如何为.NET中Google PredictionServiceClient流式调用实现重试逻辑？

需求 #

需要为Google.Cloud.AIPlatform.V1 NuGet包中的PredictionServiceClient实现重试逻辑，触发重试的场景包括：

• 请求超出配额限制（HTTP状态码429）
• 服务不可用（gRPC状态码Unavailable）
  要求所有PredictionServiceClient的调用使用统一的重试和超时配置。

参考实现（Azure OpenAI） #

同一应用中使用Azure OpenAI的Azure.AI.OpenAI包时，可在客户端初始化阶段直接配置重试参数，示例代码如下：

OpenAIClientOptions opts = new OpenAIClientOptions();
opts.Retry.MaxRetries = 2;
opts.Retry.Delay = TimeSpan.FromSeconds(60);
opts.Retry.NetworkTimeout = TimeSpan.FromSeconds(100);
opts.Retry.Mode = (RetryMode)connDetails.AzureOpenAIRetryMode.Value;
client = new OpenAIClient(new Uri("sampleUrl"), new AzureKeyCredential("sampleKey"), opts);

当前实现与报错 #

当前使用PredictionServiceClient进行流式内容生成的代码：

Google.Cloud.AIPlatform.V1.PredictionServiceClient.StreamGenerateContentStream response = myPredictionServiceClient.StreamGenerateContent(generateContentRequest);

尝试通过PredictionServiceSettings配置全局重试规则，但调用StreamGenerateContent时抛出错误：

PredictionServiceSettings settings = new PredictionServiceSettings
{
    CallSettings = CallSettings.FromRetry(RetrySettings.FromExponentialBackoff(
                                    maxAttempts: 2,
                                    initialBackoff: TimeSpan.FromSeconds(1),
                                    maxBackoff: TimeSpan.FromSeconds(10),
                                    backoffMultiplier: 2,
                                    retryFilter: RetrySettings.FilterForStatusCodes(StatusCode.Unavailable)
                                    )).WithTimeout(TimeSpan.FromSeconds(100))
};

bardClient = await new PredictionServiceClientBuilder
{
    Settings = settings,
    Endpoint = connDetails.AzureOpenAIResourceUrl
}.BuildAsync();

错误详情：

HResult=0x80131509
Message=Retry not permitted for this operation type
Source=Google.Api.Gax.Grpc
StackTrace:
 at Google.Api.Gax.Grpc.CallSettingsExtensions.ValidateNoRetry(CallSettings callSettings)
 at Google.Api.Gax.Grpc.ApiServerStreamingCall.<>c__DisplayClass0_0`2.<Create>b__1(TRequest req, CallSettings cs)
 at Google.Cloud.AIPlatform.V1.PredictionServiceClientImpl.StreamGenerateContent(GenerateContentRequest request, CallSettings callSettings) in Google.Cloud.AIPlatform.V1\PredictionServiceClientImpl.cs:line 250

问题 #

如何为PredictionServiceClient的StreamGenerateContent流式调用实现符合要求的重试逻辑？需要示例代码或实现指引。

内容的提问来源于stack exchange，提问作者singh_v