Fix infer_trained_model API

[pquentin]

It can return list of lists of integers and floats.

[davidkyle]

I can't find any places in the Elasticsearch code where an array of doubles or ints is returned. Do you have an example pls?

Why isn't the list case covered by this line in the spec? Is the result a lists of lists of numbers?

elasticsearch-specification/specification/ml/_types/inference.ts

Line 484 in 7fc0896

predicted_value?: PredictedValue | PredictedValue[]

[pquentin]

@davidkyle Sorry that I missed your comment. It happens in four "3rd party deployment" ML YAML tests, for example https://github.com/elastic/elasticsearch/blob/0d38528e0e6f01010ba4a6d11d84ff7a5ddf32f4/x-pack/plugin/src/yamlRestTest/resources/rest-api-spec/test/ml/3rd_party_deployment.yml#L600-L622. In all cases, the following JSON is returned:

{
  "inference_results": [
    {
      "predicted_value": [
        [
          1,
          1
        ]
      ]
    }
  ]
}

Agreed that if predicted_value can be a list already, that should be covered somehow.

[pquentin]

Actually yes it is a list of lists, so there's a missing list somewhere, PredictedValue[] isn't enough. Reopening.

[github-actions]

Following you can find the validation results for the APIs you have changed.

API	Status	Request	Response
ml.clear_trained_model_deployment_cache	🟢	1/1	1/1
ml.close_job	🟢	64/64	63/63
ml.delete_calendar_event	🟢	4/4	4/4
ml.delete_calendar_job	🟢	3/3	3/3
ml.delete_calendar	🟢	5/5	5/5
ml.delete_data_frame_analytics	🟢	2/2	2/2
ml.delete_datafeed	🟢	3/3	3/3
ml.delete_expired_data	🟢	5/5	5/5
ml.delete_filter	🟢	27/27	27/27
ml.delete_forecast	🟢	3/3	3/3
ml.delete_job	🟢	47/47	47/47
ml.delete_model_snapshot	🟢	2/2	2/2
ml.delete_trained_model_alias	🟢	3/3	3/3
ml.delete_trained_model	🟢	5/5	5/5
ml.estimate_model_memory	🟢	16/16	16/16
ml.evaluate_data_frame	🟢	15/15	15/15
ml.explain_data_frame_analytics	🟢	7/7	7/7
ml.flush_job	🟢	16/16	16/16
ml.forecast	🟢	2/2	2/2
ml.get_buckets	🟢	14/14	14/14
ml.get_calendar_events	🟢	17/17	17/17
ml.get_calendars	🟢	17/17	17/17
ml.get_categories	🟢	12/12	12/12
ml.get_data_frame_analytics_stats	🟢	12/12	12/12
ml.get_data_frame_analytics	🔴	17/17	14/17
ml.get_datafeed_stats	🟢	27/27	27/27
ml.get_datafeeds	🔴	20/20	9/20
ml.get_filters	🟢	13/13	13/13
ml.get_influencers	🟢	11/11	11/11
ml.get_job_stats	🔴	32/32	30/32
ml.get_jobs	🔴	31/31	22/31
ml.get_memory_stats	🟢	6/6	6/6
ml.get_model_snapshot_upgrade_stats	🟢	3/3	3/3
ml.get_model_snapshots	🟢	18/18	18/18
ml.get_overall_buckets	🟢	16/16	15/15
ml.get_records	🟢	8/8	8/8
ml.get_trained_models_stats	🔴	17/17	10/17
ml.get_trained_models	🔴	36/37	32/37
ml.infer_trained_model	🟢	10/10	10/10
ml.info	🔴	10/10	2/10
ml.open_job	🟢	83/83	83/83
ml.post_calendar_events	🟢	21/21	21/21
ml.post_data	🔴	10/12	15/19
ml.preview_data_frame_analytics	🟢	3/3	3/3
ml.preview_datafeed	🟢	17/17	17/17
ml.put_calendar_job	🟢	12/12	12/12
ml.put_calendar	🟢	135/135	135/135
ml.put_data_frame_analytics	🔴	32/33	32/33
ml.put_datafeed	🔴	70/71	53/71
ml.put_filter	🟢	27/27	27/27
ml.put_job	🔴	219/227	223/225
ml.put_trained_model_alias	🟢	13/13	13/13
ml.put_trained_model_definition_part	🟢	1/1	1/1
ml.put_trained_model_vocabulary	🟢	1/1	1/1
ml.put_trained_model	🔴	15/16	6/16
ml.reset_job	🟢	2/2	2/2
ml.revert_model_snapshot	🟢	2/2	2/2
ml.set_upgrade_mode	🟢	6/6	6/6
ml.start_data_frame_analytics	🟢	1/1	1/1
ml.start_datafeed	🟢	24/24	24/24
ml.start_trained_model_deployment	🔴	14/14	4/14
ml.stop_data_frame_analytics	🟢	5/5	5/5
ml.stop_datafeed	🟢	17/17	17/17
ml.stop_trained_model_deployment	🟢	10/10	10/10
ml.update_data_frame_analytics	🟢	2/2	2/2
ml.update_datafeed	🔴	7/7	2/7
ml.update_filter	🟢	3/3	3/3
ml.update_job	🔴	4/5	5/5
ml.update_model_snapshot	🟢	3/3	3/3
ml.update_trained_model_deployment	🔴	4/4	2/4
ml.upgrade_job_snapshot	🟢	3/3	3/3
ml.validate_detector	🟢	2/2	2/2
ml.validate	🟢	3/3	3/3

You can validate these APIs yourself by using the make validate target.

[davidkyle]

LGTM

Thanks @pquentin, the yml test you linked to uses the special pass_through type that returns the raw response. I'd forgotten about that one sorry

[pquentin]

Merging, assuming that only ints and floats can be lists of lists. Happy to open another PR if that's not the case.

Thank you!

[davidkyle]

Yes only ints and floats