Standard build artifact paths about hashdist

[certik]

How does git handle collisions? Or it assumes that with 32 chars the collision will never happen?

[dagss]

First, git has 40 chars (both git and hashdist use 160 bits, but bit uses base-16 and we base-32).

And with 160 bits, collisions will not happen (the chance is so low it's almost not an assumption; except if there are mathematical weaknesses in SHA-1 (git) or SHA-2 (hashdist)).

(If we simply include all 32 characters in the main artifact prefix path itself, we don't need to worry either, but I really like the shorter path.)

Here's what happens with git: If you have a collision, the first object to hit your .git/objects is the one that wins. Another object with the same commit simply never gets fetched or downloaded.

[dagss]

I now believe this must be done.

UI when having a full collision, instead of growing hash we do:

$ hit build
...
The target directory ~/.hashdist/ba/numpy/qmplawvo3kyy already exists.
This may be because of an earlier sudden crash, or because somebody
else is currently performing the build on a shared filesystem. Please run "hit purge numpy/qmplawvo3kyy" to manually clean out the directory.

Collision on short part only:

$ hit build
WARNING: An artifact with a hash that agrees in the first 12 characters
is already installed. The two hashes are:

numpy/qmplawvo3kyyvgo54pcazh4pnvfyjz53 (installed)
numpy/qmplawvo3kyyh4pnvfyf8s3vgo54pcaz (wants to install)

The odds of this happening due to chance are very low. Please get in touch with the Hashdist developer mailing list.

[certik]

How did you manage to hit the hash collision? Or is the probability not that low.

In any case, yes, it should get done.

[dagss]

What? I just made those examples up. I haven't hit any collisions, no.