On 2/23/2018 2:33 PM, Junio C Hamano wrote:
Derrick Stolee <sto...@gmail.com> writes:
The (unlikely, but possible) race condition involves two processes (P1
and P2):
1. P1 reads from graph-latest to see commit graph file F1.
2. P2 updates graph-latest to point to F2 and deletes F1.
3. P1 tries to read F1 and fails.
I could explicitly mention this condition in the message, or we can
just let P2 fail by deleting all files other than the one referenced
by 'graph-latest'. Thoughts?
The established way we do this is not to have -latest pointer, I
would think, and instead, make -latest be the actual thing. That is
how $GIT_DIR/index is updated, for example, by first writing into a
temporary file and then moving it to the final destination. Is
there a reason why the same pattern cannot be used?
My thought was that using a fixed name (e.g.
.git/objects/info/commit-graph) would block making the graph
incremental. Upon thinking again, this is not the case. That feature
could be designed with a fixed name for the small, frequently-updated
file and use .../info/graph-<hash>.graph names for the "base" graph files.
I'll spend some time thinking about the ramifications of this fixed-name
concept. At the moment, it would remove two commits from this patch
series, which is nice.
Thanks,
-Stolee
