋이다. JDK read-idle watchdog이 공용 ForkJoinPool에서 돌아 그 풀이 바쁘면 300ms timeout이 6초 가까이 늦었다. 붙여 쓴 -x<task>를 verifyStrictLaneStages가 읽지 않아 lane을 뺀 workflow가 통과했다. 사용자 전송이 중간 클래스를 거쳐 @SystemCaller bean을 부르면 ArchUnit 규칙을 지났다. 잠금 대기 검사기는 FK·trigger·REINDEX와 lock_timeout = 0을 못 봤고, 쿼터 보정 SQL은 코드와 다른 모양의 행을 넣었다. 방법: 영역별 4묶음이 맡은 Minor만 test-driven-development 절차로 고쳤다(재리뷰 README 12절). - watchdog은 전용 데몬 스레드 하나에서 돌고 스트림을 닫으면 작업을 취소한다. read당 비용은 그대로다(21–24ns). - 사용자 전송 → @SystemCaller는 직접·간접 의존 모두 거부한다. 리뷰가 권한 직접 의존 검사는 중간 클래스 하나로 다시 열려서 버렸다. - 검사기를 보강하면서 새로 걸린 이미 배포된 마이그레이션 5개는 내용을 바꾸지 않고 동결 목록에 더했다(main과 blob이 같다). - 비밀 목록 bean은 규칙보다 먼저 만들어 실제 클래스로 검사한다. 위험·한계: 회복 필요 예외 메시지 끝에 scope digest가 붙고 기동 실패 문구가 바뀌어, 문구 전체에 맞추던 로그 매처는 고쳐야 한다. 코드가 든 Redis 계정 참조의 fixture- 검사는 문서로만 막았다(코드로 막으려면 새 의존 edge가 필요하다). failure-retention ≥ replay-ttl 기동 검사는 넣지 않았다. -mx 같은 짧은 플래그 묶음과 graphql 전송 fixture는 남았다. spotbugsTestFixtures(게이트 밖)의 정규식 경고가 15건으로 늘었다. 검증: ci, 두 빌드의 verifyStrictLaneStages --dry-run, mainCheck, :integrationCheck, -p optional-platforms ci integrationCheck performanceCheck, :performanceCheck, :releaseCheck(-PgitRevision), verify-compose-profile-contracts(18 lane) 첫 실행 통과. watchdog은 공용 풀을 채운 상태에서 300ms 설정이 300ms에 끝나는 것을 쟀다. Refs: docs/reviews/2026-11-03-rereview/README.md 11·12절, round3/
17 KiB
HTTP Client Platform — Configuration Reference
Every outbound call resolves exactly one Named Client Profile. The whole capability lives under
the app.httpclient prefix: profiles under app.httpclient.clients[N], Dynamic Target policies
under app.httpclient.dynamic-targets[N].
Design §30.1 forbids a production profile from inheriting large framework defaults. Anything a
production deployment must decide has either no default or an unusable one, and
HttpClientStartupValidator fails the context rather than guessing.
The master switch
| Property | Type | Default | Environment |
|---|---|---|---|
app.httpclient.enabled |
boolean | true |
APP_HTTPCLIENT_ENABLED |
The module is on when present: a service adds runtimeOnly project(':adapter:outbound:httpclient')
to app-bootstrap/build.gradle, and the module's own config/module-defaults/httpclient.yml (loaded at the lowest
precedence by the support leaf's ModuleDefaultsEnvironmentPostProcessor) supplies enabled: true. Absent from the
build is the ordinary off state. false is the explicit off: HttpClientPlatformAutoConfiguration
is then never processed and neither is anything it imports — no property is bound, and no transport
provider, connection pool, TLS context, credential, thread, gateway or actuator endpoint exists. A
malformed HTTP client setting cannot fail the startup of a deployment that switched it off.
Anything that is not exactly true — yes, 1, blank — turns the platform off. On with no client
declared is a startup failure carrying HTTPCLIENT_ACTIVE_WITHOUT_CLIENTS: a platform with nothing
to call still holds transport providers and gateways no caller can reach.
