Retrieve Offsets of inner_hits

[JKatzwinkel]

• Elasticsearch: 8.6.2
• spring-data-elasticsearch: 5.1.0-M3

Hi, thanks again for the excellent work, I enjoy it very much!

I'm having an issue while trying to obtain the correct offset values from inner_hits of a nested query via SearchHit#getNestedMetaData. Even though Elasticsearch responds with the expected offsets when I run my queries directly against it, the #getOffset() return values from inner hits via #getInnerHits() are always the same, regardless of how many inner hits are returned within an individual top-level hit.

Allow me to illustrate this based on your integration test(s). They run a (nested) nested query and test the offsets of the inner hits, of which there are 1 on each nesting level:

		Inhabitant john = new Inhabitant("John", "Smith");
		Inhabitant carla = new Inhabitant("Carla", "Miller");
		House cornerHouse = new House("Round the corner", "7", Arrays.asList(john, carla));
		City metropole = new City("Metropole", Arrays.asList(cornerHouse));

		Inhabitant jack = new Inhabitant("Jack", "Wayne");
		Inhabitant emmy = new Inhabitant("Emmy", "Stone");
		House mainStreet = new House("Main Street", "42", Arrays.asList(jack, emmy));
		City village = new City("Village", Arrays.asList(mainStreet));

		operations.save(Arrays.asList(metropole, village));
	}

	@Test
	@Order(java.lang.Integer.MAX_VALUE)
	void cleanup() {
		operations.indexOps(IndexCoordinates.of(indexNameProvider.getPrefix() + "*")).delete();
	}

	@Test
	void shouldReturnInnerHits() {

		Query query = buildQueryForInnerHits("inner_hit_name", "hou-ses.in-habi-tants", "hou-ses.in-habi-tants.first-name",
				"Carla");

		SoftAssertions softly = new SoftAssertions();
		SearchHits<City> searchHits = operations.search(query, City.class);

		softly.assertThat(searchHits.getTotalHits()).isEqualTo(1);

		SearchHit<City> searchHit = searchHits.getSearchHit(0);
		softly.assertThat(searchHit.getInnerHits()).hasSize(1);

		SearchHits<?> innerHits = searchHit.getInnerHits("inner_hit_name");
		softly.assertThat(innerHits).hasSize(1);

		SearchHit<?> innerHit = innerHits.getSearchHit(0);
		Object content = innerHit.getContent();
		assertThat(content).isInstanceOf(Inhabitant.class);
		Inhabitant inhabitant = (Inhabitant) content;
		softly.assertThat(inhabitant.getFirstName()).isEqualTo("Carla");
		softly.assertThat(inhabitant.getLastName()).isEqualTo("Miller");

		NestedMetaData nestedMetaData = innerHit.getNestedMetaData();
		softly.assertThat(nestedMetaData.getField()).isEqualTo("houses");
		softly.assertThat(nestedMetaData.getOffset()).isEqualTo(0); // <-- this seems right
		softly.assertThat(nestedMetaData.getChild().getField()).isEqualTo("inhabitants");
		softly.assertThat(nestedMetaData.getChild().getOffset()).isEqualTo(1);  // this makes sense too

		softly.assertAll();

However ☝, imagine they would look like this, now finding 2 inner hits that (as far as I understand) should have different offsets:

		Inhabitant john = new Inhabitant("John", "Smith");
		Inhabitant carla1 = new Inhabitant("Carla", "Miller");
		Inhabitant carla2 = new Inhabitant("Carla", "Nguyen");  // <-- second match/inner hit!!
		House cornerHouse = new House("Round the corner", "7", Arrays.asList(john, carla1, carla2));
		City metropole = new City("Metropole", Arrays.asList(cornerHouse));

		Inhabitant jack = new Inhabitant("Jack", "Wayne");
		Inhabitant emmy = new Inhabitant("Emmy", "Stone");
		House mainStreet = new House("Main Street", "42", Arrays.asList(jack, emmy));
		City village = new City("Village", Arrays.asList(mainStreet));

		operations.save(Arrays.asList(metropole, village));
	}

	@Test
	@Order(java.lang.Integer.MAX_VALUE)
	void cleanup() {
		operations.indexOps(IndexCoordinates.of(indexNameProvider.getPrefix() + "*")).delete();
	}

	private static void testInnerHit(
		SoftAssertions softly, SearchHit<?> innerHit,
		String firstName, String lastName,
		int nestedOffsetLvl1, int nestedOffsetLvl2
	) {
		Object content = innerHit.getContent();
		assertThat(content).isInstanceOf(Inhabitant.class);
		Inhabitant inhabitant = (Inhabitant) content;
		softly.assertThat(inhabitant.getFirstName()).isEqualTo(firstName);
		softly.assertThat(inhabitant.getLastName()).isEqualTo(lastName);

