When starting the cloning phase of a migration, the donor shard will insert a document into the config.migrationCoordinators collection. If this fails and throws an exception, it will trigger the MigrationSourceManager::cleanupOnError() scope guard, which will try to complete the migration by persisting an abort decision through an update to the document that failed to be inserted, which will fail because there is no matching document. Persisting the decision retries on errors until a stepdown or shutdown, so until that happens, the migration will hang trying to update the non-existent document.

UPDATE fixing this issue exposed another problem, if a migration coordinator document is left without decision (like for example, because the document insert failed to honor the majority write concern) then another migration on the same session that increases the transaction number would cause the bump of the txnNumber to fail during the next recovery with a TransactionTooOld error, as can be seen on the linked BF.