feat(unreachable-cooldown): 连接层不可达客户端级冷却,四家判定下沉共用

BACKLOG B24(窗长 300s 已拍板):连续 2 次连接未建立即对该 host 进
冷却窗,窗内调用零探针快速失败(归因文案与调用内短路逐字一致);到期
后下一条真实调用兼作探针,失败重进、成功恢复。isConnectFailure 四处
副本(BrandCheckClient/AppearancePatentLlmClient/SimilarAsinLlmClient/
TaskOwnerForwardService)下沉 common/retry 共用。三家客户端接入同一
冷却表(两个 LLM 客户端共用 ai.t8star.org 一格,16890 独立一格),
窗长可配(0=关闭退回调用内短路)。另修 CollectDataBrandBatchFilter
首批预算竞态(1ms 预算下 GC 停顿误切首批,照 BrandCheckClient 首轮
不掐契约补 start>0 守卫)。变异验证:冷却禁用 6 例判红;定向 47/47 绿。
This commit is contained in:
2026-10-23 20:58:32 +08:00
parent 6af19b51ba
commit 09dfb01f9c
22 changed files with 426 additions and 99 deletions
@@ -0,0 +1,103 @@
package com.nanri.aiimage.common.retry;
import org.springframework.stereotype.Component;
import java.net.ConnectException;
import java.net.NoRouteToHostException;
import java.net.UnknownHostException;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLong;
/**
* 连接层不可达的客户端级冷却表(按上游 host 分键,Spring 单例)。
*
* <p>背景(BACKLOG B24,2026-10-23 拍板):三家同型客户端({@code BrandCheckClient}/
* {@code AppearancePatentLlmClient}/{@code SimilarAsinLlmClient})的连接层短路计数
* 生命周期都只到单次调用,网关长断时每个新调用/分块仍要各自真发连接探针才知道"还是断的"。
* 本表把「连续 2 次连接未建立」的证据跨调用累计,达到阈值即对该 host 进入冷却窗
* (默认 300s,同仓两处先例:{@code pt_worker/proxy_pool.py} 与 16890 服务侧的钥匙冷却),
* 窗内调用方不发请求直接快速失败;到期后下一条真实调用兼作探针,失败即重进、成功即恢复。
*
* <p>调用方契约:{@link #isCoolingDown} 在发请求前问一次;连接层失败记
* {@link #recordConnectFailure}(返回 true 的那方打一条进入冷却的 warn);
* 连接成功或"连上了没答好"(读超时/5xx/解析失败)记 {@link #recordConnected} 清零证据。
* 窗长 {@code <= 0} 表示关闭冷却,全部方法退化为无操作——调用方退回各自的调用内短路行为。
*/
@Component
public class UnreachableCooldown {
/**
* 进入冷却的证据数:与三家客户端各自的调用内短路阈值同数字——一次 refused 可能是抖动,
* 两次同形状且中间没有任何一次成功,才是服务在掉。
*/
private static final int EVIDENCE_THRESHOLD = 2;
private final ConcurrentHashMap<String, State> states = new ConcurrentHashMap<>();
private static final class State {
final AtomicInteger evidence = new AtomicInteger();
final AtomicLong cooldownUntilMillis = new AtomicLong();
}
/**
* 只认连接层故障(顺着 cause 链找):连接被拒/无路由/主机名解不了。
* 读超时与 5xx 是"连上了没答好",故意不算——那类抖动重试是有意义的。
* 原 BrandCheckClient/AppearancePatentLlmClient/SimilarAsinLlmClient 各有一份同款,
* B24 下沉至此,四家(含 TaskOwnerForwardService)共用同一判定。
*/
public static boolean isConnectFailure(Throwable error) {
for (Throwable t = error; t != null && t != t.getCause(); t = t.getCause()) {
if (t instanceof ConnectException || t instanceof NoRouteToHostException
|| t instanceof UnknownHostException) {
return true;
}
}
return false;
}
/** 该 host 是否正处冷却窗内:true 时调用方不发请求,直接按既有不可达归因文案快速失败。 */
public boolean isCoolingDown(String host, long windowMillis) {
if (windowMillis <= 0) {
return false;
}
State state = states.get(normalizeHost(host));
return state != null && state.cooldownUntilMillis.get() > System.currentTimeMillis();
}
/**
* 连接层失败证据 +1;累计到阈值即对该 host 开(或重开)冷却窗。
*
* @return true 仅当"本次调用让冷却窗从无到有"(并发下只会有一个调用方拿到 true),
* 调用方据此打一条进入冷却的 warn,期中快失败不刷屏
*/
