[Bug]: [new_indexes] The searched id after group by is not equal with the top1 id searched using the expression of "group_by_field_name==each unique value" when creating HNSW_FLAT/SQ/PQ/PRQ on float16/ bfloat16 vector

[binbinlv]

Is there an existing issue for this?

• I have searched the existing issues

Environment

- Milvus version:master-20241024-0dbf9482-amd64
- Deployment mode(standalone or cluster):all
- MQ type(rocksmq, pulsar or kafka):    both
- SDK version(e.g. pymilvus v2.0.0rc2): 2.5.0rc99
- OS(Ubuntu or CentOS): 
- CPU/Memory: 
- GPU: 
- Others:

Current Behavior

The searched id after group by is not equal with the top1 id searched using the expression of "group_by_field_name==each unique value" when creating HNSW_FLAT/SQ/PQ/PRQ on float16/ bfloat16 vector

[2024-10-25 12:10:00 - INFO - ci_test]: INT64 on FLOAT16_VECTOR dismatch_item, top1_grpby_dis: 33.54237365722656, top1_expr_dis: 14.902473449707031 (test_mix_scenes.py:2746)
[2024-10-25 12:10:00 - INFO - ci_test]: INT64 on FLOAT16_VECTOR  top1_dismatch_num: 15, results_num: 15, dismatch_rate: 1.0 (test_mix_scenes.py:2747)

Expected Behavior

The searched id after group by is always equal with the top1 id searched using the expression of "group_by_field_name==each unique value" when creating HNSW_FLAT/SQ/PQ/PRQ on float16/ bfloat16 vector

Steps To Reproduce

    @pytest.mark.tags(CaseLabel.L2)
    # @pytest.mark.parametrize("support_field", [DataType.INT8.name,  DataType.INT64.name,
    #                                            DataType.BOOL.name, DataType.VARCHAR.name])
    @pytest.mark.parametrize("support_field", [DataType.INT64.name])
    def test_search_group_by_supported_scalars_new_hnsw_index(self, support_field):
        """
        verify search group by works with supported scalar fields
        """
        nq = 1
        limit = 15
        for j in range(len(self.vector_fields)):
            j = 0
            search_vectors = cf.gen_vectors(nq, dim=self.dims[j], vector_data_type=self.vector_fields[j])
            search_params = {"params": cf.get_search_params_params(self.index_types[j])}
            res1 = self.collection_wrap.search(data=search_vectors, anns_field=self.vector_fields[j],
                                               param=search_params, limit=limit,
                                               group_by_field=support_field,
                                               output_fields=[support_field])[0]
            for i in range(nq):
                grpby_values = []
                dismatch = 0
                results_num = 2 if support_field == DataType.BOOL.name else limit
                for l in range(results_num):
                    top1 = res1[i][l]
                    top1_grpby_pk = top1.id
                    top1_grpby_value = top1.fields.get(support_field)
                    expr = f"{support_field}=={top1_grpby_value}"
                    if support_field == DataType.VARCHAR.name:
                        expr = f"{support_field}=='{top1_grpby_value}'"
                    grpby_values.append(top1_grpby_value)
                    res_tmp = self.collection_wrap.search(data=[search_vectors[i]], anns_field=self.vector_fields[j],
                                                          param=search_params, limit=1, expr=expr,
                                                          output_fields=[support_field])[0]
                    top1_expr_pk = res_tmp[0][0].id
                    if top1_grpby_pk != top1_expr_pk:
                        dismatch += 1
                        log.info(f"{support_field} on {self.vector_fields[j]} dismatch_item, top1_grpby_dis: {top1.distance}, top1_expr_dis: {res_tmp[0][0].distance}")
                log.info(f"{support_field} on {self.vector_fields[j]}  top1_dismatch_num: {dismatch}, results_num: {results_num}, dismatch_rate: {dismatch / results_num}")
                baseline = 1 if support_field == DataType.BOOL.name else 0.2    # skip baseline check for boolean
                assert dismatch / results_num <= baseline
                # verify no dup values of the group_by_field in results
                assert len(grpby_values) == len(set(grpby_values))

Milvus Log

https://grafana-4am.zilliz.cc/explore?orgId=1&panes=%7B%226mJ%22:%7B%22datasource%22:%22vhI6Vw67k%22,%22queries%22:%5B%7B%22refId%22:%22A%22,%22expr%22:%22%7Bcluster%3D%5C%22devops%5C%22,namespace%3D%5C%22chaos-testing%5C%22,pod%3D~%5C%22test-null-new-1-kwwuq.%2A%5C%22%7D%22,%22datasource%22:%7B%22type%22:%22loki%22,%22uid%22:%22vhI6Vw67k%22%7D%7D%5D,%22range%22:%7B%22from%22:%22now-1h%22,%22to%22:%22now%22%7D%7D%7D&schemaVersion=1

Anything else?

collection_name : "TestGroupSearchNewHNSWIndex_badYC7EF"

[binbinlv]

This issue not exists for FLOAT_VECTOR vector.

[alexanderguzhva]

@binbinlv zilliztech/knowhere#913 should fix the problem

[xiaofan-luan]

/assign @binbinlv

[binbinlv]

working on verification.

[binbinlv]

Verified and fixed in master branch.
pymilvus: 2.5.0rc106
milvus: master-20241101-0ac8b166

[binbinlv]

But this issue still exists in 2.5 branch:

milvus: 2.5-20241101-116bf501-amd64

results:

[2024-11-02 01:19:28 - INFO - ci_test]: VARCHAR on BFLOAT16_VECTOR dismatch_item, top1_grpby_dis: 0.6354575157165527, top1_expr_dis: 0.7928284406661987 (test_mix_scenes.py:2775)
[2024-11-02 01:19:28 - INFO - ci_test]: VARCHAR on BFLOAT16_VECTOR  top1_dismatch_num: 15, results_num: 15, dismatch_rate: 1.0 (test_mix_scenes.py:2776)

[alexanderguzhva]

@binbinlv different knowhere version, maybe?

[binbinlv]

The 2.5 branch is deleted and master branch has been verified, so close this issue.