Adding json field in output_fields is slowing search search by 20%.

[Rajat376]

On a Collection I was searching with and without a scalar json field 'data' in output_fields.
it was reducting the search by 20%.

Queries were
client.search('collection',texts[x:x+1000],limit=10,output_fields=['data'], anns_field='sparse',
search_params={
'params': {'drop_ratio_search': 0},
})
client.search('collection',texts[x:x+1000],limit=10,output_fields=[], anns_field='sparse',
search_params={
'params': {'drop_ratio_search': 0},
})

Any idea why this is happening?
And are there any configs/changes which can improve search with 'data' field.

[yhmo]

It is expected that output_fileds will increase the latency of search/query.
With output_fields=['data'], the query node needs to read the data from memory or mmap file. With more data returned, the network latency also increases.

[Rajat376]

but including it also reduces the throughput by similar amount.
That shouldn't

[xiaofan-luan]

when you output field this could introduce extra cpus, or disk ios since reverse data from columar storage is expected to be slow 。
it is also related to the topk you are asking for.

did you have a profiling result or any performance numbers? I think 20% makes sense for me but right now it's hard for me to recommend how you can improve