Uploaded image for project: 'TinkerPop'
  1. TinkerPop
  2. TINKERPOP-2169

Responses exceeding maxContentLength cause subsequent queries to hang

    Details

    • Type: Bug
    • Status: Closed
    • Priority: Major
    • Resolution: Fixed
    • Affects Version/s: 3.4.0
    • Fix Version/s: 3.3.6, 3.4.1
    • Component/s: driver
    • Labels:
      None
    • Environment:
      Client: EC2 Amazon Linux t3.medium
      Server: Amazon Neptune r4.2xlarge

      Description

      Gremlin Driver replaces connections on the channel when it receives Exceptions that are instances of IOException or CodecException (including CorruptedFrameException). When CorruptedFrameException is thrown because response length is greater than the maxContentLength value (32kb by default), driver thinks the host might be unavailable and tries to replace Connection.

      If Connection is shared among multiple requests (its pending queue is > 1), other WSConnection goes stale after connection replacement, while keeping the server executor threads busy.

      Keeping the exec threads busy for stale connections prevents server from picking up new tasks for subsequent requests from the request queue. Additionally since there is a new connection added in Client, it can accept more requests and similar errors can lead to a build up in request queue. When many concurrent requests gets into this situation server become unresponsive to the new requests.

      Steps to repro

       

      1. Have a gremlin server

      2. Connect it using java driver with setting the maxContentLength pretty low, i.e. using the config below:

       

      Cluster.Builder builder = Cluster.build();
              builder.addContactPoint(endpoint[0]);
              builder.port(8182);
              builder.maxConnectionPoolSize(100);
              builder.maxSimultaneousUsagePerConnection(100);
              builder.maxInProcessPerConnection(50);
              builder.maxContentLength(32); // <-- this is reduced from 32k 
              builder.keepAliveInterval(0);
      

       

      3. Issue concurrent requests using the cluster, where the response would be greater than 32 bytes.

      Ideas on a possible solution

      One possible solution to this is to not consider channel as dead when request length exceeds maxContentFrame length.
      {{}}

      final class Connection {
          ...
          public ChannelPromise write(final RequestMessage requestMessage, final CompletableFuture<ResultSet> future) {
          ...
      
              // FIX HERE: Do not consider CorruptedFrameException as non-recoverable exception.
              if ((t instanceof IOException || t instanceof CodecException) && (! (t instanceof CorruptedFrameException))) {
              ...
              }
          }
      }
      

      Another fix could be the request can be deleted from the Connections' pending request map, and if there are other pending requests on the connection, close them before replacing the connection, or not replace the connection at all: 

       

      final class Connection {void replaceConnection(final Connection connection) {
          ...
          // FIX HERE: Do not replace connection if there are pending requests on it. 
          if (!connection.getPending().isEmpty()) {
              return; // prevent replacing the connection while there are pending requests.
          }
      
          considerNewConnection();
          definitelyDestroyConnection(connection);
          }
      }

       

       

        Attachments

          Activity

            People

            • Assignee:
              spmallette stephen mallette
              Reporter:
              yigitk Yigit Kiran
            • Votes:
              0 Vote for this issue
              Watchers:
              7 Start watching this issue

              Dates

              • Created:
                Updated:
                Resolved: