Why Random.jl is fixed to version 0.0.0?

General Usage
, ,
1

I’m trying to explicitly specify the compatibilities of some standard packages as recently required by the General registry. I added Random = "1.9" (or "1", I tried both, resulting in the same error) to Project.toml of NNlib.jl but the CI ended up with the following error in an integration test:

ERROR: LoadError: Unsatisfiable requirements detected for package Random [9a3f8284]:
 Random [9a3f8284] log:
 ├─possible versions are: 0.0.0 or uninstalled (package in sysimage!)
 ├─restricted to versions 1.6.0-1 by Lux [b2108857] — no versions left
 │ └─Lux [b2108857] log:
 │   ├─possible versions are: 0.5.9 or uninstalled
 │   └─Lux [b2108857] is fixed to version 0.5.9
 └─Random [9a3f8284] is fixed to version 0.0.0

(see Set compatibilities for standard packages · FluxML/NNlib.jl@25d4eca · GitHub for the complete messages)

Random.jl 1.6.0 or newer is required by Lux.jl and it was run on Julia 1.9.3, so there must be a natural compatibility. However, Random.jl is fixed to version 0.0.0 for some reason and we cannot satisfy the compatibility. What is the problem of this? Or do I misunderstand the error message?

2

Basically, this happens if either:

  1. You are using Julia 1.3 or earlier.
  2. You are using Julia 1.4+, and you do Pkg operations inside your test suite.

For almost all stdlibs (including Random), the fix is:

[compat]
StdlibName = "<0.0.1, 1"

For the SHA stdlib specifically, the fix is:

[compat]
SHA = "<0.0.1, 0.7, 1"
7 Likes
3

Just had to use this fix (thanks Dilum!). However I noticed this fix also downgrades many other packages to the version they had before the [compat] entries for stdlib was implemented. The only way this fix could work in practice is if all packages had the <0.0.1 fix. But nearly all packages do not :confused:

Is there any solution that fixes this issue without forcing a downgrade of any package that does not use the <0.0.1 spec?

4

We have to go back and retroactively modify the compat entries of old package versions.

5

Unfortunately I have 110 indirect dependencies :confused: So seems infeasible to correct all the compat. Are there any alternatives to this?

6

Or could we do this en masse in the registry?

7

Seems like it would be better to special-case this in the Pkg system — if a package depends on a stdlib and doesn’t have a compat entry, it should be automatically interpreted as "<0.0.1, 1" etc.

(Though that would leave older versions of Julia in the lurch?)

8

Right, and the issue is only seen on older versions of Julia. One possible fix is to automatically add the <0.0.1 to every stdlib [compat] in the registry, since then it would propagate to the older Julia as well.

9

Yeah, we should likely just make a single mass PR to the registry.

2 Likes