Use the correct type to widen the sort fields when merging top field docs

[bugmakerrrrrr]

Describe the bug

To enable numeric sort optimization support for all numeric types, we use double to widen the sort fields in #6424. However, when the sort value exceeds the MAX_SAFE_INTEGER (2^53 – 1), it may result in incorrect sorting.

Related component

Search

To Reproduce

public class FieldSortIT extends ParameterizedDynamicSettingsOpenSearchIntegTestCase {
    public void testSortMixedNumericField() throws Exception {
            internalCluster().ensureAtLeastNumDataNodes(3);
            index("long", Long.MAX_VALUE);
            index("integer", Integer.MAX_VALUE);
            SearchResponse searchResponse = client().prepareSearch("long", "integer")
                .setQuery(matchAllQuery())
                .setSize(10)
                .addSort(SortBuilders.fieldSort("field").order(SortOrder.ASC).sortMode(SortMode.MAX))
                .get();
            assertNoFailures(searchResponse);
            long[] sortValues = new long[10];
            for (int i = 0; i < 10; i++) {
                sortValues[i] = ((Number) searchResponse.getHits().getAt(i).getSortValues()[0]).longValue();
            }
            for (int i = 1; i < 10; i++) {
                assertThat(Arrays.toString(sortValues), sortValues[i - 1], lessThan(sortValues[i]));
            }
        }
    
        private void index(String type, long end) throws Exception {
            assertAcked(
                prepareCreate(type).setMapping(
                    XContentFactory.jsonBuilder()
                        .startObject()
                        .startObject("properties")
                        .startObject("field")
                        .field("type", type)
                        .endObject()
                        .endObject()
                        .endObject()
                ).setSettings(Settings.builder().put("index.number_of_shards", 3).put("index.number_of_replicas", 0))
            );
            ensureGreen(type);
            for (int i = 0; i < 5; i++) {
                client().prepareIndex(type).setId(Integer.toString(i)).setSource("{\"field\" : " + (end - i) + " }", XContentType.JSON).get();
            }
            client().admin().indices().prepareRefresh(type).get();
        }
}

Expected behavior

When the sort values are all integer values, we can use the Long type to widen the sort fields.

To be discussed: 1. When sort values contain both integer values and float values, should we throw an exception or use double to widen sort fields? 2. Should we check the compatibility between sort types when merging top docs?

Additional Details

No response

[reta]

@bugmakerrrrrr I believe this is listed as the limitation of this data type and is explicitly documented [1]

[1] https://opensearch.org/docs/latest/field-types/supported-field-types/unsigned-long/#limitations

[bugmakerrrrrr]

@bugmakerrrrrr I believe this is listed as the limitation of this data type and is explicitly documented [1]

[1] https://opensearch.org/docs/latest/field-types/supported-field-types/unsigned-long/#limitations

@reta Thanks for the update. Actually, this is not only related to unsigned_long. The example I gave above only contains long and integer. Also, as I explained in this comment, the results can be wrong even when sorting only for unsigned_long fields in a single index.