HPO JSON schema validation

[agarwalrounak]

We have 3 JSON objects coming in to the HPO module:

• In API 1, RM does a POST with the JSON object below

  {
      "id" : "sha256sum of autotune object",
      "url" : "http://localhost:8080/searchSpace",
      "operation" : "EXP_TRIAL_GENERATE_NEW | EXP_TRIAL_GENERATE_SUBSEQUENT"
  }
  

  The corresponding schema for the above JSON is:

  {
      "type" : "object",
      "properties" : {
          "id" : {"type" : "string"},
          "url" : {
              "type" : "string",
              "format": "uri"
          },
          "operation" : {"type" : "string"}
      },
      "required": ["id", "url", "operation"]
   }
  
  

• In API 2, RM does a POST with the JSON object below

  {
      "id" : "sha256sum of autotune object",
      "trial_number": "xyz",
      "trial_result": "success | failure | error",
      "result_value_type": "double",
      "result_value": "abc",
      "operation" : "EXP_TRIAL_RESULT"
  }
  

  The corresponding schema for the above JSON is:

  {
      "type" : "object",
      "properties" : {
          "id" : {"type" : "string"},
          "trial_number" : {"type" : "integer"},
          "trial_result" : {
              "enum": [
                  "success",
                  "failure",
                  "error"
              ]
          },
          "result_value_type" : {"type" : "string"},
          "result_value" : {"type" : "number"},
          "operation" : {"type" : "string"}
      },
      "required": ["id", "trial_number", "trial_result", "result_value_type", "result_value", "operation"]
  }
  

• The searchspace JSON object

  {
      "id": "auto123",
      "application_name": "petclinic-deployment-6d4c8678d4-jmz8x",
      "objective_function": "transaction_response_time",
      "value_type": "double",
      "direction": "minimize",
      "hpo_algo_impl": "optuna_tpe",
      "tunables": [
          {
              "name": "cpu_request",
              "value_type": "double",
              "upper_bound": 6,
              "lower_bound": 1,
              "step": 0.01
          },
          {
              "name": "memory_request",
              "value_type": "integer",
              "upper_bound": 1024,
              "lower_bound": 100,
              "step": 1
          }
      ]
  }
  

  The corresponding schema for the above JSON is:

  {
      "type" : "object",
      "properties" : {
          "id" : {"type" : "string"},
          "application_name" : {"type" : "string"},
          "objective_function" : {"type" : "string"},
          "value_type" : {
              "enum": [
                  "double",
                  "integer",
                  "categorical"
              ]
          },
          "direction" : {
              "enum": [
                  "minimize",
                  "maximize",
              ]
          },
          "hpo_algo_impl" : {"type" : "string"},
          "tunables": {
              "type": "array",
              "items": {
                  "type": "object",
                  "properties": {
                      "name": {"type": "string"},
                      "value_type": {"type": "string"},
                      "upper_bound": {"type": "number"},
                      "lower_bound": {"type": "number"},
                      "step": {"type": "number"}
                  },
                  "required": ["name", "value_type", "upper_bound", "lower_bound", "step"]
              }
          }
      },
      "required": ["id", "application_name", "objective_function", "value_type", "direction", "hpo_algo_impl", "tunables"]
  }
  

[agarwalrounak]

@dinogun The value type of a tunable will be one of double, integer, categorical and boolean.

What about the value type of the overall objective function(throughput/response time)? Will it always be a numerical value?