Web lists-archives.com

Re: [RFE] Inverted sparseness




From: "Randall S. Becker"  :December 03, 2017 11:44 PM
On December 3, 2017 6:14 PM, Philip Oakley wrote a nugget of wisdom:
From: "Randall S. Becker" <rsbecker@xxxxxxxxxxxxx>
Sent: Friday, December 01, 2017 6:31 PM
On December 1, 2017 1:19 PM, Jeff Hostetler wrote:
On 12/1/2017 12:21 PM, Randall S. Becker wrote:
I recently encountered a really strange use-case relating to sparse
clone/fetch that is really backwards from the discussion that has
been going on, and well, I'm a bit embarrassed to bring it up, but I
have no good solution including building a separate data store that
will end up inconsistent with repositories (a bad solution).  The
use-case is as
follows:

Given a backbone of multiple git repositories spread across an
organization with a server farm and upstream vendors.
The vendor delivers code by having the client perform git pull into
a specific branch.
The customer may take the code as is or merge in customizations.
The vendor wants to know exactly what commit of theirs is installed
on each server, in near real time.
The customer is willing to push the commit-ish to the vendor's
upstream repo but does not want, by default, to share the actual
commit contents for security reasons.
Realistically, the vendor needs to know that their own commit id was
put somewhere (process exists to track this, so not part of the
use-case) and whether there is a subsequent commit contributed >by
the customer, but the content is not relevant initially.

After some time, the vendor may request the commit contents from the
customer in order to satisfy support requirements - a.k.a. a defect
was found but has to be resolved.
The customer would then perform a deeper push that looks a lot like
a "slightly" symmetrical operation of a deep fetch following a prior
sparse fetch to supply the vendor with the specific commit(s).

Perhaps I'm not understanding the subtleties of what you're
describing, but could you do this with stock git functionality.

Let the vendor publish a "well known branch" for the client.
Let the client pull that and build.
Let the client create a branch set to the same commit that they fetched.
Let the client push that branch as a client-specific branch to the
vendor to indicate that that is the official release they are based on.

Then the vendor would know the official commit that the client was using.
This is the easy part, and it doesn't require anything sparse to exist.

If the client makes local changes, does the vendor really need the SHA
of those -- without the actual content?
I mean any SHA would do right?  Perhaps let the client create a second
client-specific branch (set to  the same commit as the first) to
indicate they had mods.
Later, when the vendor needs the actual client changes, the client
does a normal push to this 2nd client-specific branch at the vendor.
This would send everything that the client has done to the code since
the official release.

What I should have added to the use-case was that there is a strong
audit requirement (regulatory, actually) involved that the SHA is
exact, immutable, and cannot be substitute or forged (one of the
reasons git is in such high regard). So, no I can't arrange a fake SHA
to represent a SHA to be named later. It SHA of the installed commit
is part of the official record of what happened on the specific server,
so I'm stuck with it.

I'm not sure what you mean about "it is inside a tree".

m---a---b---c---H1
         `---d---H2

d would be at a head. b would be inside. Determining content of c is
problematic if b is sparse, so I'm really unsure that any of this is
possible.

I think I get the jist of your use case. Would I be right that you don't
have a true working
solution yet? i.e. that it's a problem that is almost sorted but falls down
at the last step.

If one pretended that this was a single development shop, and the various
vendors, clients
and customers as being independent devolopers, each of whom is over
protective of their
code, it may give a better view that maps onto classic feature development
diagrams.
(i.e draw the answer for local devs, then mark where the splits happen)

In particular, I think you could use a notional regulator's view that the
whole code base is
part of a large Git heirarchy of branches and merges, and that some of the
feature loops
are only available via the particular developer that worked on that
feature.

This would mean that from a regulatory overview there is a merge commit in
the 'main'
(master) heirachy that has the main and feature commits listed, and the
feature commit
is probably an --allow-empty commit (that has an empty tree if they are
that paranoid) that
says 'function X released' (and probably tagged), and that release commit
then has, as its
parent, the true release commit, with the true code tree. The latter commit
isn't actually being
shown to you!

At this point the potential for using the graft capability comes in (as a
regulated method!).
Locally the graft records the missing line of pearls for that paranoid
dev/vendor/customer/client.
The whole git heirachy still works.

The question is how to get that  release commit with its empty tree, and
its tag, to you from
the dev. I'd guess that a fast-export of just that tag/commit/empty tree
would allow you to
bring in that sentinel point to your heirachy (initially as a
psuedo --root), and then graft it
on. (I haven't checked if fast-export allows such specificity, but it's a
method)

You can now form the merge commit and have regulatory oversight and the
full git validation
and verification capability, as long as your web of trust extends to the
regulator looking
effectively across the air gap. It's a fresh way of seeing the web of
trust.

Thus you/they have various "shallow clones", but with gaps and islands in
the shallowness....
and those gaps are spanned by grafts (which are audited). The `git-replace`
may also be an
option, but I don't think it's quite right for this case. You just have a
temporary gap in
the history, and with fast export

If using the empty tree part doesn't pass muster (i.e. showing nothing
isn't sufficient),
then the narrow clone could come into play to limit what parts of the trees
are widely
visible, but mainly its using the grafts to cover the regulatory gap, and
(for the moment)
using fast-export to transfer the singleton commit / tags

Oh Just remembered, there is the newish capability to fetch random blobs,
so that may help.

Randall said>
I think you hit the nail on the head pretty well. We're currently at 2.3.7,
with a push to 2.15.1 this week, so I'm looking forward to trying this. My
two worries are whether the empty tree is acceptable (it should be to the
client, and might be to the vendor), and doing this reliably
(semi-automated) so the user base does not have to worry about the gory
details of doing this. The unit tests for it are undoubtedly going to give
me headaches.

Thanks for the advice. Islands of shallowness are a really descriptive image
for what this is. So identifying that there are shoals (to extend the
metaphor somewhat), will be crucial to this adventure.
---

An overnight sleep remined me that the ideal way of transferring across the air gap is *Obviously* the use of `git bundle`

Bundle allows you to specify the exact revisions (the tag and commits) in the bundle that is sneakernet transferred (or email) between the repos.

I'm also thinking that the vendor/client/customet should also have, on their side, one of the empty merge commits that shows both the original fork-point (aka merge base) and their current (empty) release commit. This provides the authentication and verification that they have used the right base commit for their ox-bow lake of disconnected development (Oh the metaphors just keep coming). It also provies a place for the automated/scripted graft to get the two ends of the graft from, and check they are valid.

It would be very easy for transcription errors to sneak in at that step of recording the fork-point (which would be a bit of unexpected river capture - https://phys.org/news/2017-04-retreating-yukon-glacier-river.html), so making the client also do it removes that concern.

The creation of client side and your side empty-merges should also create a criss-cross plaiting that locks the two processes together - it's almost a block-chain!

Hope it goes well. It would be great to hear the result.
--
Philip