Re: Identical files across subsequent package revisions

[Julien Lepiller <julien@lepiller.eu>]

Le 23 décembre 2020 09:07:23 GMT-05:00, zimoun <zimon.toutoune@gmail.com> a écrit :
>Hi,
>
>On Wed, 23 Dec 2020 at 14:10, Miguel Ángel Arruga Vivas
><rosen644835@gmail.com> wrote:
>
>> Probably you're already aware of it, but I want to mention that
>> Tridgell's thesis[1] contains a very neat approach to this problem.
>
>This thesis is a must to read! :-)
>
>
>> A naive prototype would be copying of the latest available nar of the
>> package on the client side and using it as the destination for a copy
>> using rsync.  Either the protocol used by the rsync application, or a
>> protocol based on those ideas, could be implemented over the HTTP
>layer;
>> client and server implementation and cooperation would be needed
>> though.
>
>I could misunderstand and miss something, one part of the problem is
>how
>to detect “latest”; other said how to know it is different.  From my
>memories, and I have drunk couple of beers since I read the thesis :-),
>the ’rsync’ approach uses timestamp and size.  And if you switch to
>checksum instead, the performances are poor, because of IO.  Well, it
>depends on the number of files and their size, if this checksum are
>computed ahead, etc.
>
>
>> Another idea that might fit well into that kind of protocol---with
>> harder impact on the design, and probably with a high cost on the
>> runtime---would be the "upgrade" of the deduplication process towards
>a
>> content-based file system as git does[2].  This way the a description
>of
>> the nar contents (size, hash) could trigger the retrieval only of the
>> needed files not found in the current store.
>
>Is it not related to Content-Addressed Store?  i.e, «intensional
>model»?
>
>Chap. 6: <https://edolstra.github.io/pubs/phd-thesis.pdf>
>Nix FRC:
><https://github.com/tweag/rfcs/blob/cas-rfc/rfcs/0062-content-addressed-paths.md>

I think this is different, because we're talking about sub-element content-addressing. The intensional model is about content-addressing whole store elements. I think the idea would be to save individual files in, say, /gnu/store/.links, and let nar or narinfo files describe the files to retrieve. If we are missing some, we'd download them, then create hardlinks. This could even help our deduplication I think :)

>
>
>Cheers,
>simon