Declaring clients from the environment
Clients are an indexed list carrying their own name, not a map keyed by name. A map key becomes a
segment of the environment variable and the relaxed binder normalises it, so payment-api and
payment_api would arrive as one entry with nothing said about the one that was lost. Both a
duplicate name and a name that collides once normalised fail startup.
APP_HTTPCLIENT_ENABLED=true
APP_HTTPCLIENT_CLIENTS_0_NAME=payment
APP_HTTPCLIENT_CLIENTS_0_BASE_URL=https://payment.example
APP_HTTPCLIENT_CLIENTS_0_ALLOWED_HOSTS_0=payment.example
APP_HTTPCLIENT_CLIENTS_0_ALLOWED_PORTS_0=443
APP_HTTPCLIENT_CLIENTS_0_REQUEST_MAX_BODY_BYTES=1048576
APP_HTTPCLIENT_CLIENTS_0_TLS_PROFILE_ID=payment
APP_HTTPCLIENT_DYNAMIC_TARGETS_0_NAME=webhook
APP_HTTPCLIENT_DYNAMIC_TARGETS_0_ALLOWED_SCHEMES_0=https
docs/httpclient/env-fields.yaml is the registry of accepted variable names. It is
derived from the settings record and held to it in both directions, and the platform refuses to
start on an APP_HTTPCLIENT_ variable that is not in it — so
APP_HTTPCLIENT_CLIENTS_0_TIMEUOT_TOTAL_CALL fails startup instead of silently leaving the client
on its default budget. Unknown keys supplied through a configuration file rather than the
environment are refused by strict binding for the same reason.
Only APP_HTTPCLIENT_ENABLED appears in src/.env and docs/registries/env-keys.yaml. It is the
one key with a deployment-independent value; templating an indexed client in application.yml would
materialise a nameless client in every deployment, which the aggregate validation refuses.
app.httpclient.clients[N]
| Property | Type | Default | Notes |
|---|---|---|---|
name |
string | — | Required, unique, and distinct from every other name once normalised for the environment |
mode |
TRUSTED | DYNAMIC |
TRUSTED |
A dynamic profile may not carry a default credential |
base-url |
URI | — | Required for a trusted profile; no userinfo, no query |
allowed-hosts |
list | empty | Required in production |
allowed-ports |
list | empty | Compared against the effective port |
transport |
APACHE | JDK | JETTY | SIMPLE |
APACHE |
SIMPLE is rejected in production; JETTY needs a deployment-registered provider |
protocols |
list | HTTP_1_1 |
The default transport is Apache, whose classic client is HTTP/1.1 only; a profile that wants HTTP/2 declares it together with a transport that can deliver it. HTTP_3 requires the experimental acknowledgement |
experimental-acknowledgement |
string | — | Must equal I_ACCEPT_HTTP3_EXPERIMENTAL_SEMANTICS |
pool
| Property | Default | Meaning |
|---|---|---|
max-total-connections |
50 |
Socket ceiling for the runtime |
max-connections-per-route |
25 |
Per-upstream ceiling |
max-pending-acquires |
100 |
Waiting-request memory ceiling |
pending-acquire-timeout |
200ms |
Pool or stream wait ceiling |
max-idle-time |
30s |
Idle eviction |
max-life-time |
5m |
Picks up DNS, load-balancer, and certificate changes |
validate-after-inactivity |
5s |
Stale and half-open detection |
requires-route-pool |
false |
Set when route-scoped limits are mandatory; the JDK transport then refuses the profile |
requires-bounded-pending-queue |
false |
Same, for a bounded pending queue |
timeout
| Property | Default | Meaning |
|---|---|---|
dns |
300ms |
Hostname resolution |
connect |
500ms |
Socket connect |
tls-handshake |
1s |
TLS and ALPN |
response-header |
2s |
Until final response headers |
read-idle |
3s |
Between response chunks |
total-call |
4s |
The whole logical call, including retry backoff |
total-call must not be shorter than connect or response-header; the validator emits
INVALID_TIMEOUT_BUDGET otherwise. Every timeout above, and pool.pending-acquire-timeout, must be
greater than zero (TIMEOUT_NOT_POSITIVE): both engines read zero as "wait forever".
total-call also bounds the setup of a streaming download — connect, send, and the wait for
response headers. Once the stream is handed over, each gap of the body is bounded by read-idle
instead, so a long download is not cut at an arbitrary total (design §15.3). This holds on both
transports. The JDK transport keeps its exchange-wide timeout off a streamed body, and a per-stream
watchdog bounds each read's wait by read-idle, because the JDK client has no socket read timeout.
