The bugfix PR for https://jira.mongodb.org/browse/SERVER-93058 (https://github.com/10gen/mongo/pull/25784) caused some testsuite to fail.

https://jira.mongodb.org/browse/BF-34620 was opened to track these test failures.

The test that fails is

        uncaught exception: Error: command failed: {
                "ok" : 0,
                "errmsg" : "pipeline is required to run on mongoS, but cannot :: caused by :: $merge must write to disk",
                "code" : 20,
                "codeName" : "IllegalOperation",
                "$clusterTime" : {
                        "clusterTime" : Timestamp(1724143331, 19),
                        "signature" : {
                                "hash" : BinData(0,"AAAAAAAAAAAAAAAAAAAAAAAAAAA="),
                                "keyId" : NumberLong(0)

The aggregate command that triggers the failure is the following:

        db.aggregate([
            {$documents: [updatedDoc]},
            {
                $merge: {
                    into: {db: coll.getDB().getName(), coll: coll.getName()},
                    on: "_id",
                    whenMatched: "merge",
                    whenNotMatched: "fail"
                }
            }
        ]);

It is an aggregate which does not run on a particular collection, but runs on the database with a $documents input.

The database `db` at the point of execution of the aggregate command is a database named `test`, which does not exist.

When executing the aggregation command on the replica set, there is no collection routing information (CRI) for the db `test`, so the following code decides that the merge must run on mongos:

https://github.com/10gen/mongo/blob/c7df2842470f7ef497684f781a4203d69bc7aac6/src/mongo/s/query/planner/cluster_aggregation_planner.cpp#L773

But as the merge uses the `$out` operator, it cannot be executed on mongos, and the command fails.

Changing the aggregate command in the test so that it uses `testDB` instead of `db` fixes the issue. The reason is that `testDB` points to an existing database when the test executes.

It is current unclear if the test is supposed to work when running the aggregate on a db that does not exist.

The manual (https://www.mongodb.com/docs/manual/reference/operator/aggregation/merge/) states that

• If the output collection does not exist, $merge creates the collection:

• • For a replica set or a standalone, if the output database does not exist, $merge also creates the database.

• • For a sharded cluster, the specified output database must already exist.

But here the problem is not the output collection, but rather the input.