On 3/31/2023 1:06 PM, Junio C Hamano wrote:

"Derrick Stolee via GitGitGadget" [off-list ref] writes:

quoted

From: Derrick Stolee <redacted>

When fetch.bundleURI is set, 'git fetch' downloads bundles from the
given bundle URI before fetching from the specified remote. However,
when using non-file remotes, 'git fetch --all' will launch 'git fetch'
subprocesses which then read fetch.bundleURI and fetch the bundle list
again. We do not expect the bundle list to have new information during
these multiple runs, so avoid these extra calls by un-setting
fetch.bundleURI in the subprocess arguments.

It is a good observation, if we assume that all these remotes want
to share the same expectation on what the list of bundles described
in bundleURI should be.

I expected that people use multiple remotes and do a "fetch --all"
before starting work from a cronjob or while fetching coffee first
time in the morning to fetch from repositories holding work from
different folks and manged by different groups, and these groups do
not tightly share the object management recipes at what bundles to
pre-package and list in the list served at a bundleURI.  If there is
such an arrangement between repositires to share the object
management, even if the repositories fetched with the "--all" option
are truly multiple places, it may make sense to assume that these
repositories you are fetching from want you to use the same set of
bundles that are managed the same way to be used.  But I am not sure
if that assumption holds true.

Where does fetch.bundleURI come originally?  If we set only one
globally for the local repository at "git clone" time, perhaps that
is a problem?  IOW, instead of fetch.bundleURI, we would want to
have something per remote, e.g. remote.<name>.bundleURI, instead?

fetch.bundleURI is either set by 'git clone --bundle-uri' when the
list advertises a bundle.heuristic value, or can be set manually by
the user.

This indicates the use of a "standalone" bundle server that is
managed independently from a Git remote. The bundle list is downloaded
from a static URI instead of during protocol v2 in the 'bundle-uri'
command.

In that sense, the static bundle URI is not paired with a remote,
but instead presents a place where objects can be downloaded perhaps
faster than any Git remote (by downloading from a static file, or
because the bundle server has a faster link to the client machine).
The objects added to the client object store can then reduce the
size of the fetch from each other Git server remote, of course only
if the objects in the bundle(s) are also present on those remotes.

If there are remote-specific bundle lists, then the protocol v2
command will be a good way to communicate bundle lists that are
custom to those remotes. However, even in that case it is unlikely
that sourcing bundles from multiple sources will be useful, since
the bundle downloads will likely have significant common data.
 

Putting that design level tangent aside, everything I see in this
patch makes sense, assuming that the fetch_bundle_uri() call done
fairly early in the parent process is sufficient to tell the child
processes that deal with individual repositories to reuse the info
that was already retrieved that call.

Thanks for looking closely.

-Stolee