This was most likely caused by the automatic orientation detection. This feature now defaults to disable since PDF are not always compliant. In 3.0 it is possible to select all the document in a page (40) and delete the transformations. A entry can be added to the "Tools" menu to clear all transformation in the system. In the meantime the only solution would be to launch a Python shell inside the Docker instance using:
docker exec -ti <docker container name> /bin/bash
The launcher a Python shell connected to Mayan.
/usr/local/python shell
execute the following:
from documents.models import Document
from converter.models import Transformation
[Transformation.objects.get_for_model(document).delete() for document in Document.objects.all()]
I'll push the option of adding an entry to the tools menu to clear all transformation in the queue for next version.