Add option for polygon/linestring in results

[red-fenix]

At the moment, photon only returns the point of a location, and not the polygon (see #259). This PR will add the option to add the polygon (i.e. geometry) to the Elasticsearch Index and a API parameter polygon to return said polygon. If no polygon exists, the point is returned.

WARNING: This will increase the Elasticsearch Index size! (~575GB for a Planet import).

To enable: add the command line argument -use-geometry-column whilst importing and add &polygon=true to the API call.

[lonvia]

I haven't done a full review yet but I do have some general thoughts on the implementation:

• This really needs to be implemented for the OpenSearch version because the ES version of Photon is on its way out. Note that the OpenSearch variant does not use mappings.json. It defines its mapping in https://github.com/komoot/photon/blob/master/app/opensearch/src/main/java/de/komoot/photon/opensearch/IndexMapping.java.
• As long as the geometry isn't used for lookup, indexes should be disabled on the new field to save a bit of disk space.
• We already have a field extent which contains the bounding box. The geometry should replace this field, i.e. once full geometries are enabled, do not save extent and derive the extent from the geometry field when returning a result. We need to keep the centroid because it is not necessarily the geometric centroid of the geometry. (Note that extent was missing from the mapping specification so far. So likely it was in fact saved as a text field, which really is an oversight but can't really be fixed right now without creating an incompatible database version.)
• This would be the second optional database feature after add support for structured queries (opensearch only) #815. Before we become too prolific with command-line arguments, I'd lean towards adding a single parameter -extra-db-features which takes a list of features (right now: structured, geometries).
• I agree on the introduction of the polygon parameter but would prefer to disable it when the extended geometries are not available in the database. Otherwise there will be an endless stream of bug reports on photon.komoot.io, why the results return a point instead of polygon. You can save the state of the feature in the property table and load it from there on start, see add support for structured queries (opensearch only) #815 for an example on how to do it. This property also comes in handy during updates of the database. It would be useful to have exactly the same behaviour as on import then.
• This needs some tests for the import.

Two other considerations come in mind but they are easily deferred to follow-up PRs:

• Once we have full geometries, we'd want to use them for reverse lookup. See Discussion for PR to improve accuracy of reverse geocoding #357.
• It might be worth to slightly simplify the geometries before importing them, or at least make that an option. Nominatim always keeps the original OSM geometries which sometimes can have a lot more support points than necessary. Simplification might help to further reduce database size.

[red-fenix]

I haven't done a full review yet but I do have some general thoughts on the implementation:

• This really needs to be implemented for the OpenSearch version because the ES version of Photon is on its way out. Note that the OpenSearch variant does not use mappings.json. It defines its mapping in https://github.com/komoot/photon/blob/master/app/opensearch/src/main/java/de/komoot/photon/opensearch/IndexMapping.java.

Thanks. I will update the PR with the changes in this file soon.

One question though about the 'Elasticsearch is on it's way out': I've been planning to update the Elastic client to the Java API so you can use an existing Elasticsearch cluster instead of the internal one (newer versions of Elasticsearch don't support the Transport client). Is my new PR still a good idea?

• As long as the geometry isn't used for lookup, indexes should be disabled on the new field to save a bit of disk space.

Will take this along as well.

• This would be the second optional database feature after add support for structured queries (opensearch only) #815. Before we become too prolific with command-line arguments, I'd lean towards adding a single parameter -extra-db-features which takes a list of features (right now: structured, geometries).

Agreed

• I agree on the introduction of the polygon parameter but would prefer to disable it when the extended geometries are not available in the database. Otherwise there will be an endless stream of bug reports on photon.komoot.io, why the results return a point instead of polygon. You can save the state of the feature in the property table and load it from there on start, see add support for structured queries (opensearch only) #815 for an example on how to do it. This property also comes in handy during updates of the database. It would be useful to have exactly the same behavior as on import then.

OK. I will make the default to return the polygon when it's available in the index. The option 'polygon=false' will return the centroid instead.

• This needs some tests for the import.

Will do.

Two other considerations come in mind but they are easily deferred to follow-up PRs:

• Once we have full geometries, we'd want to use them for reverse lookup. See Discussion for PR to improve accuracy of reverse geocoding #357.

I have to look into this issue.

