TFIDF model very slow to Apply?

[ericchaucl]

Hi,

I've successfully created a TFIDF model with 100 features.

If I apply the model to 4000 records, it takes ~7s to show the result. That's pretty good. The problem is, it also takes ~7s if I just apply it to 1 single record!

We plan to use the TFIDF model to provide nearly instant feedback to enquiry of new input short text (say, 20 words). Appreciate for any suggestions to improve this, say, system settings or some workaround, thanks!

P.S. I assume that is not server hardware concern as it runs satisfactorily with large # of records.

[ericchaucl · Accepted solution]

If just apply one or few records in a ML model, can disable the optimzation using "| noop search_optimization=false" and will perform significantly faster. In my cases, that reduces from 7s to 1s.

[ericchaucl]

Some more info, when looking into the Job Inspection, here are the steps that takes most of the time, no matter I apply the model into 1 record or 4000 records:

1.74s - dispatch.evaluate
1.74 s - dispatch.evaluate.apply
1.74 s - dispatch.optimize.FinalEval
1.76 s - dispatch.optimize.matchReportAcceleration
1.77 s - dispatch.optimize.toJson

[ericchaucl]

If just apply one or few records in a ML model, can disable the optimzation using "| noop search_optimization=false" and will perform significantly faster. In my cases, that reduces from 7s to 1s.

[ericchaucl]

Appreciate if any suggestions in improving the performance? Or I should give up expecting Splunk to give responsive results when using ML model?