[rabbitmq-discuss] python client support for channel.flow

[Brian Sullivan]

Hi,

I wanted to follow up on an issue that's been on our radar for a while, but
I haven't seen any updates on it and wanted to see what our options might be
here.

We have been leveraging the py-amqplib client library with rabbitmq, which
has performed quite well in our application with one main limitation.  The
main issue we have with it is that it does not currently support the
Channel.Flow message from an AMQP broker - which means when the server tries
to prevent a producing client app from overloading the system, the client
ignores the message and continues sending.  Since RabbitMQ will continue to
process until it allocates all of its memory, it dies a horrible death and
all of the messages in the broker are lost (reproduction steps below).

Two basic questions:
1) What are my alternatives here from a python client side?  I have
periodically checked into this with Barry for the py-amqplib, but it appears
to involve a pretty significant code restructuring (if I remember correctly,
it had to do with the synchronous processing of messages).  We need a fully
baked client, and Barry's client works well in normal situations - we just
need it to be more fault-tolerant in the heavy saturation cases to prevent
production failures.

2) If there are no viable options from a python client right now, what
prevents you guys from adding some code to the rabbitmq broker from simply
failing a connection that refuses to honor the Channel.Flow message when it
has been broadcast?  I asked this back in 1.5.0 (or around there), and think
this would be a great safety mechanism to prevent a single rogue client from
bringing down a broker.

Any help here would be greatly appreciated.  This has bitten us a few times
in the past and is the weakest part of our pipeline right now.  We are able
to buffer our messages on disk on the producer side, so any stoppage at the
rabbitmq side would simply cause our producer app to retry a connection and
resume transmission.

I don't read this list as actively as I would like, but I didn't find any
recent posts on this from any searches I was able to come up with.
Hopefully the answer is out there and I have just missed it.  Please let me
know if I need to give more info here to help drive the discussion.

Thanks,
Brian

==== Reproduction steps ====

Attached is my failure case test script.  In a nutshell, the test is as
follows:
- Create non-auto-delete queue, bound to a topic
- Spam that topic using the attached script until it maxes out the memory on
the machine (luckily, my test box is only 2GB so this does not take too
long)
- You can see the alerts fire in the /var/log/rabbitmq/rabbit.log file, but
no active flag is set on the Channel object on the Python client side.
- One note is that I am talking to a 1.5.3 broker, but the Channel.Flow
should be working fine in that version, this is the client's handling that I
am focusing on.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.rabbitmq.com/pipermail/rabbitmq-discuss/attachments/20091227/d9ae1653/attachment.htm 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: rmq-spam.py
Type: application/octet-stream
Size: 2525 bytes
Desc: not available
Url : http://lists.rabbitmq.com/pipermail/rabbitmq-discuss/attachments/20091227/d9ae1653/attachment.obj