gRPC 客户端不使用服务器端流
Posted
技术标签:
【中文标题】gRPC 客户端不使用服务器端流【英文标题】：gRPC client not consuming server side stream 【发布时间】：2021-03-01 10:13:23 【问题描述】：我创建了一个 gRPC 服务器端流服务（在 Rust 中），它从 cluster A 上的 kafka 主题流式传输消息。每次外部客户端发起 websocket 连接时，cluster B 上的 Go 客户端都会连接到 rpc 服务；消息通过此连接流式传输到外部客户端。
然而，Go 客户端似乎并不能可靠地从 rust 服务器消费（即使消息确实已经提交）。即使 websocket 肯定是从外部客户端读取信息；有时它似乎不会从 grpc 流中消耗。我已尝试阅读文档并确定可能导致此问题的任何资源泄漏，但我无法解决问题。
我认为问题可能在于上下文处理或关闭连接（由于流最初成功，然后如果您关闭 websocket 并重新打开它，则消息无法发送）
Go 客户端代码
type KafkaChannel struct
Client bridge.KafkaStreamClient
func (kc *KafkaChannel) Consume(topic string) (<-chan []byte, func() error)
readChan := make(chan []byte)
// Context with cancellation to close the routine
ctx, cancel := context.WithCancel(context.Background())
stream, err := kc.Client.Consume(ctx, &bridge.ConsumeRequest
Topic: topic,
)
if err != nil
log.Fatalf("Error creating cosumer stream: %v", err)
// Launch listener in new thread
go func(reader *chan []byte, consumeStream *bridge.KafkaStream_ConsumeClient)
// Recover from a panic (resources consumed)
defer func()
if recover() != nil
log.Println("Consumer routine closed")
()
for
response, err := stream.Recv()
if err != nil
log.Printf("Error creating cosumer stream: %v", err)
break
switch data := response.OptionalContent.(type)
case *bridge.KafkaResponse_Content:
*reader <- *&data.Content
default:
break
(&readChan, &stream)
// Create a callback func that frees the resources
closeCallback := func() error
err := stream.CloseSend()
close(readChan)
cancel()
return err
return readChan, closeCallback
网络套接字
type ChatHandler struct
Database *db.Queries
Client *grpc.ClientConn
Context context.Context
SessionChannel chan []byte
func (handler *ChatHandler) GetChatConnection(c *websocket.Conn)
//initialisation...
consumer, closeConsume := kChannel.Consume(topic)
for msg := range consumer
log.Printf("Received message from bridge: %s", string(msg))
writeMessageStart := time.Now()
if err = c.WriteMessage(1, msg); err != nil
log.Printf("Error writing message: %v", err)
writeMessageElapsed := time.Since(writeMessageStart)
log.Printf("Write time elapsed error: %s", writeMessageElapsed)
if errors.Is(err, syscall.EPIPE)
log.Printf("Sys error: %v", err)
//continue
closeConsume()
handler.close(c)
return
writeMessageElapsed := time.Since(writeMessageStart)
log.Printf("Write time elapsed no error: %s", writeMessageElapsed)
Rust 服务器端代码
为了完整性
async fn consume(
&self,
request: Request<ConsumeRequest>,
) -> Result<Response<Self::ConsumeStream>, Status>
let (tx, rx) = mpsc::unbounded_channel();
info!("Initiated read-only stream");
tokio::spawn(async move
let message = match Some(request.get_ref())
Some(x) => x,
None => return,
;
let topic = message.topic.clone();
info!("Consuming on topic: ", topic);
let consumer = create_kafka_consumer(topic);
loop
let result = consumer.stream().next().await;
match result
None =>
warn!("Received none-type from consumer stream");
continue;
Some(Err(e)) =>
error!("Error consuming from kafka broker: :?", e);
continue;
Some(Ok(message)) =>
let payload = match message.payload_view::<str>()
None =>
warn!("Recived none-type when unpacking payload");
continue;
Some(Ok(s)) =>
info!("Received payload: :?", s);
s
Some(Err(e)) =>
error!("Error viewing payload contents: ", e);
return;
;
info!("Received message from broker in read-only stream");
if payload.len() > 0
info!("Sending payload :?", payload);
match tx.send(Ok(KafkaResponse
success: true,
optional_content: Some(
bridge::kafka_response::OptionalContent::Content(
(*payload).as_bytes().to_vec(),
),
),
))
Ok(_) => info!("Successfully sent payload to client"),
Err(e) =>
error!("GRPC error sending message to client :?", e);
return;
else
warn!("No content detected in payload from broker");
match consumer.commit_message(&message, CommitMode::Async)
Ok(_) => (),
Err(e) =>
error!("Error commiting a consumed message: :?", e);
return;
);
Ok(Response::new(Box::pin(
tokio_stream::wrappers::UnboundedReceiverStream::new(rx),
)))
【问题讨论】：
【参考方案1】：问题是 Go 客户端的资源泄漏，在某种程度上，Rust 服务器也是如此。
优雅地关闭 Go 客户端流：
对于名为 Consume 的流式 rpc
ctx, cancel := context.WithCancel(context.Background())
stream, err := handler.Client.Consume(ctx, &bridge.ConsumeRequest
Topic: topic,
)
closeCallback := func()
stream.CloseSend()
cancel()
c.Close() // where c := *websocket.Conn
defer closeCallback()
使用可取消的上下文初始化流很重要，这样运行时可以释放与多路复用 rpc 关联的资源。调用cancel() 会关闭客户端的连接（这将在服务器尝试发送新消息时显示为错误）。
在服务器端检测client.CloseSend()。
在 Go 客户端中调用 stream.CloseSend() 不会通知服务器终止连接，它只是表示客户端将停止发送消息。在这种特殊情况下，我在流 rpc 连接中的异步线程中工作，因此为了在客户端与 websocket 断开连接时优雅地关闭连接，检测CloseSend 数据包并关闭@ 非常重要987654328@频道以关闭连接。
请注意，这样做的动机是服务器将继续尝试将数据包发送到“死”接收器，因为否则它会认为连接仍然存在。
以下代码改编自this thread：
pub struct DropReceiver<T>
rx: mpsc::UnboundedReceiver<T>,
impl<T> Stream for DropReceiver<T>
type Item = T;
fn poll_next(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Option<Self::Item>>
Pin::new(&mut self.rx).poll_recv(cx)
// The Drop trait detects the `CloseSend`
impl<T> Drop for DropReceiver<T>
fn drop(&mut self)
info!("Connection has been droped");
self.rx.close(); // Shutdown the receiver
// Within the rpc
let (tx, rx) = mpsc::unbounded_channel();
let rx = DropReceiver rx: rx ;
tokio::spawn(async move /** work */);
Ok(Response::new(Box::pin(rx))) // return this
【讨论】：
以上是关于gRPC 客户端不使用服务器端流的主要内容，如果未能解决你的问题，请参考以下文章