On Tue, Feb 18, 2020 at 01:29:24AM +0100, Tomas Mudrunka wrote:
> is there safe way to garbage collect old commits from git repository? Lets
> say that i want to always keep only last 100 commits and throw everything
> older away. To achieve similar goal as git clone --depth=100, but on the
> server side. I had partial success with doing shallow clone and then
> converting to bare repo while removing the shallow flag from .git/config.
> But i didn't liked that solution and wasn't really sure what consequences in
> terms of data integrity and forward compatibility with newer git versions
> might be.
I can't say for sure, but what you did with the shallow file is likely
to bite you later. Shallow repositories are supposed to know where their
boundary cutoffs are, and that information is stored in that file.
The normal answer here is that you'd want to rewrite the history using
grafts and git-filter-branch, or the new git-filter-repo. But...
> To tell you more about my USE CASE:
>
> I want to create free opensource sofware similar to dropbox, but based on
> git. My idea is following:
...I think the rewrite would defeat the purpose, since you're relying on
the stability of the hashes to let all sides of the conversation figure
out when they're in sync.
> I am aware that this is not something which git was designed for, but to me
> it seems like it should be more than doable. Do you think, any of you can
> give me some hints on how to approach this problem please?
I don't know that there's an easy way. Git is close to what you want,
but really is designed to assume the other side has all the reachable
objects. Shallow clones are the feature that's closest to what you
want, but:
- I haven't had good experiences with repeatedly fetching into a
shallow clone. I believe the shallow list can grow because the
client doesn't realize which commits are reachable from others (and
hence are redundant). And I have seen shallow cuts crossing merge
boundaries cause a lot of extra objects to be transferred.
- It's really designed for _some_ repository to have all of the
objects. Fetching out of a shallow repository does work, I think,
but I would guess isn't very well exercised. So I have no idea what
kind of dragons you'd encounter.
Another option would be to periodically rewrite the history to a point
that you think all clients have synced, and then somehow communicate the
rewrite to them (outside of Git, but it sounds like you'd have software
wrapping Git). And then they could all do the identical rewrite and keep
going.
Also look at git-annex if you haven't, which I think supports this kind
of history truncation. I don't remember how it works exactly, but I
think that it may not put the blobs into Git at all, but rather just
pointers. So you're free to remove the actual data, but the pointer
remains (and will later say "sorry, I can't get that data for you").
-Peff
