Document Understanding - Public Endpoints regional deprecation

:wave: Hello to all friends of Document Understanding, we have an important announcement to make!

Some of our public endpoints for pre-trained Machine Learning document extraction models have been deprecated as called out on the documentation page here. This has now taken effect as of Dec 6th 2021.

For those users who have data residency constraints requiring them to process data within one specific region, you have the option to use AI Center as part of Automation Cloud, on a tenant in your region. You create a tenant in that region, then you create the ML Package of your choice and you deploy the minor version 0 (zero) of that ML Package. The minor version 0 corresponds to the identical model to the one deployed in the corresponding Public Endpoint.

Thank you and Happy Holidays! :christmas_tree:
The Document Understanding Team


Thanks for sharing!

Was this notification sent out in any formal notification to production level license holders? I do not recall seeing this notification and had to find out via all my machine learning extractions failing. I am in the US.

Hi @potchak , great question,

There was a notification sent during month of November to all customers who had processed more than 20 pages during the whole month of October through the public endpoints hosted in regions other than West Europe. The volumes we were seeing on those endpoints were very low, not compatible with any kind of production usage, and also the public endpoints get automatically updated on every release with the new models. The behavior on any one particular usecase is unpredictable however, because we cannot test on the data of every customer separately - and customers generally cannot tolerate that kind of unpredictability in their production automations.

Consequently, we strongly recommend that for any production usage you use ML Skills deployed in AI Center, which you have full control over, rather than public endpoints which can be updated at any time. Public endpoints are targeted at Community users, or for testing, PoCs, or demos for Enterprise users, or for Prelabelling usage as part of labelling training or evaluation datasets in Data Manager. But these are all very low volume uses, which do not justify maintaining the full production infrastructure for all models (of which there are already 15 with more coming) on all 5 regions totalling 75 services.


1 Like