The watchdog wakes on one daemon thread of its own (httpclient-jdk-read-idle), not on the common
ForkJoinPool, so a service that fills that pool with blocking work does not delay it (re-review
2026-11-04 RR3-03-01). Before review 2026-11-03 RR-04-03 every JDK download longer than
response-header was cut mid-body. A body failure after hand-over is thrown from read as HttpBodyReadException — an IOException, as
InputStream promises, so catch (IOException) still sees it — whose failure() is
HttpResponseTimeoutException (a wait longer than read-idle) or HttpResponseTruncatedException
(the connection ended early), stage RESPONSE_BODY, evidence PARTIAL_RESPONSE (re-review
2026-11-04 RR-03-14). On the JDK transport read-idle bounds each read's wait, skip included, and
time the caller spends between reads does not count (RR-03-12).
Removed settings
Six keys were removed because they were bound and reached no transport (review 2026-11-01 05-15).
Setting one now fails startup with a bind error naming the key (The elements [app.httpclient.clients[0].timeout.request-write-idle] were left unbound.), instead of being
accepted and ignored.
| Removed key | Why |
|---|---|
timeout.request-write-idle |
Neither engine can bound a blocking socket write |
timeout.proxy-connect, proxy.connect-timeout |
The TCP connect to a proxy is governed by timeout.connect |
pool.eviction-interval |
Apache's idle evictor runs at max-idle-time |
pool.shutdown-timeout |
A retired runtime drains for app.httpclient.rotation-drain-timeout |
observability.operation-name-required |
An operation name is mandatory in the API; the switch could only be true |
circuit-breaker
What counts as a failure is fixed: a 5xx, 408 or 429 response — whether the call throws or returns it as a result — and every transport failure. Every other 4xx is a healthy answer. These settings only say how many failures, out of how many calls, open the circuit. Defaults are Resilience4j's own, which every profile ran with before the keys existed.
Two deadlines can end an attempt in flight, and they are counted differently (review 2026-11-03 RR-04-01):
| Ended by | Breaker | Why |
|---|---|---|
The caller's own deadline (HttpOperation.deadline(), when earlier than total-call) |
Not recorded; the permission is given back | The caller ran out of time. Counting it opened the circuit on a healthy upstream whenever callers passed short budgets |
The profile's timeout.total-call |
Recorded as a failure | total-call is the operator's statement of how long a healthy upstream may take; a slow-drip upstream that never trips a per-gap timeout is seen only through it |
The first row has a blind spot (re-review 2026-11-04 RR-03-16). The attempt is let off whenever the
caller's deadline was the earlier one and had passed when the attempt failed — the code does not ask
whether the upstream was still answering within the profile's limits. If every caller of a profile
passes a budget shorter than connect or response-header, an upstream that hangs (drops packets,
never answers) ends every attempt at the caller's deadline first, so no attempt is recorded and the
circuit never opens; every call waits out its own budget. A failure that comes before the caller's
deadline — a refused connection — is still counted. For such a profile, set connect and
response-header shorter than the shortest caller budget.
| Property | Default | Meaning |
|---|---|---|
failure-rate-threshold |
50 |
Percent of failed calls in the window that opens the circuit, 1–100 |
minimum-number-of-calls |
100 |
Calls recorded before the rate is evaluated; at most sliding-window-size |
sliding-window-size |
100 |
The last N calls the rate is computed over (count-based) |
wait-duration-in-open-state |
60s |
How long an open circuit rejects before trial calls; positive |
permitted-calls-in-half-open-state |
10 |
Trial calls while half-open; at least 1 |
An out-of-range value fails startup with CIRCUIT_BREAKER_SETTING_INVALID. The breaker belongs to
the profile, not to a runtime generation, so a certificate rotation keeps its state.
