Missing link between model and test data (if you don't store the predictions)

[nanounanue]

Running triage without storing the predictions (in CLI: --no-save-predictions flag) apparently "breaks" the connection between model and test matrix.

What do I mean by "breaks" is that (again, apparently) there is no way to see, in the DB in which matrix the model was tested.

Originally, one possible query to do that was:

select distinct
    model_id,
    model_group_id, train_end_time,
    model_hash,
    pr.matrix_uuid as test_matrix_uuid,
    ma.num_observations as observations,
    ma.lookback_duration as feature_lookback_duration,  ma.feature_start_time
from
    model_metadata.models as mo
    join
    test_results.predictions as pr using (model_id)
    join
    model_metadata.matrices as ma on pr.matrix_uuid = ma.matrix_uuid
order by
    model_group_id, train_end_time asc;

Any ideas?

[thcrock]

This brings up a good point. I think that the evaluation tables should have the matrix uuid just like the predictions table has.

As far as how to achieve this now, I think you could do a JSON join from specific fields in the matrix_metadata column in the matrices table and columns from the evaluation table.

[nanounanue]

I totally agree with you, Should I rename the issue?

[thcrock]

Sure, and add it to the project in the "let's do" column.