public boolean recordConnectFailure(String host, long windowMillis) {
if (windowMillis <= 0) {
return false;
}
State state = states.computeIfAbsent(normalizeHost(host), key -> new State());
if (state.evidence.incrementAndGet() < EVIDENCE_THRESHOLD) {
return false;
}
long now = System.currentTimeMillis();
long previous = state.cooldownUntilMillis.get();
if (previous > now) {
return false;
}
return state.cooldownUntilMillis.compareAndSet(previous, now + windowMillis);
}
/** 连接成功或"连上了没答好":证据清零(已开的冷却窗不受影响,等它自然到期)。 */
public void recordConnected(String host) {
State state = states.get(normalizeHost(host));
if (state != null) {
state.evidence.set(0);
}
}
/** host 归一:两个 LLM 客户端配的同一网关必须落进同一格,别因尾斜杠/大小写裂成两份。 */
static String normalizeHost(String host) {
String normalized = host == null ? "" : host.trim().toLowerCase();
return normalized.endsWith("/") ? normalized.substring(0, normalized.length() - 1) : normalized;
}
}
@@ -19,7 +19,6 @@ import org.springframework.web.client.RestClient;
import org.springframework.web.util.ContentCachingRequestWrapper;
import java.io.IOException;
import java.net.ConnectException;
import java.util.Enumeration;
import java.util.List;
import java.util.Set;
@@ -89,7 +88,7 @@ public class TaskOwnerForwardService {
.retrieve()
.toEntity(byte[].class);
} catch (ResourceAccessException accessError) {
if (!isConnectFailure(accessError)) {
if (!com.nanri.aiimage.common.retry.UnreachableCooldown.isConnectFailure(accessError)) {
throw accessError;
}
lastError = accessError;
@@ -103,17 +102,6 @@ public class TaskOwnerForwardService {
throw lastError;
}
private static boolean isConnectFailure(Throwable error) {
Throwable cursor = error;
while (cursor != null) {
if (cursor instanceof ConnectException) {
return true;
}
cursor = cursor.getCause();
}
return false;
}
private static void sleepQuietly(long millis) {
try {
Thread.sleep(millis);
@@ -46,6 +46,13 @@ public class AppearancePatentProperties {
* 每行每个 LLM 请求的重试次数(含首次)
*/
private int llmRetryTimes = 3;
/**
* 连接层不可达的客户端级冷却窗长(0=关闭,退回调用内短路)。连续 2 次连接未建立即进窗,
* 窗内调用不发请求直接快速失败;到期后下一条真实调用兼作探针。拍板口径见 BACKLOG B24
* (300s 同仓两处先例:pt_worker 代理池与 16890 服务侧的钥匙冷却)。
*/
private int llmUnreachableCooldownMillis = 300000;
/**
* 单行最多被「轮次」发起多少次(一轮 = 一次批量派发;每轮内部另有 llmRetryTimes 次 HTTP 重试)。
* <=0 关闭本闸。
@@ -34,4 +34,11 @@ public class BrandCheckProperties {
private int totalTimeoutMillis = 90000;
private int connectTimeoutMillis = 10000;
private int readTimeoutMillis = 60000;
/**
* 连接层不可达的客户端级冷却窗长(0=关闭,退回批内短路)。连续 2 次连接未建立即进窗,
* 窗内整批品牌不发请求直接按服务不可达判失败;到期后下一批的真实请求兼作探针。
* 拍板口径见 BACKLOG B24(300s 同仓两处先例:pt_worker 代理池与 16890 服务侧的钥匙冷却)。
*/
private int unreachableCooldownMillis = 300000;
}
@@ -197,6 +197,13 @@ public class SimilarAsinProperties {
private int llmFirstAttemptReadTimeoutMillis = 300000;
private int llmRetryTimes = 3;
/**
* 连接层不可达的客户端级冷却窗长(0=关闭,退回调用内短路)。连续 2 次连接未建立即进窗,
