In SERVER-46077, we created a new passthrough suite which runs the tests in jstests/aggregation/ against a replica set. It configures causal consistency and runs all read operations against a secondary. Our intention was to reuse $out/$merge test coverage to make sure that these operations still work when they run on a replica set secondary.

Since the implementation of $out/$merge running on a secondary differs depending on whether the cluster is a single replica set or is sharded, we need to get similar test coverage in the sharded case. For this purpose, we should introduce a variant of the aggregation_secondary_reads passthrough which runs against a sharded cluster. I would recommend that we test against a 2-shard cluster (where each shard is a 2-node or 3-node replica set) and that we implicitly shard accessed collections.