Synthetic source should not load fields that are not required

[iverase]

The current implementation synthesise the full document regardless if the value is required on the final response doing unnecessary work that can be very expensive on slow disk.

Consider the following example:

public void testSyntheticSourceAllFields() throws Exception {

        XContentBuilder indexSpec = XContentBuilder.builder(XContentType.JSON.xContent()).startObject();
        {
            indexSpec.startObject("_source").field("mode", "synthetic").endObject();
            indexSpec.startObject("properties")
                .startObject("kwd1").field("type", "keyword").endObject()
                .startObject("kwd2").field("type", "keyword").endObject()
                .startObject("kwd3").field("type", "keyword").endObject()
                .startObject("kwd4").field("type", "keyword").endObject()
            .endObject();
        }
        assertAcked(admin().indices().prepareCreate("synthetic").setMapping(indexSpec.endObject()).get());
        indexRandom(true, client().prepareIndex("synthetic").setSource("kwd1", "val1", "kwd2", "val2", "kwd3", "val3", "kwd4", "val4"));
        SearchResponse response = client().prepareSearch("synthetic").addFetchField("kwd3").setFetchSource(false).get();
    }

The query is not asking for source and only requesting the value of one of the fields. Still in this case we synthesise source with all the fields, accessing 4 doc values. We should be able to read just the requested value in this case.

[elasticsearchmachine]

Pinging @elastic/es-analytics-geo (Team:Analytics)

[martijnvg]

Note that the source is completely synthesized if source filtering is active. For example in this case:

public void testSyntheticSourcePartialSource() throws Exception {
        XContentBuilder indexSpec = XContentBuilder.builder(XContentType.JSON.xContent()).startObject();
        //indexSpec.startObject("mappings");
        {
            indexSpec.startObject("_source").field("mode", "synthetic").endObject();
            indexSpec.startObject("properties")
                .startObject("kwd1").field("type", "keyword").endObject()
                .startObject("kwd2").field("type", "keyword").endObject()
                .startObject("kwd3").field("type", "keyword").endObject()
                .startObject("kwd4").field("type", "keyword").endObject()
                .endObject();
        }
        assertAcked(admin().indices().prepareCreate("synthetic").setMapping(indexSpec.endObject()).get());
        indexRandom(true, client().prepareIndex("synthetic").setSource("kwd1", "kwd1", "kwd2", "kwd2", "kwd3", "kwd3", "kwd4", "kwd4"));

        SearchResponse response = client().prepareSearch("synthetic").setFetchSource("kwd3", null).get();
        logger.info("response=" + response);
    }

Ideally only fields that match with the source filtering should be loaded.

[elasticsearchmachine]

Pinging @elastic/es-storage-engine (Team:StorageEngine)