When Grit writes a loose object via the LooseStorage class,
it just opens the object file and starts writing. This works
most of the time, but can be a problem in some corner cases,
1. If another process tries to write the same object
simultaneously, the writes may be interleaved and the
object can be corrupted.
2. If another process tries to read the object
simultaneously, it may see the object in a half-written
3. If the process or machine crashes during the write, we
may leave a half-written corrupt object.
This can be solved by writing the object to a temporary file
and linking it into place. This is the same strategy used by
When we write a loose object to disk, we simply close the
file object before moving it into place. If the machine
crashes shortly after our write, the contents may not have
been committed to disk (depending your filesystem, usually
the metadata is, and you end up with a corrupt, zero-length
loose object file).
This is especially bad because we report that the object is
successfully written, which means we may have updated refs
to point to it. A corrupt object at that point means not
only does the operation fail, but the repository is left in
a corrupted and unusable state.
We can fix this by calling fsync on the object file before
linking it into place. Between this and the previous commit,
our object writing should now behave exactly like git's