During timeseries retriable updates, we choose an executor based on the cluster role from the server global parameters. If the node has been started with --shardsvr but has not yet written a shard identity document (as is the case when transitioning a replica set to sharded cluster before the addShard command has been run) the attempt to get the the executor from the grid will invariant that the grid is not yet initialized.

Note that I don't think using the other executor (the ReplicaSetNodeProcessInterface one) is a viable alternative as we don't start it up if the node is started with --shardsvr.