Evaluate API fails when there is type mismatch between actual and predicted fields.

[przemekwitek]

If fields (actualField, predictedField) are misconfigured and it happens that their mapping types differ (e.g. long vs boolean) then Evaluate API fails with non-intuitive error message mentioning script failure (the script we are using internally to calculate the metrics).

Example failures:

{
  "error" : {
    "root_cause" : [
      {
        "type" : "script_exception",
        "reason" : "runtime error",
        "script_stack" : [
          "java.base/java.lang.Class.cast(Class.java:3604)",
          "doc['mal_count'].value == doc['ml.label_prediction'].value",
          "                                                    ^---- HERE"

{
  "error" : {
    "root_cause" : [
      {
        "type" : "illegal_argument_exception",
        "reason" : "Can't parse boolean value [1337x], expected [true] or [false]"

[elasticmachine]

Pinging @elastic/ml-core (:ml)

[przemekwitek]

I closed this issue as now the Evaluate API does not throw any strange script-related errors in face of type mismatch situation.

This approach may be revisited in the future. I can imagine adding a validation where we would check upfront what the actual and predicted fields types are and deciding whether this evaluation request makes sense or not.
Also, we may think of matching in a smart manner. E.g.: matching a boolean field (that only has true or false values) with a numeric field (that only has 1 or 0 values).