On 21 Nov 2013, at 4:50 PM, Plüm, Rüdiger, Vodafone Group <ruediger.pl...@vodafone.com> wrote:
> If this filter sets aside all stuff until eos (or a flush bucket) is seen it > might > cause a huge amount of memory consumption. So we possibly need some limit for > memory consuming > buckets which when passed causes us to send the stuff we have down the chain > even without an eos > or flush bucket. As far as I know the current core output filter also only > buffers a limited > amount of "memory" buckets until it decides to go for blocking write. In theory it would be ideal if flow control could be handled at the start of the protocol stack rather than at the end. In other words, the protocol used for flow control in core be moved into a dedicated filter. We will then be able to pass file buckets through without a problem, and properly handle morphing and memory buckets up to the sensible limits we have now. Life then becomes simple, memory constrained (and hopefully exclusively async) downstream from this filter. Regards, Graham --
