I have been exploring using Vertex AI for my machine learning workflows. Because deploying different models to the same endpoint utilizing only one node is not possible in Vertex AI, I am considering a workaround. With this workaround, I will be unable to use many Vertex AI features, like model monitoring, feature attribution etc., and it simply becomes, I think, a managed alternative to running the prediction application on, say, a GKE cluster. So, besides the cost difference, I am exploring if running the custom prediction container on Vertex AI vs. GKE will involve any limitations, for example, only N1 machine types are available for prediction in Vertex AI
There is a similar question, but I it does not raise the specific questions I hope to have answered.
I am not sure of the available disk space. In Vertex AI, one can specify the machine type, such as n1-standard-2 etc., but I am not sure what disk space will be available and if/how one can specify it? In the custom container code, I may copy multiple model artifacts, or data from outside sources to the local directory before processing them so understanding any disk space limitations is important.
For custom training in Vertex AI, one can use an interactive shell to inspect the container where the training code is running, as described here. Is something like this possible for a custom prediction container? I have not found anything in the docs.
For custom training, one can use a private IP for custom training as described here. Again, I have not found anything similar for custom prediction in the docs, is it possible?
If you know of any other possible limitations, please post.
we don't specify a disk size, so default to 100GB
I'm not aware of this right now. But if it's a custom container, you could just run it locally or on GKE for debugging purpose.
are you looking for this? https://cloud.google.com/vertex-ai/docs/predictions/using-private-endpoints
Related
I have the task of optimizing search on the website. The search should be for pictures and for text by text query. I have already developed, trained, tested and selected a machine learning model that transforms images and text into a feature vector (Python, based on OpenAI CLIP). This feature vector will be transferred to Elastic Search. Elastic Search will be configured by another specialist.
The model will be used first to determine the feature vector on all existing images and texts, and then be used whenever new content is added or existing content is changed.
There is a lot of existing content (approximately several tens of millions of pictures and texts together). About 100-500 pieces of content are added and changed per day.
I haven't worked much with AWS, but in this case the model needs to be deployed to AWS somehow. Of course, I have the model and the entire project locally, I can write an API app and make a Docker container.
The question is, what is the best method to deploy this application on AWS? The best in terms of speed and ease of implementation (for me as an AWS beginner), as well as cost optimization, taking into account the number of requests for the application.
I've seen different possibilities, from simply deploying the application on EC2 (probably the easiest option) to using SageMaker. Also Kubernetes and ECS...
I'd recommend using SageMaker Hosting endpoint if you need to be able to run vectorization in near-real time any time of the day, or in a SageMaker Training job if you can run vectorization batched, for example once every few hour.
For both systems you can use pre-defined Framework containers and SDK to which you pass a Python code and optionally requirements.txt, or you can create your own image.
I have a set of calculations that needs to run in a batch, and the workload is easily parallelized across machines. The work to be done is already done within a Docker container. I'm trying to understand the easiest way for me to run this workload in a highly parallel way on AWS. However, in trying to figure out where to begin I'm having trouble finding the right entrypoint. I read about AWS Batch and AWS Fargate, but each time I try to go down one of those paths to learn about them in more detail, more AWS services start popping up (Lamdas, Step Functions, ECS, AutoScaling groups), with each article having a different combination. Furthermore, I start thinking about the problem as a Batch vs Fargate problem, and then I find another article that talks about Batch + Fargate, or X + ECS + ....
I'm having trouble finding the appropriate introduction to the choices so I can get started with setting something up and getting some experience. Any pointers on which direction I might go or some resources for me to look at?
AWS containers services team member here. Your question triggers all my button cause I have been working on a deliverable to address some of this confusion ("where do I start with xyz?"). I can try to answer your question briefly here but if you want to read more (perhaps way more than you'd need feel free to contact me offline (mreferre at amazon dot com will work).
First and foremost it's not a Vs but it's an AND. Think of all these products you mention being distributed at different layers of the stack (this is a draft visual in the deliverable):
Fargate represents capacity (where your container is running), ECS represents a core containers orchestrator and Batch is one of the provisioners on top of the container orchestrator. Lambda is something separate and that live on its own. The options for your specific use case seem to be:
Lambda
ECS/Fargate
Batch/ECS/Fargate
Step Functions/ECS/Fargate (this one is outside of analysis and you don't see it in my visual - wondering if I should add it).
As others have hinted you probably want to use Lambda if your model is event-driven (e.g. if you want to fire up a dedicated function for every event like a new file uploaded to S3).
You probably do not want to use a naked ECS/Fargate solution because it would require more work to deal with the triggering and the scheduling of your batch jobs.
