README.md

Shared Node Pools

What are they

When you deploy a recipe via Corrino, the underlying infrastructure is spun up (separate node pool for each recipe within the Corrino cluster) and the application layer is deployed on top of that infrastructure. Once you are done with recipe and undeploy it, the application layer and the infrastructure gets spun down (the node pool is deleted). This creates an issue when you want to quickly deploy and undeploy recipes onto infrastructure that requires a long recycle time (such as bare metal shapes) or you want to deploy multiple recipes onto the same underlying infrastructure (ex: recipe A uses 2 GPUs and recipe B uses 2 GPUs on a shape with 4 GPUs).

In come shared node pools, where you can launch infrastructure independent of a recipe. You can launch a shared node pool, and deploy/undeploy recipes onto the same node pool (underlying infrastructure) - removing the need to spin up new infrastructure for every recipe that is deployed.

How to use them

You can create a shared node pool with a selector or without a selector. A selector is a naming convention that you can use to ensure specific recipes land on specific shared node pools.

With selector

1. Create a shared node pool with a selector (this would be the payload to the /deployment POST API):

{
	"deployment_name": "BM.GPU4.8 shared pool",
	"recipe_mode": "shared_node_pool",
	"shared_node_pool_size": 1,
	"shared_node_pool_shape": "BM.GPU4.8",
	"shared_node_pool_boot_volume_size_in_gbs": 500,
	"shared_node_pool_selector": "selector_1"
}

2. Wait for the shared node pool to be deployed (can check deployment_logs for status)
3. Deploy any recipe that you want to be deployed on the BM.GPU4.8 "selector_1" shared node pool (this would be the payload to the /deployment POST API):

{
	"recipe_id": "example",
	"recipe_mode": "service",
	"deployment_name": "echo1 selector_1",
	"recipe_use_shared_node_pool": true,
	"recipe_shared_node_pool_selector": "selector_1",
	"recipe_image_uri": "hashicorp/http-echo",
	"recipe_container_command_args": [
	"-text=corrino"
	],
	"recipe_container_port": "5678",
	"recipe_node_shape": "BM.GPU4.8",
	"recipe_replica_count": 2
}

Note that the parameters:

• recipe_use_shared_node_pool ensures that we are using a shared node pool for this recipe (and not launching new infrastructure
• recipe_shared_node_pool_selector ensures that we are deploying this recipe onto the BM.GPU4.8 shared node pool we deployed in Step 1
• recipe_node_shape needs to match the shape of the shared node pool we launched in step 1 (regardless of including the selector or not)

** If no shared node pool of shape BM.GPU4.8 with selector selector_1 exisited, the recipe would wait to be deployed until that shared node pool was created

Without Selector

1. Create a shared node pool without a selector (this would be the payload to the /deployment POST API):

{
	"deployment_name": "BM.GPU4.8 shared pool2",
	"recipe_mode": "shared_node_pool",
	"shared_node_pool_size": 1,
	"shared_node_pool_shape": "BM.GPU4.8",
	"shared_node_pool_boot_volume_size_in_gbs": 500,
}

2. Wait for the shared node pool to be deployed (can check deployment_logs for status)
3. Deploy any recipe that you want to be deployed on the BM.GPU4.8 "selector_1" shared node pool (this would be the payload to the /deployment POST API):

{
	"recipe_id": "example",
	"recipe_mode": "service",
	"deployment_name": "echo2 selector_1",
	"recipe_use_shared_node_pool": true,
	"recipe_image_uri": "hashicorp/http-echo",
	"recipe_container_command_args": [
	"-text=corrino"
	],
	"recipe_container_port": "5678",
	"recipe_node_shape": "BM.GPU4.8",
	"recipe_replica_count": 2
}

Note that the parameters:

• recipe_use_shared_node_pool ensures that we are using a shared node pool for this recipe (and not launching new infrastructure
• recipe_node_shape needs to match the shape of the shared node pool we launched in step 1 (regardless of including the selector or not)

This recipe will be deployed onto any BM.GPU4.8 shared node pool since no selector was included in the recipe For example, if you had two BM.GPU4.8 shared node pools, then it will randomly select one for deployment

Considerations

• If you do not have a shared node pool deployed but try to deploy a recipe using the recipe_use_shared_node_pool, Corrino will wait to deploy the recipe until it has the shared node pool to launch the recipe onto
• Bare metal shape shared node pools require the shared_node_pool_boot_volume_size_in_gbs parameter
• Any recipe is compatible with shared node pools

Sample Recipes

shared_node_pool_A10_BM