New pricing model for AutoML - how to undeploy

[Gordon Henderson]

I'm so disappointed in the announcement for the pricing change for AutoML.  It will price me right out of the Google solution, unfortunately.  It seems you are only interested in large clients that need nodes running 24x7.  I'll have to find another solution, unfortunately.

How can I undeploy models now?  I do not want to be stuck with a large bill in September and I'd like to shut them down now.

Bummed,

-- Gordon

[Tahamina Yusuf]

Hi,

You can follow the public reference [1] regarding how to undeploy your model.

[1] https://cloud.google.com/vision/automl/object-detection/docs/undeploy#automl_vision_object_detection_undeploy_model-web 

[Wouter Dierckx]

Related: I keep getting emails that one of my google projects is affected, it appears to "have active Cloud models which need to be reviewed:".

The email mentions it's about "Google Cloud AutoML Vision & AutoML Vision Edge" services

However, when I open this project, and navigate to AutoML Vision/Models tab: there's nothing there: only an "Enable API" model for this service.

I've done this review now a few times, and I don't have any bills related to these services - is Google going to keep sending emails regardless?

[Sujoy Basu]

Hi Wouter, 

For the benefit of others who may not have read this message, I will point to our release notes dated November 12.

Please check whether the Datasets tab has any Image Classification dataset. We sent the third and final message for pricing change, which necessitated the inclusion of these customers, as they may train and deploy models again. We realized that customers are slowly migrating their models from the old prediction service. In the section on action to be taken, we pointed out that there is no action if the  Integrated UI does not display any model as deployed with ‘0 nodes’. 

Thanks,

Sujoy

[Rob Gordon]

I wanted to also echo my disappointment in the new pricing model - it prices me out as well - even as a hobbyist developer.  A real shame!

[Camilo Rodriguez]

Just a warning for other users :

I trained some models on auto ml and I forgot they were deployed and never used them.

Then suddenly out of nowhere (one month later), I receive an alert :

You have reached 50% of your billing budget

then a couple of hours later

you have reached 90% of your billing budget

then some hours late:

You have reached 100% of your billing budget.

I connected as soon as I could and undeployed the models, apparently they had been running for 120 hours.

Thank God I activated the budget alerts, otherwise the bill would come without me even noticing it.

I don't think this is very honest from Google Cloud, I mean, I agree that it is my fault to leave a model unattended, but it is unfair that the billing does not happen gradually, but on one shot.

This does not give users visibility on their infrastructure costs, you guys should definetly take a look at this...

[Jun (Cloud Platform Support)]

Hi Camilo, 

I suggest that you should consider reach out to the Billing and Payments Support to see if you can get further assistance from them regarding your situation and concerns. 

[RALPH ARAUJO]

Hi Sujoy,

This can help other people:

I'm trying to use AUTOML API ( Classification models) in a ON LINE mode.

At begining I forgot the models DEPLOYED and was billed for 24 hours a day !!!!! 

I read that I need to DEPLOY and after use UNDEPLOY the model for reduce costs of nodes. And pay only for the efective use of API (images classified)

I test two process and none works:

1- Deploy model manually at console and after : 

     -   Request Classification of ONE image

     

Do not works

2 - Remains undeployed models at console and then:

   - DEPLOYMODEL ( using NODEJs code recomended)

But this not work too !!

Can you help with this problem .

tks  

[Sujoy Basu]

Hi Ralph, 

Please use the private issue tracker, and specify your project ID, model ID, and steps to reproduce the problem.  

Thanks,

Sujoy

[RALPH ARAUJO]

Hi , Sujoy

These last few days we have studied and carried out numerous tests and we have been able to advance the processing of ON LINE transactions, with CANCELLATION after each use of the models to interrupt the collection.

Routines are running as intended, less than the TIME required to run the processes.

We are verifying that the time to DEPLOY and UNDEPLOY the model is making its use unfeasible (the processes are taking MINUTES!). The documentation says that each NODE can execute between 3/4 queries per SECOND!

That is, an online consultation is made, and if a new one is submitted then an ERROR 9 is returned - indicating that a process is still RUNNING and therefore NOT ready for a new consultation.

Recalling that our mobile application provides many users with the option to submit images for consultations, without having to control the time between consultations.

How to resolve this issue?

How to reduce the time of DEPLOYMENT and DEPLOYMENT of the models, so that it is technically acceptable and commercially viable?

[Sujoy Basu]

Hello, 

The service is not designed to support deploy / undeploy operation being called at high frequency. If you keep the model deployed, and then multiple deployment requests are sent, an error may still be returned for one of them. In that case, the request must be retried according to the "Back-off Requirements" of our Service Level Agreement.

Since the documentations shares the number of queries per second your model can support, the number of nodes deployed can be adjusted based on volume of queries from users of your mobile application. This will minimize errors. However, if that user traffic has a spiky pattern and you need to adjust node count frequently, the use case is not a good match for this product. Deploying a model takes about 20 minutes. The system is not designed for rapid change in the node count of a single model. 

Thanks,

Sujoy

[Sujoy Basu]

I need to fix a typo:  If you keep the model deployed, and then multiple online prediction requests are sent, an error may still be returned for one of them. In that case, the request must be retried according to the "Back-off Requirements" of our Service Level Agreement.