when sorting, break ties with lexical comparison

[krancour]

The manner in which comparisons are evaluated is correct. This PR doesn't seek to change that. 1.24 is equivalent to 1.24.0, however, when sorting lists that contain both these values, those two may appear in either order in relation to one another.

This PR breaks ties (only when sorting) through a lexical comparison of the original strings, thereby enabling deterministic sort results.

[mattfarina]

I looked at this with testing and I noticed that the order of the output is based on the order of the input.

For example, if I have an input of the following:

raw := []string{
	"5.6.0-alpha1",
	"1.2.3",
	"1.3",
	"1.0",
	"1.3.0",
	"2",
	"0.4.2",
	"5.6-alpha1",
}

The output ordering in the present codebase would be:

[
    0.4.2
    1.0
    1.2.3
    1.3
    1.3.0
    2
    5.6.0-alpha1
    5.6-alpha1
]

The first on the input list is the first on the output list.

Why change from that?

[krancour]

@mattfarina thanks for the response.

I looked at this with testing and I noticed that the order of the output is based on the order of the input.

I believe you may have lucked into that result. Admittedly, my best efforts to engineer a counter-example are failing, but your package doesn't offer any guarantees of the behavior you are observing. I don't know how it could since you do not implement the sorting algorithm yourself. You only implement comparison operations used in sorting.

Go currently uses the PDQ sort -- a variation on a quicksort -- and I cannot see where that algorithm would guarantee the behavior you are observing either. Even if it coincidentally did, it's only as of fairly recently that Go uses PDQ and who's to say that their default sorting algorithm cannot change again in the future?

Beyond this, I would strongly assert that the whole point of sorting is to impose a deterministic order on unordered data and that the original order of the input should have no bearing on the result. In other words, why shouldn't sorting [1.1, 1.1.0] and [1.1.0, 1.1] yield the same result? To emphasize this further, consider that the order of the input cannot be controlled in all cases.

I can also point to an actual occurrence of the present behavior causing a bug in the wild:

argoproj-labs/argocd-image-updater#375

I can only speculate what conditions may have precipitated that. Perhaps the initial order of the input was itself not consistent between runs, but as I said, it defeats the point of sorting if the order you start with matters.

[krancour]

@mattfarina I don't mean to nag, but have I convinced you well-enough that the order of the input, which often cannot be controlled anyway, should have no bearing on the results of a sort? Possible to get this merged?

[chengfang]

@mattfarina It will be great to get fix merged. In argocd-image-updater project, we've seen at least 3 reported issues caused by this.

[krancour]

@mattfarina please?