* 窗内调用不发请求直接快速失败;到期后下一条真实调用兼作探针。拍板口径见 BACKLOG B24
* (300s 同仓两处先例:pt_worker 代理池与 16890 服务侧的钥匙冷却)。
*/
private int llmUnreachableCooldownMillis = 300000;
/** 批内行级并发上限:每行最多 2 次图片对比 + 1 次合规 + 3 次类目匹配。 */
private int llmRowConcurrency = 5;
@@ -3,6 +3,7 @@ package com.nanri.aiimage.modules.appearancepatent.client;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.nanri.aiimage.common.retry.LlmRetryBackoff;
import com.nanri.aiimage.common.retry.UnreachableCooldown;
import com.nanri.aiimage.config.AppearancePatentProperties;
import com.nanri.aiimage.config.HttpClientPool;
import com.nanri.aiimage.metrics.ExternalCallMetricsRecorder;
@@ -21,9 +22,6 @@ import org.springframework.web.client.RestClient;
import java.io.IOException;
import java.io.InputStream;
import java.net.ConnectException;
import java.net.NoRouteToHostException;
import java.net.UnknownHostException;
import java.nio.charset.StandardCharsets;
import java.time.Duration;
import java.util.ArrayList;
@@ -158,6 +156,7 @@ public class AppearancePatentLlmClient {
private final ExternalCallMetricsRecorder externalCallMetrics;
private final BrandCheckClient brandCheckClient;
private final UserSecretUsageService userSecretUsageService;
private final UnreachableCooldown unreachableCooldown;
private volatile RestClient sharedRestClient;
/** 首次尝试专用客户端(更短读超时,快速失败);配置未生效时与 sharedRestClient 相同。 */
@@ -647,15 +646,30 @@ public class AppearancePatentLlmClient {
String apiKey,
String responseFormat) throws Exception {
int attempts = Math.max(1, properties.getLlmRetryTimes());
long cooldownMillis = Math.max(0, properties.getLlmUnreachableCooldownMillis());
String host = properties.getLlmHost();
if (unreachableCooldown.isCoolingDown(host, cooldownMillis)) {
// 冷却期快失败不发请求、不记密钥次数(B24)
log.debug("[appearance-patent] 客户端级冷却中,快速失败 model={} window={}ms", model, cooldownMillis);
throw new IllegalStateException(REASON_LLM_UNREACHABLE + ";lastErr=冷却中(连续连接未建立,暂不打上游)");
}
Exception lastFailure = null;
AtomicInteger connectFailures = new AtomicInteger();
for (int attempt = 1; attempt <= attempts; attempt++) {
try {
return invokeChatOnce(model, system, userText, images, apiKey, responseFormat, attempt);
String content = invokeChatOnce(model, system, userText, images, apiKey, responseFormat, attempt);
unreachableCooldown.recordConnected(host);
return content;
} catch (Exception ex) {
lastFailure = ex;
if (isConnectFailure(ex)) {
if (connectFailures.incrementAndGet() >= LLM_CONNECT_FAILURE_EVIDENCE) {
if (UnreachableCooldown.isConnectFailure(ex)) {
boolean entered = unreachableCooldown.recordConnectFailure(host, cooldownMillis);
if (entered) {
log.warn("[appearance-patent] 连接层证据累计 {} 次,进入客户端级冷却 {}ms host={}",
LLM_CONNECT_FAILURE_EVIDENCE, cooldownMillis, host);
}
if (connectFailures.incrementAndGet() >= LLM_CONNECT_FAILURE_EVIDENCE
|| unreachableCooldown.isCoolingDown(host, cooldownMillis)) {
