Rewrite Git history to prune large 'old files'?

[pwaller]

Update summary, 23/11/2018: This repository currently requires ~10MiB of download, which isn't ideal considering the source is only a few hundreds of kilobytes. @mewmew and I propose to shrink it to ~800kiB, to give a faster "Go install" experience for anyone using the repository.

The reason for the blowup is that there were some large test cases (including sqlite) which measure in the 10's of MiBs, and various other bits relating to parsing were also quite large. Those have now moved into other repositories in the llir organization, so we don't need to download those anymore if you just want to import llir.

---

Original issue text.

I just saw @mewmew's comment in ec48d54 but thought it would be easier to have a separate issue for discussion - the commit itself is very long so if I commented on the commit the discussion would be way down at the bottom!

First, can I clarify the question - are you asking how to remove lots of old large assets from the history of the repository?

If that is the question, the answer is, yes you can do it, but anyone who cloned the repository needs to know about it otherwise they might get in a mess, since it requires rewriting history. At least, that's the best I know. See github's guidance on the issue.

[mewmew]

Thanks for creating the issue, I agree, it's easier to keep the discussion here.

First, can I clarify the question - are you asking how to remove lots of old large assets from the history of the repository?

Exactly!

On the other hand, I did a git clone just now to check the size of the repo, and it wasn't as bad as I had thought. Perhaps we don't need to do this after all.

[u@x1 ~]$ time git clone https://github.com/llir/llvm
Cloning into 'llvm'...
remote: Enumerating objects: 20, done.
remote: Counting objects: 100% (20/20), done.
remote: Compressing objects: 100% (14/14), done.
remote: Total 10760 (delta 10), reused 15 (delta 6), pack-reused 10740
Receiving objects: 100% (10760/10760), 9.03 MiB | 1.47 MiB/s, done.
Resolving deltas: 100% (6437/6437), done.

real	0m7.715s
user	0m1.810s
sys	0m0.521s

[u@x1 ~]$ du -hs llvm
11M	llvm

I think we could prune the repo down to 1 MB or so instead of 11 MB, but the question is if it's worth it, given that it requires a force push.

However, should we decide to do this, then having it ready before the v0.3.0 release seems like a perfect time.

[pwaller]

I like the idea in principle, but I think this repository has enough followers that it may cause harm to drop the old content.

In practice, the repository is quite small even at 11MiB.

If you decided to do it, you could keep a copy of the old repository around at llir/llvm-legacy, analogously to https://github.com/go-gl-legacy/gl - and that way if anyone does need the old content (e.g, they were depending on a specific git hash) then at least it still exists for the purposes of figuring out what the git hash is in the new repository.

Maybe it's possible to keep the old history around in a separate git ref which doesn't get cloned by default. But in that case I guess the content would be harder to discover.

[mewmew]

I like the idea in principle, but I think this repository has enough followers that it may cause harm to drop the old content.

In practice, the repository is quite small even at 11MiB.

Agreed. Had the repo been at 100 MB, then we probably would have done it, but at this size it does not seem worth the potential harm to users. (The idea with shrinking the repo was of course to make it easier for users to make the initial download, especially those who happen to be on a slow Internet connection, as may be common in parts of Asia, etc).

So, for now. I'm fine with keeping it as it is, and just being careful when adding large content in the future. Closing this issue for now. We can always refer back and re-open at a later point.

[mewmew]

Maybe it's possible to keep the old history around in a separate git ref which doesn't get cloned by default. But in that case I guess the content would be harder to discover.

Also, if Go ever does shallow Git clone, this issue would be resolved I think. (upstream issue golang/go#13078)

[pwaller]

I just learned that Go did this to their repository recently, the discussion in there and how they went about it is pretty interesting:

golang/go#28899

I think it probably doesn't change anything with respect to what we might do to this repository.

[mewmew]

I just learned that Go did this to their repository recently, the discussion in there and how they went about it is pretty interesting:

Thanks for the link! It was an interesting read to see how they resolved it.

I think it probably doesn't change anything with respect to what we might do to this repository.

Most likely not. If we end up doing a pruning, then I'd suggest we use bfg as suggested on the GitHub link you posted. Also, if we do this, then perhaps in the next few weeks, as the intention is to have v0.3.0 released some time in early December.

I'm kind of still a bit on the fence. I don't think we need the rewrite. However, should we ever do one, now is basically the perfect time to. As we move from v0.2 to v0.3, since users will have to do manual changes to get the latest release anyways (updating to the latest API, etc).

[mewmew]

Until we decide for sure. I'll re-open the issue. Also, this may help get input from other users of the repo who it may affect. I'll also re-name the title to include a mention of Git history rewrite.

[pwaller]

Some large paths:

 git rev-list --objects --all | git cat-file --batch-check='%(objectsize:disk) %(objectname) %(objecttype) %(rest)' | grep ' blob ' | awk '{print $4" "$1}' | awk '{
    arr[$1]+=$2
   }
   END {
     for (key in arr) printf("%s\t%s\n", arr[key], key)
   }' | sort -nr | awk '{print $2"\t"$1}' | column -t -s$'\t' | head
old/asm/internal/testdata/sqlite/sqlite3.ll                                 3404085
old/asm/internal/testdata/sqlite/sqlite3.c                                  1726782
asm/internal/parser/actiontable.go                                          472010
old/asm/internal/parser/actiontable.go                                      246534
asm/internal/parser/gototable.go                                            149186
old/asm/internal/parser/gototable.go                                        67735
asm/testdata/DebugInfo/COFF/big-type.ll                                     61648
asm/internal/ll.bnf                                                         59883
asm/ll/ll.tm                                                                59610
asm/testdata/c4.ll                                                          55555

This graph shows how much space will be saved, assuming you eliminate large file paths:

[sbinet]

@pwaller why matplotlib? gonum/plot is so much better :P

[pwaller]

kill_ids.csv

[mewmew]

The current intention is to clone llir/llvm into llir/llvm-legacy, to preserve the complete history. Then, to start clean, we will keep any fine currently in HEAD, and it's entire history at that path. Since we need to do a force push anyway, this seem to be the time to really get the size of the repo down.

If anyone currently using the repo has some input or feedback, feel welcome to contribute your thoughts.