Randomizer::pickArrayKeys() has intransparent external behavior

[TimWolla]

Description

The following code:

<?php

$r1 = new Random\Randomizer(new Random\Engine\Xoshiro256StarStar(1));
$r2 = new Random\Randomizer(new Random\Engine\Xoshiro256StarStar(1));

$a = [ 'a' => 1, 'b' => 2, 'c' => 3, 'd' => 4, 'e' => 5, 'f' => 6, ];
unset($a['b']);

$b = [ 'a' => 1, 'c' => 3, 'd' => 4, 'e' => 5, 'f' => 6, ];

var_dump($a === $b); // bool(true)

var_dump(
	$r1->pickArrayKeys($a, 1), // [ 0 => 'e' ]
	$r2->pickArrayKeys($b, 1), // [ 0 => 'd' ]
);

Resulted in this output:

bool(true)
array(1) {
  [0]=>
  string(1) "e"
}
array(1) {
  [0]=>
  string(1) "d"
}

But I expected this output instead:

The same key selected, because the arrays are identical.

PHP Version

Current git master

Operating System

No response

[TimWolla]

Not sure if this is really fixable without breaking compatibility one way or another; and without adding performance issues. I'm primarily filing this issue so that this is tracked once somewhere and can be added to the documentation as a caveat.

[cmb69]

array_rand() already exhibited that behavior for quite some time (https://3v4l.org/g0uNE). I agree that it is not worth fixing, and I'm not even sure if we need to document that (although it won't hurt).

[TimWolla]

I'm not even sure if we need to document that (although it won't hurt).

During the RFC process it was a very important aspect for me were the promises the "Randomizer" gives to the user. In the end it was decided that any externally-visible change to the Randomizer's behavior is considered breaking and needs to be approved by another RFC: Users should be able to rely on their tests remaining stable for as long as possible. For engines the promise is even stronger: They must match their reference implementation 1:1 (that's why they return a bytestring instead of an integer).

As such I consider it important to document that pickArrayKeys()'s return value depends on the internal array structure (e.g. the construction order) and might not be 100% reproducible.

[cmb69]

Fair enough. Transferring to doc-en.