Semantic versioning

[juliohm]

This package is a dependency in Meshes.jl and we recently started having issues with a non-working resize! function. I think this is probably related to semantic versioning misunderstanding...

When you updated from 1.1 to 1.2 to 1.3 you adopted the following rule? https://semver.org/spec/v2.0.0.html#spec-item-7

Our tests are failing now downstream. Could you please double check and help fix the issues?

Also, I reiterate the importance of placing this package in an organization where multiple members can help maintain and double check these things before submitting disruptive new releases.

[serenity4]

The bug in question can be reproduced in 1.3.0:

julia> using CircularArrays

julia> vec = CircularVector([1, 2])
2-element CircularVector(::Vector{Int64}):
 1
 2

julia> push!(vec, 3)
ERROR: MethodError: no method matching resize!(::CircularVector{Int64, Vector{Int64}}, ::Int64)
Closest candidates are:
  resize!(::Vector{T} where T, ::Integer) at array.jl:1101
  resize!(::BitVector, ::Integer) at bitarray.jl:810
Stacktrace:
 [1] _append!(a::CircularVector{Int64, Vector{Int64}}, #unused#::Base.HasLength, iter::Tuple{Int64})
   @ Base ./array.jl:989
 [2] append!(a::CircularVector{Int64, Vector{Int64}}, iter::Tuple{Int64})
   @ Base ./array.jl:981
 [3] push!(a::CircularVector{Int64, Vector{Int64}}, iter::Int64)
   @ Base ./array.jl:982
 [4] top-level scope
   @ REPL[7]:1

but can also be reproduced in 1.1.0. I am not sure why it worked before on Meshes and why it no longer does. Anyway, the problem seems to come from resize! not being forwarded to the wrapped array.

[serenity4]

I see that pushing/appending to a circular array does not work and is even tested for not working. The problem seems to come deeper from a change in behavior on broadcasting, where the result of a broadcast yields a value of type CircularArray whereas before it would produce a vector.

[Vexatos]

Calling push! on a CircularVector is indeed supposed to not work. Here is the test case for it. For the same reason, resize! was never supposed to work. The reason for this decision when I made this package was that being able to resize a circular array would make the circular behaviour no longer trivially predictable. It's not really certain whether push! should add elements to the start or end of such an array as, conceptually, it should not be possible to determine its start or end at all (superficially). You can of course always resize the parent() of the circular array if you really need to, which is probably the recommended way to do it anyway.

1.3.0 introduced a fix to broadcast behaviour where broadcasting over a circular array would produce a circular array where applicable (#15). Broadcasting not returning a circular array was a bug and relying on the type conversion during a broadcast operation is something that should arguably be fixed in the code downstream. The update introduced a larger number of methods, so if any regressions occur, be sure to let me know. However, without any reproduction steps or examples, I cannot do much.

[serenity4]

Indeed we found that we were relying on broadcasting giving a vector, which was a happy coincidence that made the code work but definitely not a better behavior. The rationale behind not allowing push! on a CircularVector makes total sense, using parent seems to do the trick.

@juliohm I think we can close this issue.

[juliohm]

Thank you all, let's close the issue and try to fix it downstream then. 👍🏽