log.warn("[appearance-patent] LLM_UNREACHABLE 连续 {} 次连接未建立,判定 LLM 服务不可达,"
+ "停止重试 attempt={} model={} err={}",
LLM_CONNECT_FAILURE_EVIDENCE, attempt, model, failureMessage(ex));
@@ -667,6 +681,7 @@ public class AppearancePatentLlmClient {
} else {
// 连上了只是这次没答好(读超时/解析失败/5xx):连接层证据不作数
connectFailures.set(0);
unreachableCooldown.recordConnected(host);
}
if (attempt >= attempts) {
break;
@@ -681,17 +696,11 @@ public class AppearancePatentLlmClient {
}
/**
* 只认连接层故障(顺着 cause 链找):连接被拒/无路由/主机名解不了。
* 读超时与 5xx 是"连上了没答好",故意不算——那类抖动重试是有意义的。
* 连接层判定下沉 {@code UnreachableCooldown.isConnectFailure}(B24):判定口径与
* BrandCheckClient/SimilarAsinLlmClient 共用一份,不再各留包内副本。
*/
static boolean isConnectFailure(Throwable error) {
for (Throwable t = error; t != null && t != t.getCause(); t = t.getCause()) {
if (t instanceof ConnectException || t instanceof NoRouteToHostException
|| t instanceof UnknownHostException) {
return true;
}
}
return false;
return UnreachableCooldown.isConnectFailure(error);
}
private String invokeChatOnce(String model,
@@ -77,16 +77,19 @@ public class BrandCheckClient {
private final BrandCheckProperties properties;
private final com.nanri.aiimage.metrics.ExternalCallMetricsRecorder externalCallMetrics;
private final UserApiSecretService userApiSecretService;
private final com.nanri.aiimage.common.retry.UnreachableCooldown unreachableCooldown;
/** 提交线程上的拒绝标记:队列满时由拒绝策略置位,提交方据此立即降级(不新增 catch 出口)。 */
private final ThreadLocal<AtomicBoolean> rejectedSubmission = new ThreadLocal<>();
private volatile RestClient sharedRestClient;
public BrandCheckClient(BrandCheckProperties properties,
com.nanri.aiimage.metrics.ExternalCallMetricsRecorder externalCallMetrics,
UserApiSecretService userApiSecretService) {
UserApiSecretService userApiSecretService,
com.nanri.aiimage.common.retry.UnreachableCooldown unreachableCooldown) {
this.properties = properties;
this.externalCallMetrics = externalCallMetrics;
this.userApiSecretService = userApiSecretService;
this.unreachableCooldown = unreachableCooldown;
}
// 有界队列 + 拒绝即降级:大批量(几千品牌)检查时队列洪峰不无界堆内存。
@@ -267,6 +270,18 @@ public class BrandCheckClient {
}
return outcomes;
}
long cooldownMillis = Math.max(0, properties.getUnreachableCooldownMillis());
if (unreachableCooldown.isCoolingDown(properties.getBaseUrl(), cooldownMillis)) {
// 客户端级冷却(B24):上一批已确认连接层不可达,本批连首轮探针都不发。
// 结论与批内短路版逐字一致(每行仍判查询失败、reason 同文案),差别只是不再为
// "确认还是断的"每批各烧一轮连接探针;到期后下一批的真实请求兼作探针。
log.warn("[brand-check] SERVICE_UNREACHABLE 客户端级冷却中,整批不打上游 brands={} window={}ms",
brands.size(), cooldownMillis);
for (String brand : brands) {
outcomes.put(brand, new BrandCheckOutcome(List.of(), List.of(brand), REASON_SERVICE_UNREACHABLE));
}
return outcomes;
}
long firstAttemptDeadlineNanos = deadlineNanos == Long.MAX_VALUE
? Long.MAX_VALUE
: deadlineNanos + FIRST_ATTEMPT_QUEUE_GRACE_MILLIS * 1_000_000L;
@@ -363,6 +378,7 @@ public class BrandCheckClient {
