On --depth=funny value
From: Junio C Hamano <hidden>
Date: 2016-06-15 22:55:42
Here to outline my current thinking. Note that this is unrelated to
the "git clone --bottom=v1.2.3" to say "I do not care about anything
that happened before that version".
* First, let's *not* do "git fetch --depth=inf"; if you want to
unplug the bottom of your shallow clone, be more explicit and
introduce a new option, e.g. "git fetch --unshallow", or
something.
* Make "git fetch" and "git clone" die() when zero or negative
number is given with --depth=$N, for the following reasons:
- The --depth option is describe as:
("git clone") ... a 'shallow' clone with a history
truncated to the specified number of revisions.
("git fetch") Limit fetching to ancestor-chains not longer
than n.
It is fairly clear from the above that negative $N does not
make any sense.
Technically, fetching the commits that were explicitly asked
for and not there parents is the only possible ancestor-chain
that is not longer than -4, so "git fetch --depth=-4" ought to
behave just like "git fetch --depth=0", but you have to be very
sick to read the documentation and expect things to work that
way. Also there is no way to misread the "git clone"
documentation and expect "git clone --depth=-4" to create a
history truncated to negative number of revisions.
Which means that it is the right thing to do to die() when a
negative number is given to --depth for these commands.
- As people count from one, the natural way to ask for the tip
commit without any history ought to be "--depth=1". Let's
declare the current behaviour of "--depth=1" that gives the tip
and one commit behind it a bug.
Which means that these commands should be updated to die() when
zero is given to their --depth option. "Do not give me any
history" is inherenty incompatibe with "clone" or "fetch".
Because there is no configuration variable "fetch.depth" (or
"clone.depth") that forces all your cloned repositories to be
shallow, "git clone --depth=0" or "git fetch --depth=0"
couldn't have been ways for existing users to ask to defeat any
funny configured depth value and clone or fetch everything.
When they wanted to clone or fetch everything, they would have
just used the command without any "--depth" option instead.
Which means that nobody gets hurt if we change these commands
to die() when zero is given to their --depth option.
* We would like to update "clone --depth=1" to end up with a tip
only repository, but let's not to touch "git fetch" (and "git
clone") and make them send 0 over the wire when the command line
tells them to use "--depth=1" (i.e. let's not do the "off-by-one"
thing).
Instead, fix "upload-pack" (the bug is in get_shallow_commits()
in shallow.c, I think), so that it counts correctly. When the
other end asks for a history with 1-commit deep, it should return
a history that is 1-commit deep, and tell the other end about the
parents of the returned history, instead of returning a history
that is 2 commmits deep. So when talking with a newer server,
clients will get correct number of commits; when talkng with an
older server, clients will get a bit more than they asked, but
nothing will break.
Can people sanity check the reasoning outlined here? Anything I
missed?
The above outline identifies three concrete tasks that different
people can tackle more or less independently, each with updated
code, documentation and test:
1. "git fetch --unshallow" that gives a pretty surface on Duy's
"--depth=inf";
2. Making "git fetch" and "git clone" die on "--depth=0" or
"--depth=-4";
3 Updating "upload-pack" to count correctly.
I'll refrain from saying "Any takers?" for now.