redirect, request, response
| Property | Default | Meaning |
|---|---|---|
redirect.enabled |
false |
Engine redirect handling is always off; the platform follows hops itself |
redirect.max-hops |
0 |
Enabling redirects with zero hops is a configuration error |
redirect.allow-cross-origin |
false |
When enabled, credentials are stripped on the hop |
request.max-body-bytes |
0 |
Required in production |
request.compression |
false |
|
response.max-wire-bytes |
5242880 |
Bytes on the wire |
response.max-decoded-bytes |
10485760 |
Bytes after decoding; hard maximum is 64 MiB |
response.allowed-content-types |
JSON + problem+json | Empty means "any" |
authentication
| Property | Default | Meaning |
|---|---|---|
type |
NONE |
One of the design §20.1 methods |
registration-id |
— | Required for OAuth2 |
scopes |
empty | Part of the token cache key |
audience |
— | Part of the token cache key |
header-name |
— | Required for API_KEY_HEADER; must be on the allowlist |
secret-reference |
— | Resolved by the deployment's secret loader, never a literal |
retry
| Property | Default | Meaning |
|---|---|---|
policy |
none |
Named policy for reporting |
max-attempts |
1 |
Attempts, not retries; at most 5 (RETRY_MAX_ATTEMPTS_EXCEEDED) |
base-backoff |
50ms |
|
max-backoff |
200ms |
|
jitter |
FULL |
NONE | FULL | DECORRELATED |
retry-after |
HONOR |
HONOR | IGNORE | CAP |
budget |
— | Shared token bucket name. A retrying profile that names none gets its own bucket, with the same pool-derived capacity (max-total-connections / 10, at least 1, refilled each minute) |
tls
| Property | Default | Meaning |
|---|---|---|
profile-id |
— | Required in production; the only TLS identifier the actuator exposes |
protocols |
TLSv1.3, TLSv1.2 |
Anything else is rejected |
hostname-verification |
true |
Setting it false fails startup |
trust-all |
false |
Exists only so the unsafe intent is rejectable; nothing acts on true |
allow-plain-http |
false |
Plaintext fallback fails startup in production |
trust-material-reference |
— | Custom CA, resolved by the secret loader |
key-material-reference |
— | Client certificate for mTLS |
proxy and observability
| Property | Default | Meaning |
|---|---|---|
proxy.enabled |
false |
|
proxy.host / proxy.port / proxy.type |
— / 0 / HTTP |
|
proxy.credential-provider |
— | Proxy authentication is separate from target authentication |
proxy.import-ambient-no-proxy |
false |
Ambient NO_PROXY never widens a validated profile |
observability.full-url-recording |
false |
|
observability.body-logging |
false |
app.httpclient.dynamic-targets[N]
| Property | Default | Meaning |
|---|---|---|
name |
— | Required, unique, and subject to the same normalisation rule as a client name |
allowed-schemes |
https |
|
allowed-ports |
443 |
|
allowed-host-suffixes |
empty | |
allowed-hosts |
empty | Empty means "any host that survives address validation" |
max-redirect-hops |
0 |
Each hop repeats the full validation flow |
trace-propagation |
false |
Off by default for dynamic targets |
blocked-cidrs |
empty | Organisation-defined internal ranges |
Startup violation codes
TRUSTED_BASE_URL_REQUIRED, BASE_URL_USERINFO_FORBIDDEN, BASE_URL_QUERY_FORBIDDEN,
PLAINTEXT_PRODUCTION_TARGET, ALLOWED_HOST_MISMATCH, ALLOWED_PORT_MISMATCH,
REDIRECT_POLICY_INVALID, REDIRECT_CROSS_ORIGIN_CREDENTIAL_POLICY_REQUIRED,
INVALID_TIMEOUT_BUDGET, RESPONSE_HARD_MAXIMUM_EXCEEDED, PRODUCTION_SIMPLE_FACTORY_FORBIDDEN,
