upsert not woking correctly

[vihariazure]

hi team , i am testing upsert functionality

i took 1k records

created collection
inserted data
-> flush
->created hnsw index

i used describe the index it shows 1k total_rows , indexed_rows and pending_index_rows as zero

now i loaded the collection
the same 1k records using upsert it output ({'upsert_count': 1000}
flush
when i describe the index it shows 2k total_rows , indexed_rows and pending_index_rows as zero

similarly i run it for another 3 times. every time total_rows , indexed_rows is increasing by 1k

python sdk version 2.5.4
milvus version 2.5.4

then i used this
res = self.client.query(collection_name=self.collection_name, consistency_level="Strong", output_fields=["count()"])
print("count() returns row count", res)
output :-
count() returns row count data: ["{'count()': 1000}"]

what is wrong in it , iam inserting the same data using insert first time and 2-5th times as upsert . ideally it has to 1k only even after multiple upsets

why the values are increasing total_rows , indexed_rows

[xiaofan-luan]

because indexed row and pending indexed row is meta data.
it didn't know whether the data is deleted or not, until compaction happened.

On the other side, count acutally read the data, so it is accurate.

[xiaofan-luan]

all the delete, upsert for milvus is soft delete. and data won't be cleaned instantly

[vihariazure]

@xiaofan-luan

thanks for the reply , what is the SLA for upsert to reflect. which parameter is responsible for it.
to get instantly results what are the ways and parameters need to configured

[xiaofan-luan]

what do you mean by SLA?
if you use strong consistency then upsert will be applied instantly, but search/query will be relatively slow.
see document here https://milvus.io/docs/consistency.md