Using `"lang":"painless"` in the _msearch has poor performance

[masterwang22327]

Is your feature request related to a problem? Please describe

I have a multi-index and multi-field knn vector search scenario, and the data volume reaches millions. In order to perform full-text search, I plan to use the "_msearch" interface to implement this function in one request. After studying the official documentation, I found that only the "script_score" request method is supported. In the request body I used "lang":"painless". Below is my detailed request body example, a search scenario with three indexes and nine fields. When I used this request body to query opensearch, the cashback response time and QPS were very low, so I gradually increased the vector query volume for comparative experiments. When I conducted the stress test experiment, some configurations were as follows.

Describe the solution you'd like

I want to achieve a corresponding performance of 50QPS and 3 seconds in a multi-index and multi-vector search scenario. How should I optimize the hardware configuration or optimize the query body?

Related component

Search

Describe alternatives you've considered

How should I optimize the hardware configuration or optimize the query body?

Additional context

No response

[peternied]

[Triage - attendees 1 2 3] @masterwang22327 Thanks for filing this issue. Would it be possible to get a text of the query to aid in analyzing the issue. Making a pull request would be useful to help diagnose as well.

[masterwang22327]

hello peternied. my text of the query is as followed。

[peternied]

@masterwang22327 Thanks, I wasn't clear: please copy/paste the full text of request/response as opposed to a screenshot of text.

[masterwang22327]

@peternied It was delayed for a while due to technical issues.I put the request body and return body into two txt files respectively. git_embedding_msearch_query_body.txt git_embedding_msearch_response_body.txt