You probably want to use either Batch or Step Functions to schedule jobs on ECS/Fargate. I'd argue SF is good if you have basic workflows that you need to deal with and Batch if you need to manage complex jobs at scale. Perhaps this 35 mins presentation that I did last year can provide a bit more background on these Batch Vs SF differences.
Let me know if you have any additional questions because this discussion is super useful for the positioning I am trying to build.
When I train my model locally I use a 20% test set and then cross validation. Sagameker seems like it needs a dedicated valdiation set (at least in the tutorials I've followed). Currently I have 20% test, 10% validation leaving 70% to train - so I lose 10% of my training data compared to when I train locally, and there is some performance loss as a results of this.
I could just take my locally trained models and overwrite the sagemaker models stored in s3, but that seems like a bit of a work around. Is there a way to use Sagemaker without having to have a dedicated validation set?
Thanks
SageMaker seems to allow a single training set while in cross validation you iterate between for example 5 different training set each one validated on a different hold out set. So it seems that SageMaker training service is not well suited for cross validation. Of course cross validation is usually useful with small (to be accurate low variance) data, so in those cases you can set the training infrastructure to local (so it doesn't take a lot of time) and then iterate manually to achieve cross validation functionality. But it's not something out of the box.
Sorry, can you please elaborate which tutorials you are referring to when you say "SageMaker seems like it needs a dedicated validation set (at least in the tutorials I've followed)."
SageMaker training exposes the ability to separate datasets into "channels" so you can separate your dataset in whichever way you please.
See here for more info: https://docs.aws.amazon.com/sagemaker/latest/dg/your-algorithms-training-algo-running-container.html#your-algorithms-training-algo-running-container-trainingdata
My company is currently evaluating hyperledger(fabric) and we're using it for our POC. It looks very promising and we're targeting rolling out to production in next few months.
We're targeting AWS as our production environment.
However, we're struggling to find good tutorial/practices/recommendations about operating hyperledger network in such environment.
I'm aware that Cello is aiming to solve/ease deploying/monitoring hyperledger network but i also read that its not production ready yet. Question is, should we even consider looking at Cello at this point?
If not, what are our alternatives? Docker swarm, kubernetes?
I also didn't find information about recommended instance types. I understand this is application and AWS specific but what are the minimal system requirements
(memory&CPU&network) for example for 'peer' node (our application is not network intensive, nor a lot of transactions will be submitted per hour/day, only few of them per day).
Another question is where to create those instances on AWS from geographical&decentralization point of view. Does it make sense all of them to be created in same region? Or, we must create instances running in different regions?
Tnx a lot.
Igor.
yes, look at Cello.. if nothing else it will help you see the aws deployment model.
really nothing special..
design the desired system, peers, orderer, gateways, etc..
then decide who many ec2 instance u need to support that.
as for WHERE (region).. depends on where the connecting application is and what kind of fault tolerance you need for your business model.
one of the businesses I am working with wants a minimum of 99.99999 % availability. so, multi-region is critical. its just another ec2 instance with sockets open from different hosts..
aws doesn't provide much in terms of support for hyperledger. they have some templates which allow you to setup the VMs initially, but that's stuff you can do yourself as well.
you are right, the documentation is very light and most of the time confusing. I got to the point where I can start from scratch with a brand new VM and got everything ready and deploy my own network definition and chaincode and have the scripts to do that.
IBM cloud has much better support for hyperledger however. you can design your network visually, you can download your connection profiles, deploy and instantiate chaincode, create and join channels, handle certificates, pretty much everything you need to run and support such a network. It's light years ahead of AWS. They even have a full CI / CD pipepline that you could replicate for your own project. if you look at their marbles demo, you'll see what i mean.
Cello is definitely worth looking at, with the caveat that it's incubation meaning, not real yet, not production ready and not really useful until it becomes a fully fledged product.
Google Compute Engine lets you get a group of instances that are semantically local in the sense that only they can talk to each other and all external access has to go through a firewall etc. If I want to run Map-Reduce or other kinds of cluster jobs that are going to induce high network traffic, then I also want machines that are physically local (say, on the same rack). Looking at the APIs and initial documentation, I don't see any way to request that; does anyone know otherwise?
There is no support in GCE right now for specifying rack locality. However, we built the system to work well in the face of large numbers of instances talking to each other in a fully connected way, as long as they are in the same zone.
This is one of the things that allowed MapR to approach the record for a hadoop terasort. You can see that in action in the video for the Criag Mcluckie's talk from IO:
https://developers.google.com/events/io/sessions/gooio2012/302/
The best way to see is to test out your application and see how it works.