Since this issue is still actively referenced I'd like to shed some explicit details here. I'm using redis as my backend. As stated above, celery is not the issue. I wouldn't advise setting output buffers to 0 either. You can find a very good description for this limitation in redis docs here.

To investigate whether output buffer limits are your issue, run client list on the redis server reference. Trim all the clients where omem=0. The remaining clients will have output buffers > 0. See if any have sub > 0 or psub > 0 and exceed your servers soft/hard limits.

We have a single service that triggers .delay tasks. Delay returns an AsyncResult. This means celery has to subscribe to redis for events related to the task. As referenced above, the pub/sub clients have limits on output buffers. If you only have a few workers, it could take some time for the subscriptions to call back. In our particular case, we just need "fire and forget". So we are making use of the ignore_result feature (docs). If you need the results you'll need to balance your connections with the amount of workers doing the processing.