The $sample stage returns a different sample every time it runs. $lookup sometimes re-runs the inner pipeline per outer document, and sometimes runs it only once. This makes the behavior of $sample inside $lookup hard to predict.

For example, this query runs the sub-pipeline only once, resulting in the same sample chosen every time:

{$lookup: {
	from: 'foreign_coll',
	pipeline: [
		{$sample: {size: 5}},
	],
	as: 'docs',
}}

On the other hand, this query re-runs the sub-pipeline, choosing a different sample per outer document:

{$lookup: {
	from: 'foreign_coll',
	let: {outer: "$_id"},
	pipeline: [
		{$match: {$expr: {$lt:["$_id", "$$outer"]}}},  // correlation predicate
		{$sample: {size: 3}},
	],
	as: 'docs',
}}

Since we consider DocumentSourceSequentialDocumentCache to be an optimization, there could be other exceptions to this rule. For example, if you add a dummy correlation hoping to force the inner pipeline to re-run, it can get optimized out.

This ticket will make changes to consider any $sample stage or stage containing a $rand or $sampleRate expression to be ineligible for uncorrelated pipeline caching.