From: Nicolas Pitre <nico@fluxnic.net>
Date: Mon, 09 Sep 2013 15:01:10 GMT
Subject: Re: [PATCH v2 15/16] index-pack: use nr_objects_final as sha1_table size
Message-ID: <alpine.LFD.2.03.1309091047510.20709@syhkavp.arg>
In-Reply-To: <1378735087-4813-16-git-send-email-pclouds@gmail.com>

On Mon, 9 Sep 2013, Nguyễn Thái Ngọc Duy wrote:

> nr_objects in the next patch is used to reflect the number of actual
> objects in the stream, which may be smaller than the number recorded
> in pack header.

This highlights an issue that has been nagging me for a while.

We decided to send the final number of objects in the thin pack header 
for two reasons:

1) it allows to properly size the SHA1 table upfront which already 
   contains entries for the omitted objects;

2) the whole pack doesn't have to be re-summed again after being 
   completed on the receiving end since we don't alter the header.

However this means that the progress meter will now be wrong and that's 
terrible !  Users *will* complain that the meter doesn't reach 100% and 
they'll protest for being denied the remaining objects during the 
transfer !

Joking aside, we should think about doing something about it.  I was 
wondering if some kind of prefix to the pack stream could be inserted 
onto the wire when sending a pack v4.  Something like:

'T', 'H', 'I', 'N', <actual_number_of_sent_objects_in_network_order>

This 8-byte prefix would simply be discarded by index-pack after being 
parsed.

What do you think?


Nicolas