We should consider swapping $match with $setWindowFields. I think it's valid when:

• within each partition, the predicate is either always true or always false
• the $match doesn't depend on any 'output' field

For example, in this query:

{$setWindowFields: {
    partitionBy: ["$state", "$city"],
    output: {total: {$sum: "$x"}},
}},
{$match: {state: "NY"}},

Doing the $match first shouldn't change the result, because it drops whole partitions.

However, this could be tricky given how we desugar $setWindowFields:

{$set: {__tmp: ["$state", "$city"]}},
{$sort: {__tmp: 1}},
{$_internalSetWindowFields: {
    partitionBy: "$__tmp",
    output: {total: {$sum: "$x"}},
}},
{$unset: 'tmp'},
{$match: {state: "NY"}},

It will be hard for the optimizer to see the relationship between {state: "NY"} and partitionBy: "$__tmp". Some things that could help are:

• a new analysis (functional dependency)
• ability to $sort by expression, instead of a __tmp field
• a way to defer desugaring $setWindowFields until after some optimization