JDK_FINE_GRAINED_POOL_UNSUPPORTED, HTTP3_STABLE_FORBIDDEN,
DYNAMIC_TARGET_TRANSPORT_UNSUPPORTED, DYNAMIC_DEFAULT_CREDENTIAL_FORBIDDEN,
OAUTH2_REGISTRATION_REQUIRED, API_KEY_HEADER_NAME_REQUIRED, TRUST_ALL_FORBIDDEN,
HOSTNAME_VERIFICATION_REQUIRED, PLAINTEXT_FALLBACK_FORBIDDEN, TLS_PROTOCOL_FORBIDDEN,
RETRY_BACKOFF_REQUIRED, MISSING_PRODUCTION_SETTING, DUPLICATE_CLIENT_NAME,
HTTPCLIENT_ACTIVE_WITHOUT_CLIENTS, DYNAMIC_BASE_URL_REQUIRED,
DYNAMIC_TARGET_PROXY_UNSUPPORTED,
HTTP2_REQUIRED_TRANSPORT_UNSUPPORTED, POOL_ROUTE_EXCEEDS_TOTAL,
TLS_PROTOCOL_SET_REQUIRED,
RETRY_POLICY_CONTRADICTS_ATTEMPTS, FULL_URL_RECORDING_FORBIDDEN, BODY_LOGGING_FORBIDDEN,
DNS_TIMEOUT_UNSUPPORTED, PROXY_CREDENTIAL_UNSUPPORTED, PROXY_AMBIENT_NO_PROXY_UNSUPPORTED,
TIMEOUT_NOT_POSITIVE, RETRY_MAX_ATTEMPTS_EXCEEDED, CIRCUIT_BREAKER_SETTING_INVALID.
A key that names no setting is not a violation code: the shared strict binder
(adapter:outbound:support StrictSettingsBinder) refuses it while binding, at any depth. Spring's
own unbound-key check stops at the first level under a list element, so clients[0].timeout.<typo>
used to bind nothing and fail nothing; the binder now refuses a key under clients[N] or
dynamic-targets[N] that bound nothing, for every capability rather than for this platform alone
(review 2026-11-01 05-15). The platform's own guard for it, and its HTTPCLIENT_UNKNOWN_SETTING
code, were removed with that change.
DNS_TIMEOUT_UNSUPPORTED, PROXY_CREDENTIAL_UNSUPPORTED and PROXY_AMBIENT_NO_PROXY_UNSUPPORTED name settings the platform binds but cannot yet honour. Neither the Apache classic
client nor the JDK client exposes a DNS-resolution timeout, and no proxy-credential path exists, so
a non-default value is refused rather than accepted and ignored. Leaving the defaults alone is
unaffected — only a deliberate, unmet request fails.
Three of these are about a guarantee that used to be silently unmet rather than refused:
HTTP2_REQUIRED_TRANSPORT_UNSUPPORTED— declaringprotocols: [HTTP_2]alone states that HTTP/2 is required. No transport can be configured to offer H2 and nothing else; the JDK client treats it as a preference and negotiates HTTP/1.1, and Apache's classic client is HTTP/1.1 only.POOL_ROUTE_EXCEEDS_TOTAL— a per-route ceiling above the total is incoherent.TLS_PROTOCOL_SET_REQUIRED— an emptytls.protocolsused to pass and then let the JVM choose, so emptying the list to "tighten" a profile loosened it.RETRY_POLICY_CONTRADICTS_ATTEMPTS—retry.policywas read by nothing on the execution path, so the actuator could reportnonefor a profile retrying three times. The two settings must now agree:policy: nonerequiresmax-attempts: 1, and any other policy requires more than one.FULL_URL_RECORDING_FORBIDDEN/BODY_LOGGING_FORBIDDEN— both settings were bindable and inert. Recording an expanded URL puts path identifiers and query strings into unbounded metric tags; recording bodies puts someone else's data into logs. Representable so the intent is rejectable, refused under a production profile.
DYNAMIC_TARGET_PROXY_UNSUPPORTED is worth spelling out: a forward proxy resolves the hostname on
its own side, so the addresses this platform validated and pinned are not the addresses the
connection reaches. The SSRF defence would be present, correct, and bypassed — so the combination is
refused rather than served with a guarantee it cannot keep.