search

Azure-Samples / graphrag-accelerator

One-click deploy of a Knowledge Graph powered RAG (GraphRAG) in Azure
https://github.com/microsoft/graphrag
MIT License
1.65k stars 250 forks source link

[FEATURE] Other VM SKU options are required #112

Closed kevinzhg closed 1 month ago

kevinzhg commented 1 month ago

Is your feature request related to a problem? Please describe. it's related to the Azure deployment cost issue.

Describe the solution you'd like Have the other option less costy than E16as v5


Describe alternatives you've considered A cheaper VM Size?

Additional context No more context.

timothymeyers commented 1 month ago

@kevinzhg this is an example reference implementation with some recommended defaults. You can absolutely deploy it with whatever VM Size best meets your requirements.

rnpramasamyai commented 1 month ago

I changed the default VM size. It created AKS, but the pods were not running. I requested a higher quota, changed the VM size back to default, and then I was able to run the pods in AKS successfully.

biggernaaz commented 1 month ago

@rnpramasamyai - I experienced the same behavior when I tried to reduce VM size and ended up resizing back to the default. Curious what the dependency here is. The VMSS behind the GraphRAG pod is a substantial expense.

jgbradley1 commented 1 month ago

@rnpramasamyai - I experienced the same behavior when I tried to reduce VM size and ended up resizing back to the default. Curious what the dependency here is. The VMSS behind the GraphRAG pod is a substantial expense.

I made a PR yesterday to address the cost of the accelerator when it's deployed and not being used. I introduced a new node pool with cheaper VMs and we only use a second nodepool of more expensive VMs when performing graphrag indexing.

GraphRAG indexing currently has a high memory footprint for medium-large scale datasets and that is the reason for choosing the memory optimized VM sku's.

Also some of the default sku's in this accelerator are in high demand in certain cloud regions (i.e. eastus). If you attempt to deploy to these cloud regions, the deployment may fail because of the current high demand. Even if your subscription has quota, I'm seeing that deployments still fail for this reason.