lastFailure = ex;
response = null;
if (isConnectFailure(ex)) {
recordConnectFailure(ex);
if (connectFailures.incrementAndGet() >= CONNECT_FAILURE_EVIDENCE) {
ladderCompleted = false;
connectDown = true;
@@ -374,6 +390,7 @@ public class BrandCheckClient {
} else {
// 连上了只是这次没答好(读超时/解析失败):连接层证据不作数
connectFailures.set(0);
unreachableCooldown.recordConnected(properties.getBaseUrl());
}
if (attempt < attempts) {
long retryDelayMillis = retryDelayMillis(attempt);
@@ -385,6 +402,7 @@ public class BrandCheckClient {
}
// 拿到响应=连接是通的,证据清零,别跨到下一次真断连时凑数
connectFailures.set(0);
unreachableCooldown.recordConnected(properties.getBaseUrl());
if (response != null && nullToEmpty(response.getQueryFaildData()).isEmpty()) {
// 本次查询完成且无失败项:命中侵权(faild_data)属有效结论,不重试。
return new BrandCheckOutcome(nullToEmpty(response.getFaildData()), List.of(), "");
@@ -554,14 +572,21 @@ public class BrandCheckClient {
* <p>故意不含读超时与 5xx:那两种是"连上了但这次没答好",重试救得回来;把它们算进
* 不可达证据,就会把一次上游抖动误判成服务在掉、把同批其余品牌的重试一起掐掉。
*/
static boolean isConnectFailure(Throwable error) {
for (Throwable t = error; t != null && t != t.getCause(); t = t.getCause()) {
if (t instanceof ConnectException || t instanceof NoRouteToHostException
|| t instanceof UnknownHostException) {
return true;
}
/** 连接层失败记入客户端级冷却表:达到阈值的那方打一条进入冷却的 warn(B24)。 */
private void recordConnectFailure(Exception ex) {
long cooldownMillis = Math.max(0, properties.getUnreachableCooldownMillis());
if (unreachableCooldown.recordConnectFailure(properties.getBaseUrl(), cooldownMillis)) {
log.warn("[brand-check] SERVICE_UNREACHABLE 连接层证据累计 {} 次,进入客户端级冷却 {}ms host={}",
CONNECT_FAILURE_EVIDENCE, cooldownMillis, properties.getBaseUrl());
}
return false;
}
/**
* 连接层判定下沉 {@code UnreachableCooldown.isConnectFailure}(B24):判定口径与
* 两个 LLM 客户端共用一份,不再各留包内副本。
*/
static boolean isConnectFailure(Throwable error) {
return com.nanri.aiimage.common.retry.UnreachableCooldown.isConnectFailure(error);
}
/**
@@ -118,7 +118,10 @@ public class CollectDataBrandBatchFilter {
}
continue;
}
if (deadlineNanos != Long.MAX_VALUE && System.nanoTime() >= deadlineNanos) {
// 首批不掐(与 BrandCheckClient 的 attempt>1 同契约):预算检查只拦后续批次。
// 对首批也查的话,deadline 计算与检查之间一次 GC/调度停顿(毫秒级)就会把首批
// 整组误切成 queryFailed——1ms 预算的测试已实测抖出过这个形状。
if (start > 0 && deadlineNanos != Long.MAX_VALUE && System.nanoTime() >= deadlineNanos) {
// 预算用尽:不再为后面的批次捶上游。已缓存的品牌仍按缓存结论归类(不花代价),
// 没结论的行归 queryFailed 入补测队列——既不按 accepted 放行(可能正是侵权品牌),
// 也不按 rejected 拉黑(没查过就不该写进无效品牌表,否则永久误杀)。
@@ -3,6 +3,7 @@ package com.nanri.aiimage.modules.similarasin.client;
import com.fasterxml.jackson.databind.JsonNode;
import com.fasterxml.jackson.databind.ObjectMapper;
import com.nanri.aiimage.common.retry.LlmRetryBackoff;
import com.nanri.aiimage.common.retry.UnreachableCooldown;
import com.nanri.aiimage.config.HttpClientPool;
import com.nanri.aiimage.config.SimilarAsinProperties;
