On 3/26/2018 8:42 AM, Michael Haggerty wrote:
[...]
But there might be some tools out in the wild that have been writing
broken headers. In that case, users who upgrade Git might suddenly
find that they can't read repositories that they could read before. In
fact, a tool that we wrote and use internally at GitHub was doing
exactly that, which is how we discovered this "problem".
This patch shows what it would look like to relax the parsing again,
albeit *only* for the first line of the file, and *only* for lines
that start with '#'.
The problem with this patch is that it would make it harder for people
who implement broken tools in the future to discover their mistakes.
The only result of the error would be that it is slower to work with
the `packed-refs` files that they wrote. Such an error could go
undiscovered for a long time.
My opinion is that we shouldn't maintain back-compat with formats that
may have been written by another tool because Git wasn't strict about
it. As long as Git never wrote files with these formats, then they
shouldn't be supported.
You are absolutely right that staying strict will help discover the
tools that are writing an incorrect format.
Since most heavily-used tools that didn't spawn Git processes use
LibGit2 to interact with Git repos, I added Ed Thomson to CC to see if
libgit2 could ever write these bad header comments.
Thanks for writing this RFC so we can have the discussion and more
quickly identify this issue if/when users are broken.
Thanks,
-Stolee