• It might be worth to slightly simplify the geometries before importing them, or at least make that an option. Nominatim always keeps the original OSM geometries which sometimes can have a lot more support points than necessary. Simplification might help to further reduce database size.

I will look into this.
Another thing related to this: when someone is searching for a street Nominatim (and thus Photon) returns a street in parts because they are separate OSM id's. I'm still looking for a method to merge multiple ways (i.e. linestrings) into 1 linestring to make sure the whole street is displayed instead of a part (example)

[lonvia]

One question though about the 'Elasticsearch is on it's way out': I've been planning to update the Elastic client to the Java API so you can use an existing Elasticsearch cluster instead of the internal one (newer versions of Elasticsearch don't support the Transport client). Is my new PR still a good idea?

We'll drop ES support completely and go with OpenSearch. Note that the OS version already supports an external OpenSearch cluster. The support is just somewhat rudimentary and HTTP-only.

[karussell]

Would this PR return a Multipolygon in case for queries like "hamburg"? See the nominatim response.

Currently photon returns a slightly confusing, but correct extent as only a single extent is allowed https://photon.komoot.io/api/?q=hamburg ... or maybe we add a new field extents in the response?

[red-fenix]

Would this PR return a Multipolygon in case for queries like "hamburg"? See the nominatim response.

Currently photon returns a slightly confusing, but correct extent as only a single extent is allowed https://photon.komoot.io/api/?q=hamburg ... or maybe we add a new field extents in the response?

Yes, it would return a Polygon like Nominatim does.

[red-fenix]

I haven't done a full review yet but I do have some general thoughts on the implementation:

• This really needs to be implemented for the OpenSearch version because the ES version of Photon is on its way out. Note that the OpenSearch variant does not use mappings.json. It defines its mapping in https://github.com/komoot/photon/blob/master/app/opensearch/src/main/java/de/komoot/photon/opensearch/IndexMapping.java.

Done

• I agree on the introduction of the polygon parameter but would prefer to disable it when the extended geometries are not available in the database. Otherwise there will be an endless stream of bug reports on photon.komoot.io, why the results return a point instead of polygon. You can save the state of the feature in the property table and load it from there on start, see add support for structured queries (opensearch only) #815 for an example on how to do it. This property also comes in handy during updates of the database. It would be useful to have exactly the same behaviour as on import then.

Done

• This needs some tests for the import.

Done

I'm looking forward to your response :)

[lonvia]

Sorry for the long delay in responding.

I've decided to get a release out first before looking into this again. Because of this there is a minor merge conflict now regarding the dependencies. Also, the opensearch variant doesn't compile because setUpESWithPolygons() is only defined for the elasticsearch variant. Could you quickly fix both issues?

[lonvia]

Couple of more comments after trying this out. The OpenSearch version currently doesn't work at all. It already fails to import.

The output for the ElasticSearch version displays a wrong projection:

...
"geometry": {
  "coordinates":[9.5227103,47.1395576],
  "type":"Point",
  "crs":{"type":"name","properties":{"name":"EPSG:0"}}
}
...

Photon doesn't supply a CRS at all with the geometry. I'd leave it at that.

[red-fenix]

Sorry for the long delay in responding.

I've decided to get a release out first before looking into this again. Because of this there is a minor merge conflict now regarding the dependencies. Also, the opensearch variant doesn't compile because setUpESWithPolygons() is only defined for the elasticsearch variant. Could you quickly fix both issues?

Will do this weekend.

[red-fenix]

Couple of more comments after trying this out. The OpenSearch version currently doesn't work at all. It already fails to import.

I've made some adjustments and have been trying to get the OpenSearch version running however, it fails with an error:

Request failed: [mapper_parsing_exception] No handler for type [geo_shape] declared on field [geometry]

I've searching online, but can't seem to find the reason for OpenSearch clearly has geo_shape support.

In other news: I fixed the PR with your suggestions, but will perform some additional tests this weekend.

[lonvia]

Request failed: [mapper_parsing_exception] No handler for type [geo_shape] declared on field [geometry]

That's where I gave up my experiments with this branch as well. It's possible that the embedded version of OpenSearch that we are using is missing a dependency needed for geo_shape. Maybe try running against a stand-alone OpenSearch and see if it works then?

[otbutz]

https://github.com/opensearch-project/geospatial ? Maybe it's enough to add the OpenSearch plugin as a dependency.