import com.nanri.aiimage.metrics.ExternalCallMetricsRecorder;
@@ -17,9 +18,6 @@ import org.springframework.web.client.RestClient;
import java.io.IOException;
import java.io.InputStream;
import java.net.ConnectException;
import java.net.NoRouteToHostException;
import java.net.UnknownHostException;
import java.nio.charset.StandardCharsets;
import java.time.Duration;
import java.util.ArrayList;
@@ -52,6 +50,7 @@ public class SimilarAsinLlmClient {
private final ObjectMapper objectMapper;
private final ExternalCallMetricsRecorder externalCallMetrics;
private final UserSecretUsageService userSecretUsageService;
private final UnreachableCooldown unreachableCooldown;
private volatile RestClient sharedRestClient;
/** 首次尝试专用客户端(更短读超时,快速失败);配置未生效时与 sharedRestClient 相同。 */
@@ -60,11 +59,13 @@ public class SimilarAsinLlmClient {
public SimilarAsinLlmClient(SimilarAsinProperties properties,
ObjectMapper objectMapper,
ExternalCallMetricsRecorder externalCallMetrics,
UserSecretUsageService userSecretUsageService) {
UserSecretUsageService userSecretUsageService,
UnreachableCooldown unreachableCooldown) {
this.properties = properties;
this.objectMapper = objectMapper;
this.externalCallMetrics = externalCallMetrics;
this.userSecretUsageService = userSecretUsageService;
this.unreachableCooldown = unreachableCooldown;
}
/** 文本对话,json_object 输出。 */
@@ -85,16 +86,32 @@ public class SimilarAsinLlmClient {
String responseFormat) {
String resolvedKey = resolveApiKey(apiKey);
int attempts = Math.max(1, properties.getLlmRetryTimes());
long cooldownMillis = Math.max(0, properties.getLlmUnreachableCooldownMillis());
String host = properties.getLlmHost();
if (unreachableCooldown.isCoolingDown(host, cooldownMillis)) {
// 冷却期快失败不发请求、不记密钥次数:结论与短路版一致(调用方拿不到 LLM 结论),
// 差别只是不再为"确认还是断的"各烧一轮连接探针(B24)
log.debug("[similar-asin][llm] 客户端级冷却中,快速失败 model={} window={}ms", model, cooldownMillis);
throw new IllegalStateException(REASON_LLM_UNREACHABLE + ";lastErr=冷却中(连续连接未建立,暂不打上游)");
}
Exception lastFailure = null;
int connectFailures = 0;
for (int attempt = 1; attempt <= attempts; attempt++) {
try {
return invokeChatOnce(model, system, userText, images, resolvedKey, responseFormat, attempt);
String content = invokeChatOnce(model, system, userText, images, resolvedKey, responseFormat, attempt);
unreachableCooldown.recordConnected(host);
return content;
} catch (Exception ex) {
lastFailure = ex;
if (isConnectFailure(ex)) {
if (UnreachableCooldown.isConnectFailure(ex)) {
connectFailures++;
if (connectFailures >= LLM_CONNECT_FAILURE_EVIDENCE) {
boolean entered = unreachableCooldown.recordConnectFailure(host, cooldownMillis);
if (entered) {
log.warn("[similar-asin][llm] 连接层证据累计 {} 次,进入客户端级冷却 {}ms host={}",
2, cooldownMillis, host);
}
if (connectFailures >= LLM_CONNECT_FAILURE_EVIDENCE
|| unreachableCooldown.isCoolingDown(host, cooldownMillis)) {
