Why reproducibility is breaking by metadata?

Why reproducibility is breaking by metadata?

[znavko]

Hello, Guix Help! I am translating Guix manual and found that the author is entirely given to reproducibility. It leads to such phrases that metadata breaks reproducibility when he describes '--save-provenance' flag of 'guix pack' command here:

'This option is not enabled by default because, like timestamps, provenance information contributes nothing to the build process. In other words, there is an infinity of channel URLs and commit IDs that can lead to the same pack. Recording such “silent” metadata in the output thus potentially breaks the source-to-binary bitwise reproducibility property. '

I did not expected such a categorical statement. I think, it does not actually break reproducibility but only complicates checks. If we have to talk about reproducibility to ignoramus, saying 'this option breaks reproducibility option' have to have remark 'simply put' or 'plainly'.

Re: Why reproducibility is breaking by metadata?

[Julien Lepiller]

Le 3 juillet 2019 06:05:02 GMT+02:00, znavko@disroot.org a écrit :
>Hello, Guix Help! I am translating Guix manual and found that the
>author is entirely given to reproducibility. It leads to such phrases
>that metadata breaks reproducibility when he describes
>'--save-provenance' flag of 'guix pack' command here:
>
>'This option is not enabled by default because, like timestamps,
>provenance information contributes nothing to the build process. In
>other words, there is an infinity of channel URLs and commit IDs that
>can lead to the same pack. Recording such “silent” metadata in the
>output thus potentially breaks the source-to-binary bitwise
>reproducibility property. '
>
>I did not expected such a categorical statement. I think, it does not
>actually break reproducibility but only complicates checks. If we have
>to talk about reproducibility to ignoramus, saying 'this option breaks
>reproducibility option' have to have remark 'simply put' or 'plainly'.

I think the key here is "bitwise": it breaks reproducibility in terms of the build results not having the same hash. Does it make sense?

Re: Why reproducibility is breaking by metadata?

[Konrad Hinsen]

znavko@disroot.org writes:

> Hello, Guix Help! I am translating Guix manual and found that the
> author is entirely given to reproducibility. It leads to such phrases
> that metadata breaks reproducibility when he describes
> '--save-provenance' flag of 'guix pack' command here:
>
> 'This option is not enabled by default because, like timestamps,
> provenance information contributes nothing to the build process. In
> other words, there is an infinity of channel URLs and commit IDs that
> can lead to the same pack. Recording such “silent” metadata in the
> output thus potentially breaks the source-to-binary bitwise
> reproducibility property. '
>
> I did not expected such a categorical statement. I think, it does not
> actually break reproducibility but only complicates checks. If we have
> to talk about reproducibility to ignoramus, saying 'this option breaks
> reproducibility option' have to have remark 'simply put' or 'plainly'.

If you define reproducibility in a pragmatic way as something that can
be verified without deep knowledge of file formats, then the only
reasonable definition is bitwise identity. That's what anyone can check
with generic tools. If you want to check that two outputs of 'guix pack'
are equivalent, i.e. identical up to provenance data, then you need a
separate tool for each output format (and you have to write those tools
yourself because to the best of my knowledge they don't exist yet). In
short, I agree with the statements made in the text.

However, you are completely right that this paragraph is not written at
the right level for the typical user. Here is a proposition for
improving it:

--save-provenance

    Save provenance information for the packages passed on the command
    line. Provenance information includes the URL and commit of the
    channels in use (see Channels), which permit the recipient to locate
    the source code of the package definitions that were used.

    Provenance information is saved in the /gnu/store/…-profile/manifest
    file in the pack, along with the usual package metadata—the name and
    version of each package, their propagated inputs, and so on.

    Note that inclusion of provenance information makes the resulting
    archive non-reproducible in the sense that two archives containing
    identical binaries can be different if they were built using
    different channels or different commits.

Does this look better to you?

Philosophical side note: the right way to store provenance information
is outside of the data they refer to. Unfortunately, with file-based
storage, there is no clean way to attach the provenance information
securely to the data without putting it into the same file.

Konrad.

Re: Why reproducibility is breaking by metadata?

[Giovanni Biscuolo]

[-- Attachment #1: Type: text/plain, Size: 910 bytes --]

Hello Konrad,

Konrad Hinsen <konrad.hinsen@fastmail.net> writes:

[...]

> Philosophical side note: the right way to store provenance information
> is outside of the data they refer to. Unfortunately, with file-based
> storage, there is no clean way to attach the provenance information
> securely to the data without putting it into the same file.

Just "food for thought": git-annex metadata allows users to attach
arbitrary metadata to their annexed files, «metadata is stored in the
git-annex branch, and so is automatically kept in sync with the rest of
git-annex's state»

Being content-hashed, each file will persintently keep its metadata even
in case of renames (in the git-annex "filesystem" scope)

I doubt this could be applied to `guix pack`, but it's an intereting
approach to "sidecar" metadata

Happy Guix! Gio'.

-- 
Giovanni Biscuolo

Xelera IT Infrastructures

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]

Re: Why reproducibility is breaking by metadata?

[Konrad Hinsen]

Hi Giovanni,

> Just "food for thought": git-annex metadata allows users to attach
> arbitrary metadata to their annexed files, «metadata is stored in the
> git-annex branch, and so is automatically kept in sync with the rest of
> git-annex's state»

That's indeed a nice example of how to handle metadata in a better way.
But it does require more than a plain files, so publishing and archiving
the metadata requires additional tools (git in this case).

> I doubt this could be applied to `guix pack`, but it's an intereting
> approach to "sidecar" metadata

We could adapt the idea to the Guix store. Write the metadata to a
separate file in the store that has the same filename as the packed
archive, but an additional extension. Since the filename contains the
hash of the archive contents, the relation between data and metadata
is also maintained.

Konrad.