> > To resolve this, prefetch all missing REF_DELTA bases before attempting > > to resolve them. This both ensures that all bases are attempted to be > > fetched, and ensures that we make only one request per index-pack > > invocation, and not one request per missing object. > > Hmm. I wonder whether this can lead to *really* undesirable behavior, e.g. > with deep delta chains. The client would possibly have to fetch the > REF_DELTA object, but that would also be delivered in a thin pack with > *another* REF_DELTA object, and the same over and over again, with plenty > of round trips that kill performance really well. When the client fetches the REF_DELTA base, it won't be a REF_DELTA object itself because Git makes these fetches without any "have" lines, so the server doesn't know anything to delta against. Admittedly, this is just due how to we implemented it - if later we find a way to optimize the lazy fetches by adding "have", then we'll have to revisit this. Quoting from the commit message: > > (When fetching REF_DELTA bases, it is unlikely that > > those are REF_DELTA themselves, because we do not send "have" when > > making such fetches.) I tried to address this point with this sentence in the commit message. If you think that this should be addressed more clearly in the commit message, let me know if you have any suggestions. > Wouldn't it make more sense to introduce a new term like `promised` > (instead of `have`)? Both client and server will have to know about this, > and it would be a new capability, of course, but that way the server could > know that it has to send the entire delta chain. > > Of course, this would be quite a bit more involved than the current patch > :-( I think this can also be solved by omitting "thin-pack". We might want to do this once we optimize the lazy fetches by adding "have". Thanks for taking a look at this.
