This means that a server supporting SW might only hear of the tx data and not get the signature data for some transactions, depending on how the relay rules worked (e.g. if the SW peers had higher minrelaytxfee settings than the legacy peers). This would complicate fast block relay code like IBLTs, since we now have to check to see that the recipient has both the tx data and the witness/sig data.
The same issue might happen with block relay if we do SW as a soft fork. A SW node might see a block inv from a legacy node first, and might start downloading the block from that node. This block would then be marked as in-flight, and the witness data might not get downloaded. This shouldn't be too hard to fix by creating an inv for the witness data as a separate object, so that a node could download the block from e.g. Peer 1 and the segwit data from Peer 2.
Of course, the code would be simpler if we did this as a hard fork and we could rely on everyone on the segwit fork supporting the segwit data. Although maybe we want to write the interfaces in a way that supports some nodes not downloading the segwit data anyway, just because not every node will want that data.
I haven't had time to read sipa's code yet. I apologize for talking out of a position of ignorance. For anyone who has, do you feel like sharing how it deals with these network relay issues?
By the way, since this thread is really about SegWit and not about any other mechanism for increasing Bitcoin capacity, perhaps we should rename it accordingly?
A segwit supporting server would be required to support relaying segwit transactions, although a non-segwit server could at least inform a wallet of segwit txns observed, even if it doesn't relay all information necessary to validate.
Non segwit servers and wallets would continue operations as if nothing had occurred.
If this means essentially that a soft fork deployment of SegWit will require SPV wallet servers to change their logic (or risk not being able to send payments) then it does seem to me that a hard fork to deploy this non controversial change is not only cleaner (on the data structure side) but safer in terms of the potential to affect the user experience.
—
Regards,