log.warn("[similar-asin][llm] LLM_UNREACHABLE 连续 {} 次连接未建立,判定 LLM 服务不可达,"
+ "停止重试 attempt={} model={} err={}",
LLM_CONNECT_FAILURE_EVIDENCE, attempt, model, failureMessage(ex));
@@ -106,6 +123,7 @@ public class SimilarAsinLlmClient {
} else {
// 连上了只是这次没答好(读超时/解析失败/5xx):连接层证据不作数
connectFailures = 0;
unreachableCooldown.recordConnected(host);
}
if (attempt >= attempts) {
break;
@@ -123,20 +141,6 @@ public class SimilarAsinLlmClient {
: new IllegalStateException(lastFailure.getMessage(), lastFailure);
}
/**
* 只认连接层故障(顺着 cause 链找):连接被拒/无路由/主机名解不了。
* 读超时与 5xx 是"连上了没答好",故意不算——那类抖动重试是有意义的。
*/
static boolean isConnectFailure(Throwable error) {
for (Throwable t = error; t != null && t != t.getCause(); t = t.getCause()) {
if (t instanceof ConnectException || t instanceof NoRouteToHostException
|| t instanceof UnknownHostException) {
return true;
}
}
return false;
}
/** 只认任务/用户自己的 api_key(payload → 按 uid 的服务端密钥表已在 readApiKey 兜过);
* 为空即空、跳过 LLM,绝不回落平台公费密钥 —— 各用户的任务只能烧各自的密钥。 */
public String resolveApiKey(String apiKey) {
@@ -378,6 +378,9 @@ aiimage:
total-timeout-millis: ${AIIMAGE_BRAND_CHECK_TOTAL_TIMEOUT_MILLIS:90000}
connect-timeout-millis: ${AIIMAGE_BRAND_CHECK_CONNECT_TIMEOUT_MILLIS:10000}
read-timeout-millis: ${AIIMAGE_BRAND_CHECK_READ_TIMEOUT_MILLIS:60000}
# 连接层不可达的客户端级冷却窗长(B24,0=关闭):连续 2 次连接未建立即进窗,
# 窗内整批品牌不发请求直接按服务不可达判失败;到期后下一批的真实请求兼作探针。
unreachable-cooldown-millis: ${AIIMAGE_BRAND_CHECK_UNREACHABLE_COOLDOWN_MILLIS:300000}
appearance-patent:
no-result-upload-timeout-minutes: ${AIIMAGE_APPEARANCE_PATENT_NO_RESULT_UPLOAD_TIMEOUT_MINUTES:180}
llm-host: ${AIIMAGE_APPEARANCE_PATENT_LLM_HOST:https://ai.t8star.org}
@@ -402,6 +405,9 @@ aiimage:
# 实测空行 20.7%~51.3%,全部放行且零告警,因为旧分母按解析行标题判、用户表不带标题列时分母为 0)。
llm-coverage-min-ratio: ${AIIMAGE_APPEARANCE_PATENT_LLM_COVERAGE_MIN_RATIO:0.95}
llm-retry-times: ${AIIMAGE_APPEARANCE_PATENT_LLM_RETRY_TIMES:3}
# 连接层不可达的客户端级冷却窗长(B24,0=关闭):连续 2 次连接未建立即进窗,
# 窗内调用不发请求直接按服务不可达快速失败;到期后下一条真实调用兼作探针。
llm-unreachable-cooldown-millis: ${AIIMAGE_APPEARANCE_PATENT_LLM_UNREACHABLE_COOLDOWN_MILLIS:300000}
# 单行最多被重发几轮(<=0 关闭)。失败行会被释放、下一轮重新申领,没有这道闸时
# 上游劣化窗口内会无限重发并每轮扣用户 LLM 次数——deferRunning 不累加 retryCount,任务层那道永不触发。
llm-row-max-rounds: ${AIIMAGE_APPEARANCE_PATENT_LLM_ROW_MAX_ROUNDS:3}
@@ -440,6 +446,9 @@ aiimage:
llm-read-timeout-millis: ${AIIMAGE_SIMILAR_ASIN_LLM_READ_TIMEOUT_MILLIS:600000}
llm-first-attempt-read-timeout-millis: ${AIIMAGE_SIMILAR_ASIN_LLM_FIRST_ATTEMPT_READ_TIMEOUT_MILLIS:300000}
llm-retry-times: ${AIIMAGE_SIMILAR_ASIN_LLM_RETRY_TIMES:3}
# 连接层不可达的客户端级冷却窗长(B24,0=关闭):连续 2 次连接未建立即进窗,
# 窗内调用不发请求直接按服务不可达快速失败;到期后下一条真实调用兼作探针。
llm-unreachable-cooldown-millis: ${AIIMAGE_SIMILAR_ASIN_LLM_UNREACHABLE_COOLDOWN_MILLIS:300000}
llm-row-concurrency: ${AIIMAGE_SIMILAR_ASIN_LLM_ROW_CONCURRENCY:5}
llm-category-batch-size: ${AIIMAGE_SIMILAR_ASIN_LLM_CATEGORY_BATCH_SIZE:10}
# 合规判定一次请求合并的行数:省往返不省判定,每件商品仍各自出 is_conform/reason/category。