bug#42162: Recovering source tarballs

[zimoun <zimon.toutoune@gmail.com>]

Hi Chris,

On Mon, 13 Jul 2020 at 20:20, Christopher Baines <mail@cbaines.net> wrote:

> Going forward, being methodical as a project about storing the tarballs
> and source material for the packages is probalby the way to ensure it's
> available for the future. I'm not sure the data storage cost is
> significant, the cost of doing this is probably in working out what to
> store, doing so in a redundant manor, and making the data available.

A really rough estimate is 120KB on average* per raw tarball.  So if we
consider 14000 packages and 70% of them are url-fetch, then it leads to
14k*0.7*120K= 1.2GB; which is not significant.  Moreover, if we
extrapolate the numbers, between v1.0.0 and now it is 23 commits per day
modifying gnu/packages/ so 0.7*23*120K*365= 700MB per year.  However,
the 120KB of metadata to re-assemble the tarball have to be compared to
the 712KB of raw compressed tarball; both about the hello package.

*based on the hello package.  And it depends on the number of files in
 the tarball.  File stored not compressed: plain sexp.


Therefore, in addition to what to store, redundancy and availability,
one question is how to store?  Git-repo? SQL database? etc.



> The Guix Data Service knows about fixed output derivations, so it might
> be possible to backfill such a store by just attempting to build those
> derivations. It might also be possible to use the Guix Data Service to
> work out what's available, and what tarballs are missing.

Missing from where?  The substitutes farm or SWH?


Cheers,
simon