For $lookup inner queries that have only one query solution, we do not multiplan and therefore do not cache. However caching these plans is beneficial to performance, we are seeing an 8% overall improvement in TPCH Q11 throughput and 3% improvement in TPCH Q7.

To avoid affecting a large class of queries, we should only force the cache when the query comes from a lookup. We can set a variable via the ExpressionContext to communicate this flag.

Slack discussion