		NestedMetaData nestedMetaData = innerHit.getNestedMetaData();
		softly.assertThat(nestedMetaData.getField()).isEqualTo("houses");
		softly.assertThat(nestedMetaData.getOffset()).isEqualTo(nestedOffsetLvl1);
		softly.assertThat(nestedMetaData.getChild().getField()).isEqualTo("inhabitants");
		softly.assertThat(nestedMetaData.getChild().getOffset()).isEqualTo(nestedOffsetLvl2);
	}

	@Test
	void shouldReturnInnerHits() {

		Query query = buildQueryForInnerHits("inner_hit_name", "hou-ses.in-habi-tants", "hou-ses.in-habi-tants.first-name",
				"Carla");

		SoftAssertions softly = new SoftAssertions();
		SearchHits<City> searchHits = operations.search(query, City.class);

		softly.assertThat(searchHits.getTotalHits()).isEqualTo(1);

		SearchHit<City> searchHit = searchHits.getSearchHit(0);
		softly.assertThat(searchHit.getInnerHits()).hasSize(1);

		SearchHits<?> innerHits = searchHit.getInnerHits("inner_hit_name");
		softly.assertThat(innerHits).hasSize(2);

		testInnerHit(softly, innerHits.getSearchHit(0), "Carla", "Miller", 0, 1);
		testInnerHit(softly, innerHits.getSearchHit(1), "Carla", "Nguyen", 0, 2); // <-- the important bit!

		softly.assertAll();
	}

❌ Now the tests (./mvnw clean verify) fail because all of the inner hits have the same offset (1). This is despite of Elasticsearch itself returning the offsets I expect (1 and 2 in this case). I might be doing something wrong, but on the other hand the implementation responsible for mapping the inner hit offsets does contain something that confuses me (first and last line):

			NestedMetaData nestedMetaData = searchHits.getSearchHit(0).getContent().getNestedMetaData();
			ElasticsearchPersistentEntityWithNestedMetaData persistentEntityWithNestedMetaData = getPersistentEntity(
					mappingContext.getPersistentEntity(type), nestedMetaData);

			if (persistentEntityWithNestedMetaData.entity != null) {
				List<SearchHit<Object>> convertedSearchHits = new ArrayList<>();
				Class<?> targetType = persistentEntityWithNestedMetaData.entity.getType();

				// convert the list of SearchHit<SearchDocument> to list of SearchHit<Object>
				searchHits.getSearchHits().forEach(searchHit -> {
					SearchDocument searchDocument = searchHit.getContent();

					Object targetObject = converter.read(targetType, searchDocument);
					convertedSearchHits.add(new SearchHit<>(searchDocument.getIndex(), //
							searchDocument.getId(), //
							searchDocument.getRouting(), //
							searchDocument.getScore(), //
							searchDocument.getSortValues(), //
							searchDocument.getHighlightFields(), //
							searchHit.getInnerHits(), //
							persistentEntityWithNestedMetaData.nestedMetaData, // <-- this

So for each inner hit, the same nested matadata are being used (persistendEntityWithNestedMetaData.nestedMetaData), even though each inner inner hit should have its own metadata (_nested object with field and offset inside) in the Elasticsearch response.
There is probably a good reason for it to be implemented like this, but this seems to cause the unexpected behaviour I described above. I actually managed to fix this very easily by mapping the actual metadata of the actual inner hits respectively:

					convertedSearchHits.add(new SearchHit<>(searchDocument.getIndex(), //
							searchDocument.getId(), //
							searchDocument.getRouting(), //
							searchDocument.getScore(), //
							searchDocument.getSortValues(), //
							searchDocument.getHighlightFields(), //
							searchHit.getInnerHits(), //
							getPersistentEntity(
								mappingContext.getPersistentEntity(type),
								searchHit.getContent().getNestedMetaData()  // <-- 👀
							).nestedMetaData,
							searchHit.getExplanation(), //
							searchHit.getMatchedQueries(), //
							targetObject));
				});

✅ After this change, the tests pass again.

Am I right to assume that the way in which I amended your integration test(s) is how you are supposed to retrieve the inner hits offsets, or am I doing something wrong?

Thanks again and happy Easter!

[sothawo]

Thanks for finding, analyzing and providing a fix for this. Do you want to submit a Pull Request, otherwise I could do it as well, I copied and tested that code locally.

As for why this is so? Don't know. I probably did not seem to consider the offset when implementing it and so just missed it.

[JKatzwinkel]

thanks for the feedback and yes, I'm happy to prepare a pr I hope to get to it tomorrow!