Warmup & Cooldown

When you have predictable user behaviour in your product, or just want a model to scale up/down, you can drastically reduce costs while improving your user experience by making sure models are ready or completely down.

Mystic offers two endpoints to achieve this:

• Warmup - /v4/pipelines/{pipeline_id}/warmup
• Cooldown/v4/pipelines/{pipeline_id}/cooldown

Warmup

The warmup endpoint overrides the minimum nodes count for a given pipeline. You can view the API reference here

It takes in the following input:

{
  "minimum_nodes": 1,
  "duration": 60
}

• minimum_nodes: int: The minimum number of nodes to have for the pipeline
• duration: int (default=300): The number of seconds to have the minimum number nodes up for

Note: This controls the minimum number of nodes, so if you start performing runs that require more nodes for a pipeline it will scale up as needed

Example request to warmup SDXL on 2 nodes for 10minutes:

curl --request POST \
     --url https://www.mystic.ai/v4/pipelines/pipeline_8e79f00bfcbc4d00869fa3650060fb52/warmup \
     --header 'content-type: application/json' \
     --data '
{
  "minimum_nodes": 2,
  "duration": 600
}
'

Cooldown

The cooldown endpoint overrides the maximum nodes count for a given pipeline. You can view the API reference here.

It takes in the following input:

{
  "duration": 60
}

• duration: int (default=1800): The number of seconds force no nodes

📘

Runs will reset cooldowns

Cooldowns control the maximum number of nodes, so if you start performing runs it will scale up a pipeline as needed and default back to your set scaling configuration removing the cooldown

curl --request POST \
     --url https://www.mystic.ai/v4/pipelines/pipeline_8e79f00bfcbc4d00869fa3650060fb52/cooldown \
     --header 'content-type: application/json' \
     --data '{"duration":10}'