为什么 ZeroMQ PUSH/PULL 可以工作，而不是 PUB/SUB？

Question

Environment: NVIDIA-flavored Ubuntu 18.01 on their Jetson development board with their TX2i processor. ZMQ 4.3.2, utilizing the cppzmq C++ wrapper for ZMQ.

I've got a slew of code running using google protocol buffers with ZeroMQ, and it's all PUSH/PULL, and it works fine except I've got one case that isn't point-to-point, but 1:3. The correct solution here is to do PUB/SUB, but I cannot get messages through to my subscriber.

I shaved my code down to this simple example. If I uncomment the #define statements, the subscriber gets nothing. Commented (which compiles as PUSH/PULL instead of PUB/SUB), then the subscriber gets the message as expected. With the excessive sleep_for() times, I would expect the subscriber has ample time to be registered before the publisher performs the send.

EDIT:

Why the try/catch on the subscriber? I was getting an exception early on, and believed it was because the publisher wasn't ready. This no longer appears to be the case, so it wasn't what I thought it was.

// Publisher
#include "/usr/local/include/zmq.hpp"
#include "protobuf_namespace.pb.h"
#include <chrono>
#include <thread>


#define PUB_SUB

int main( void )
{
  zmq::context_t* m_pContext = new zmq::context_t( 1 );

#ifdef PUB_SUB
  zmq::socket_t*  m_pSocket  = new zmq::socket_t( *m_pContext, ZMQ_PUB );
#else
  zmq::socket_t*  m_pSocket  = new zmq::socket_t( *m_pContext, ZMQ_PUSH );
#endif

  std::this_thread::sleep_for( std::chrono::seconds( 1 ) );
  //m_pSocket->bind( "tcp://*:53001" );       // using '*' or specific IP doesn't change result
  m_pSocket->bind( "tcp://127.0.0.1:53001" );
  std::this_thread::sleep_for( std::chrono::seconds( 1 ) );

  // Send the parameters
  protobuf_namespace::Params params;
  params.set_calibrationdata( protobuf_namespace::CalDataType::CAL_REQUESTED ); // init one value to non-zero
  std::string        params_str = params.SerializeAsString();
  zmq::message_t     zmsg( params_str.size() );

  memcpy( zmsg.data(), params_str.c_str(), params_str.size() );
  m_pSocket->send( zmsg, zmq::send_flags::none );

  std::this_thread::sleep_for( std::chrono::seconds( 1 ) );
  m_pSocket->close();
  zmq_ctx_destroy( m_pContext );
}

// Subscriber - start me first!
#include "/usr/local/include/zmq.hpp"
#include "protobuf_namespace.pb.h"
#include <chrono>
#include <thread>
#include <stdio.h>

#define PUB_SUB


int main( void )
{
  zmq::context_t* m_pContext = new zmq::context_t( 1 );

#ifdef PUB_SUB
  zmq::socket_t*  m_pSocket  = new zmq::socket_t( *m_pContext, ZMQ_SUB );
  m_pSocket->connect( "tcp://127.0.0.1:53001" );

  int linger = 0;
  zmq_setsockopt( m_pSocket, ZMQ_LINGER, &linger, sizeof( linger ) );
  zmq_setsockopt( m_pSocket, ZMQ_SUBSCRIBE, "", 0 );
#else
  zmq::socket_t*  m_pSocket  = new zmq::socket_t( *m_pContext, ZMQ_PULL );
  m_pSocket->connect( "tcp://127.0.0.1:53001" );
#endif

  protobuf_namespace::Params params;
  zmq::message_t zmsg;
  bool retry = true;

  do {
    try {
      m_pSocket->recv( zmsg, zmq::recv_flags::none );
      retry = false;
      std::this_thread::sleep_for( std::chrono::seconds( 1 ) );
    } catch( ... ) { 
      printf("caught\n");
    }
    std::this_thread::sleep_for( std::chrono::seconds( 1 ) );
  } while( retry );

  std::string param_str( static_cast<char*>( zmsg.data() ), zmsg.size() );
  params.ParseFromString( param_str );

  if( params.calibrationdata() == protobuf_namespace::CalDataType::CAL_REQUESTED )
    printf( "CAL_REQUESTED\n" );
  else
    printf( "bad data\n" );


  std::this_thread::sleep_for( std::chrono::seconds( 1 ) );
  m_pSocket->close();
  zmq_ctx_destroy( m_pContext );
}

Answer 1

_{In case one has never worked with ZeroMQ,
one may here enjoy to first look at " ZeroMQ Principles in less than Five Seconds "
before diving into further details}

Q : Why the try/catch on the subscriber?

Because :
a ) // Subscriber - start me first! and yet having the PUB -side sleeping almost "forever" before doing the tcp:// transport-class path setup to accept any first .connect() in .bind() , here preceded by a massive nap ...
std::this_thread::sleep_for( std::chrono::seconds( 1 ) ); and

b ) the try 'd m_pSocket->recv( zmsg, zmq::recv_flags::none ); must by definition throw an exception, as there is no tcp:// transport-class path setup so far ( as the PUB -side has not returned yet from the sleep )

Q : Why will ZeroMQ PUSH/PULL work but not PUB/SUB?

Well, both will , if designed properly, respecting the published API.

Just remove any blocking sleep() -s, preventing the SUB -s to join in, making the .connect() -s able to success ASAP. Plus perhaps move into a non*blocking form of the .recv() -ops ( refactor try/catch ), as is common in distributed-computing to better reflect the nature of preventively designed .poll() -based or reactive .recv(..., ZMQ_NOBLOCK ) -based event-handling.

---

Last, but not least :

ZeroMQ v4+ ( as opposed to v2+ and pre-v3.? API ) was switched to use the PUB -side message filtering, so a due consideration of subscription management (timing/error-handling/resilience) has to take place either.

In any case of doubt, may integrate a use of ZeroMQ built-in socket_monitor tools, extending the Context() -instance, and trace / inspect each and every event, internal to the Context() -instance, well under the published API-events, down to the lowest Level-of-Detail.

Do not hesitate to read and ask more