irmin-pack: Fix failure when computing summary in tree benchmark

[zshipko]

This PR makes it so ./bench/irmin-pack/tree.exe runs successfully and is able to produce a summary.

Fixes #1965

[codecov-commenter]

Codecov Report

Merging #1968 (eb6e11b) into main (abeee12) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main    #1968   +/-   ##
=======================================
  Coverage   63.95%   63.95%           
=======================================
  Files         130      130           
  Lines       15479    15479           
=======================================
  Hits         9900     9900           
  Misses       5579     5579           

Impacted Files	Coverage Δ
src/irmin-test/store.ml	94.82% <0.00%> (-0.06%)	⬇️
src/irmin-fs/unix/irmin_fs_unix.ml	68.38% <0.00%> (+0.64%)	⬆️

📣 Codecov can now indicate which changes are the most critical in Pull Requests. Learn more

[metanivek]

I don't have enough context on broader implications, but either the flag also needs to change in octez (see here and here) or the fix for this bug is somewhere deeper in the gc code.

[zshipko]

Yeah, it's very likely that there is some deeper issue but this at least gets us some usable output from the tree benchmark.

[samoht]

That's a great insight but I am not convinced by the fix :-)

Could you try to trace when an empty tree is generated/written on disk? This is never supposed to happen. I also think we should remove that option and always fail hard when this happens as this is almost always due to a bug - unless I am missing something.

[zshipko]

Right after closing this I figured out the actual issue - sorry for the noise!

It looks like the store was being initialized with an empty tree in tree.ml: https://github.com/mirage/irmin/pull/1968/files#diff-ac9a848ef33f94ac0d84a06ad3b528d96a7b45df4ce2c699c8617f0f36154a3dL89

This PR now removes that initial empty commit and everything seems to be working - luckily this means this issue does not come from the tezos trace.

[samoht]

Saving an empty root actually seems a valid use case. I would just make it a no-op so that empty trees are actually never persisted on disk.

[samoht]

Tracked in #1969 ; your fix is great for now :-)