Erlang/OTP Forums

From: "Edwin Fine" <emofine@gmail.com (emofine@gmail.com)>
Date: 8 May 2008 22:12:07 BST
To: "Ben Hood" <0x6e6562@gmail.com (0x6e6562@gmail.com)>
Subject: More RabbitMQ Erlang client woes



Hi Ben,

Sorry for the long email. Trying to get to the bottom of the problems. I have done a LOT more investigation into the Erlang network client, and I believe there are two related bugs, one in the client, and the other I don't know where, or perhaps I am abusing how AMQP is supposed to work.

Bug #1: rabbit_writer not shut down

The first (and easier) bug, which I have fixed eventually, is that every time you start a channel, the network client creates a rabbit_writer, but never shuts it down. Therefore, each channel that is created leaves behind a rabbit_writer process.
What happens is that in the handshake, the network driver starts a rabbit_writer on channel 0. I believe this is used for the connection itself. Its pid is stored in writer_pid in the connection_state. This writer gets cleaned up properly when the connection is shut down. There is no problem with this.
Thereafter, when a channel is opened and amqp_network_driver:open_channel is called, another writer is started (correctly - need one per connection, right?). There is a singleton reader. Anyway, this writer is never terminated. The writer is registered as a peer to the channel, using amqp_channel:register_direct_peer. This causes a bug, because the registered peer is never shut down, probably because the direct peer never should be shut down... but this is the NETWORK peer.

So what I did (and you may have a better way) is to add another function, amqp_channel:register_network_peer. This sets a "direct" flag in the channel_state (which I had to add) to false. Calling register_direct_peer sets the flag to true. When amqp_channel:channel_cleanup() is eventually called (and I had to do something for this, too), it checks to see if direct is false. If so, it stops the writer in writer_pid, otherwise it leaves it alone.

I also had to add a call to channel_cleanup in the close_ok, because the cleanup was never getting called.:

amqp_channel:handle_method(ChannelCloseOk = #'channel.close_ok'{}, State) ->