Request Ready Listing


Resolution

The cluster node maintains a ready record which maps a key and
a callback perform. The bottom line is chosen relying on the precise
standards to invoke the callback.
For instance if it must be invoked each time
a message from different cluster node is acquired, it may be the
Correlation Identifer of the message.
Within the case of Replicated Log it’s the
Excessive-Water Mark. The callback handles
the response and decides if the consumer request could be fulfilled.

Think about the instance of a key-value retailer the place, knowledge is replicated
on a number of servers. Right here, Quorum can be utilized
to resolve when a replication could be thought of profitable
to provoke a response to the consumer.
The cluster node then tracks the requests despatched to different cluster nodes,
and a callback is registered with every request.
Every request is marked with a Correlation Identifer,
which is used to map response to the request.
The ready record is then notified to invoke the callback
when the response from different cluster nodes are acquired.

For the sake of this instance, let’s name our three cluster nodes
athens, byzantium and cyrene.
The consumer connects with athens to retailer “title” as “Microservices”.
Athens replicates it on byzantium and cyrene; so it sends
a request to itself to retailer the key-value and sends
requests to each byzantium and cyrene concurrently.
To trace responses, athens creates a WriteQuorumResponseCallback
and provides it to the ready record for every of the requests despatched.

For each response acquired, the WriteQuorumResponseCallback is
invoked to deal with the response. It checks whether or not the required quantity
of responses have been acquired.
As soon as the response is acquired from byzantium, the quorum is reached
and the pending consumer request is accomplished.
Cyrene can reply later, however the response could be despatched to the consumer
with out ready for it.

The code seems to be just like the pattern beneath:
Notice that each cluster node maintains its personal occasion of a ready record.
The ready record tracks the important thing and related callback and
shops the timestamp at which the callback was registered.
The timestamp is used to verify whether or not the callbacks should be expired
if responses have not been acquired throughout the anticipated time.

public class RequestWaitingList<Key, Response> {
    non-public Map<Key, CallbackDetails> pendingRequests = new ConcurrentHashMap<>();
    public void add(Key key, RequestCallback<Response> callback) {
        pendingRequests.put(key, new CallbackDetails(callback, clock.nanoTime()));
    }
class CallbackDetails {
    RequestCallback requestCallback;
    lengthy createTime;

    public CallbackDetails(RequestCallback requestCallback, lengthy createTime) {
        this.requestCallback = requestCallback;
        this.createTime = createTime;
    }

    public RequestCallback getRequestCallback() {
        return requestCallback;
    }

    public lengthy elapsedTime(lengthy now) {
        return now - createTime;
    }
}
public interface RequestCallback<T> {
    void onResponse(T r);
    void onError(Throwable e);
}

It’s requested to deal with the response or error
as soon as the response has been acquired from the opposite cluster node.

class RequestWaitingList…

  public void handleResponse(Key key, Response response) {
      if (!pendingRequests.containsKey(key)) {
          return;
      }
      CallbackDetails callbackDetails = pendingRequests.take away(key);
      callbackDetails.getRequestCallback().onResponse(response);

  }

class RequestWaitingList…

  public void handleError(int requestId, Throwable e) {
      CallbackDetails callbackDetails = pendingRequests.take away(requestId);
      callbackDetails.getRequestCallback().onError(e);
  }

The ready record can then be used to deal with quorum responses
with the implementation wanting one thing like this:

static class WriteQuorumCallback implements RequestCallback<RequestOrResponse> {
    non-public closing int quorum;
    non-public unstable int expectedNumberOfResponses;
    non-public unstable int receivedResponses;
    non-public unstable int receivedErrors;
    non-public unstable boolean completed;

    non-public closing RequestOrResponse request;
    non-public closing ClientConnection clientConnection;

    public WriteQuorumCallback(int totalExpectedResponses, RequestOrResponse clientRequest, ClientConnection clientConnection) {
        this.expectedNumberOfResponses = totalExpectedResponses;
        this.quorum = expectedNumberOfResponses / 2 + 1;
        this.request = clientRequest;
        this.clientConnection = clientConnection;
    }

    @Override
    public void onResponse(RequestOrResponse response) {
        receivedResponses++;
        if (receivedResponses == quorum && !completed) {
            respondToClient("Success");
            completed = true;
        }
    }

    @Override
    public void onError(Throwable t) {
        receivedErrors++;
        if (receivedErrors == quorum && !completed) {
            respondToClient("Error");
            completed = true;
        }
    }


    non-public void respondToClient(String response) {
        clientConnection.write(new RequestOrResponse(RequestId.SetValueResponse.getId(), response.getBytes(), request.getCorrelationId()));
    }
}

Every time a cluster node sends requests to different nodes,
it provides a callback to the ready record mapping with the Correlation Identifer
of the request despatched.

class ClusterNode…

  non-public void handleSetValueClientRequestRequiringQuorum(Listing<InetAddressAndPort> replicas, RequestOrResponse request, ClientConnection clientConnection) {
      int totalExpectedResponses = replicas.dimension();
      RequestCallback requestCallback = new WriteQuorumCallback(totalExpectedResponses, request, clientConnection);
      for (InetAddressAndPort reproduction : replicas) {
          int correlationId = nextRequestId();
          requestWaitingList.add(correlationId, requestCallback);
          strive {
              SocketClient consumer = new SocketClient(reproduction);
              consumer.sendOneway(new RequestOrResponse(RequestId.SetValueRequest.getId(), request.getMessageBodyJson(), correlationId, listenAddress));
          } catch (IOException e) {
              requestWaitingList.handleError(correlationId, e);
          }
      }
  }

As soon as the response is acquired, the ready record is requested to deal with it:

class ClusterNode…

  non-public void handleSetValueResponse(RequestOrResponse response) {
      requestWaitingList.handleResponse(response.getCorrelationId(), response);
  }

The ready record will then invoke the related WriteQuorumCallback.
The WriteQuorumCallback occasion verifies if
the quorum responses have been acquired and invokes the callback
to reply to the consumer.

Expiring Lengthy Pending Requests

Typically, responses from the opposite cluster nodes are
delayed. In these cases the ready record typically has
a mechanism to run out requests after a timeout:

class RequestWaitingList…

  non-public SystemClock clock;
  non-public ScheduledExecutorService executor = Executors.newSingleThreadScheduledExecutor();
  non-public lengthy expirationIntervalMillis = 2000;
  public RequestWaitingList(SystemClock clock) {
      this.clock = clock;
      executor.scheduleWithFixedDelay(this::expire, expirationIntervalMillis, expirationIntervalMillis, MILLISECONDS);
  }

  non-public void expire() {
      lengthy now = clock.nanoTime();
      Listing<Key> expiredRequestKeys = getExpiredRequestKeys(now);
      expiredRequestKeys.stream().forEach(expiredRequestKey -> {
          CallbackDetails request = pendingRequests.take away(expiredRequestKey);
          request.requestCallback.onError(new TimeoutException("Request expired"));
      });
  }

  non-public Listing<Key> getExpiredRequestKeys(lengthy now) {
      return pendingRequests.entrySet().stream().filter(entry -> entry.getValue().elapsedTime(now) > expirationIntervalMillis).map(e -> e.getKey()).accumulate(Collectors.toList());
  }



Source_link

Leave a Reply

Your email address will not be published.