Chunk migration today uses read preference 'PrimaryOnly' when reading the data to migrate. To enable better workload isolation between application and migration, we should expose a way have this read preference be configurable.
The current mechanism needs proper modification to support reading from a secondary. For the initial clone, read from a secondary using a snapshot at timestamp X, ensuring the secondary has replicated at least up to X before starting. For the catch-up phase, read modifications with timestamp > X from the same secondary used for the initial clone.