diff --git a/backend-java/src/main/java/com/nanri/aiimage/config/TransientStorageProperties.java b/backend-java/src/main/java/com/nanri/aiimage/config/TransientStorageProperties.java
index aab05e41..8824cf71 100644
--- a/backend-java/src/main/java/com/nanri/aiimage/config/TransientStorageProperties.java
+++ b/backend-java/src/main/java/com/nanri/aiimage/config/TransientStorageProperties.java
@@ -37,13 +37,20 @@ public class TransientStorageProperties {
     private long failureCooldownMillis = 10000;
     private int dispatcherMaxRequests = 56;
     private int dispatcherMaxRequestsPerHost = 56;
+    /**
+     * 空闲连接保留数。
+     *
+     * <p>2026-09-17 曾试过设 0（彻底不复用）来验证"unexpected end of stream 是复用死连接导致的"
+     * 这一假设——**实测照旧失败**（新容器起来后第一次请求就中招）。至此已排除公网链路、
+     * keepAlive 过长、连接复用三项；用 mc 并发压 200 个小对象也全部成功，说明服务端没问题。
+     * 剩余方向指向 MinIO Java SDK / OkHttp 与 RustFS 的协议细节，故恢复默认的连接复用。
+     */
     private int connectionPoolMaxIdle = 5;
     /**
-     * 空闲连接在池里的保留时长。默认 5 分钟（OkHttp 原值）会让客户端复用"已被 RustFS
-     * 或中间设备关掉的空闲连接"，表现为 unexpected end of stream —— 线上每天上千次，
-     * 全靠重试兜底。对象存储访问是突发型，连接复用率本就低，缩短保活几乎没有代价。
+     * 空闲连接在池里的保留时长。曾由 300000 调到 30000 试图减少 unexpected end of stream，
+     * 实测无改善（该现象与连接复用无关，见 {@link #connectionPoolMaxIdle} 的排查记录），故恢复原值。
      */
-    private long connectionPoolKeepAliveMillis = 30000;
+    private long connectionPoolKeepAliveMillis = 300000;
     private long warnPayloadBytes = 5L * 1024 * 1024;
     private long maxPayloadBytes = 50L * 1024 * 1024;
     private long maxStoredPayloadBytes = 50L * 1024 * 1024;
diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/file/service/object/RustfsObjectStorageService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/file/service/object/RustfsObjectStorageService.java
index dc7f9145..355bb1bf 100644
--- a/backend-java/src/main/java/com/nanri/aiimage/modules/file/service/object/RustfsObjectStorageService.java
+++ b/backend-java/src/main/java/com/nanri/aiimage/modules/file/service/object/RustfsObjectStorageService.java
@@ -9,6 +9,7 @@ import io.minio.MinioClient;
 import io.minio.PutObjectArgs;
 import io.minio.RemoveObjectArgs;
 import io.minio.StatObjectArgs;
+import io.minio.errors.ErrorResponseException;
 import lombok.extern.slf4j.Slf4j;
 import okhttp3.ConnectionPool;
 import okhttp3.Dispatcher;
@@ -21,6 +22,7 @@ import java.io.ByteArrayInputStream;
 import java.io.ByteArrayOutputStream;
 import java.nio.charset.StandardCharsets;
 import java.util.Objects;
+import java.util.Set;
 import java.util.concurrent.Semaphore;
 import java.util.concurrent.ThreadLocalRandom;
 import java.util.concurrent.TimeUnit;
@@ -38,6 +40,11 @@ public class RustfsObjectStorageService {
     private static final String OP_STAT = "stat";
     private static final String OP_TOTAL = "total";
 
+    /** S3 确定性错误码：重试不会改变结果，应立即失败而不是白打两次请求。 */
+    private static final Set<String> NON_RETRYABLE_S3_CODES = Set.of(
+            "NoSuchKey", "NoSuchBucket", "NoSuchVersion",
+            "AccessDenied", "InvalidAccessKeyId", "SignatureDoesNotMatch", "InvalidBucketName");
+
     private final TransientStorageProperties properties;
     private final ObjectProvider<MeterRegistry> meterRegistryProvider;
     private final ObjectProvider<RustfsDeleteRetryService> deleteRetryServiceProvider;
@@ -251,6 +258,15 @@ public class RustfsObjectStorageService {
                         last = ex;
                         recordOperation(operation, attempt < maxRetries ? "retry" : "failure", elapsedNanos(startedAt));
                         recordFailure(operation, objectKey, ex);
+                        if (isNonRetryable(ex)) {
+                            // 确定性错误（NoSuchKey / AccessDenied…）：重试多少次结果都一样。
+                            // 线上 read 一个已被清理的 chunk 就会连打 3 次请求、还被记成 ERROR。
+                            log.warn("[rustfs] 确定性错误，不重试 operation={} objectKey={} err={}",
+                                    operation, objectKey, ex.getMessage());
+                            throw ex instanceof RuntimeException runtimeException
+                                    ? runtimeException
+                                    : new IllegalStateException(ex);
+                        }
                         if (attempt < maxRetries) {
                             delayMillis = retryDelayMillis(attempt);
                             log.warn("[rustfs] operation failed, retrying operation={} objectKey={} attempt={}/{} delayMs={} err={}",
@@ -279,6 +295,27 @@ public class RustfsObjectStorageService {
         }
     }
 
+    /**
+     * 是否为「重试也没用」的确定性错误（NoSuchKey / AccessDenied / 签名错误…）。
+     *
+     * <p>只有网络类与 5xx 类失败才值得重试；确定性错误重试多少次结果都一样。
+     */
+    private static boolean isNonRetryable(Throwable error) {
+        Throwable cursor = error;
+        while (cursor != null) {
+            if (cursor instanceof ErrorResponseException responseException) {
+                String code = responseException.errorResponse() == null
+                        ? null
+                        : responseException.errorResponse().code();
+                if (code != null && NON_RETRYABLE_S3_CODES.contains(code)) {
+                    return true;
+                }
+            }
+            cursor = cursor.getCause();
+        }
+        return false;
+    }
+
     private void acquirePermit(String operation, String objectKey, Semaphore semaphore, long deadlineNanos) {
         try {
             checkDeadline(operation, objectKey, deadlineNanos);