[ML] Regression job failed in coarse_parameter_search phase if dependent_variable field is unsigned_long

[wwang500]

When testing unsigned_long data type in 7.10, DFA regression job failed in coarse_parameter_search phase if dependent_variable is unsigned_long,

Error in es log:

[2020-10-30T00:03:25,932][INFO ][o.e.x.m.d.p.AnalyticsProcessManager] [Weis-MacBook-Pro.local] [wine_ul] Started analyzing
[2020-10-30T00:03:25,932][INFO ][o.e.x.m.d.p.AnalyticsProcessManager] [Weis-MacBook-Pro.local] [wine_ul] Waiting for result processor to complete
[2020-10-30T00:03:26,172][INFO ][o.e.x.m.p.AbstractNativeProcess] [Weis-MacBook-Pro.local] [wine_ul] State output finished
[2020-10-30T00:03:26,172][ERROR][o.e.x.m.p.AbstractNativeProcess] [Weis-MacBook-Pro.local] [wine_ul] analytics process stopped unexpectedly:
[2020-10-30T00:03:26,173][ERROR][o.e.x.m.p.l.CppLogMessageHandler] [Weis-MacBook-Pro.local] [controller/5526] [CDetachedProcessSpawner.cc@184] Child process with PID 9510 was terminated by signal 11
[2020-10-30T00:03:26,173][ERROR][o.e.x.m.p.l.CppLogMessageHandler] [Weis-MacBook-Pro.local] [controller/5526] [CDetachedProcessSpawner.cc@99] Will not attempt to kill process 9510: not a child process
[2020-10-30T00:03:26,173][ERROR][o.e.x.m.p.l.CppLogMessageHandler] [Weis-MacBook-Pro.local] [controller/5526] [CCommandProcessor.cc@115] Failed to kill process with PID 9510 in command with ID 524
[2020-10-30T00:03:26,174][ERROR][o.e.x.m.d.p.AnalyticsProcessManager] [Weis-MacBook-Pro.local] [wine_ul] Failed to kill process
java.io.IOException: ML controller failed to execute command [524]: [Failed to kill process with PID 9510]

Job configure:

{
  "id": "wine_ul",
  "create_time": 1604030604645,
  "version": "8.0.0",
  "description": "",
  "source": {
    "index": [
      "wine_ul"
    ],
    "query": {
      "match_all": {}
    }
  },
  "dest": {
    "index": "dest_wine_ul",
    "results_field": "ml"
  },
  "analysis": {
    "regression": {
      "dependent_variable": "quality",
      "num_top_feature_importance_values": 0,
      "prediction_field_name": "quality_prediction",
      "training_percent": 95,
      "randomize_seed": 95,
      "loss_function": "mse"
    }
  },
  "analyzed_fields": {
    "includes": [
      "alcohol",
      "chlorides",
      "citric acid",
      "density",
      "fixed acidity",
      "free sulfur dioxide",
      "pH",
      "quality",
      "residual sugar",
      "sulphates",
      "total sulfur dioxide",
      "volatile acidity"
    ],
    "excludes": []
  },
  "model_memory_limit": "125mb",
  "allow_lazy_start": false,
  "max_num_threads": 1
}

dataset csv file can be found here:
wine_quality_white_ul.csv.zip

Step to reproduce:

• use data viz to import wine_quality_white_ul.csv file
• manual update the quality mapping to unsigned_long
• create a DFA job by using the above job configure
• Run job

[droberts195]

if dependent_variable is unsigned_long

I strongly suspect that is a misleading statement. In the test data all the values are huge and close enough that within the accuracy of what float can represent they're all the same. I bet if your unsigned long values had been 1, 2, 3, 4, 5, 6 and 7 then this would not have happened.

This bug is far more likely to be "DFA regression job failed in coarse_parameter_search phase if all values of dependent_variable are the same when cast to float".