Polygons with rank_address = 0 are only used in search and (rarely)
for reverse lookup. Geometries do not need to be precise for that
because topology does not matter. OSM has some very large polygons
of natural features with sizes of more than 10MB. Simplify these
polygons to keep the database and indexes smaller.
Use JSON arrays which can have mixed types and therefore have
a more logical structure than separate arrays. Avoid JSON dicts
because of their verboseness.
Before postgis 3.4 ST_Project required a geography as input and seemed
to have implicitly converted to geography. Since 3.4 geometry input
is supported but leads to a completely different result.
So far the SQL logic used the information from the address field
to determine if an address is attached to a street or place.
This changes the logic to use the information provided in the
token_info. This allows sanitizers to enforce a certain parenting
without changing the visible address information.
Snapping a line to a point before splitting was meant to ensure
that the split point is really on the line. However, ST_Snap() does
not always behave well for this case. It may shorten the interpolation
line in some cases with the result that two points housenumbers
suddenly fall on the same point. It might also shorten the line down
to a single point which then makes ST_Split() crash.
Switch to a combination of ST_LineLocatePoint and ST_LineSubString
instead, which guarantees to keep the original geometry. Explicitly
handle the corner cases, where the split point falls on the beginning
or end of the line.
Add a special index that contains the place nodes buffered by their
respective area according to their search rank. This replaces the
maximum area search for place nodes and reduces drastically the number
of place nodes that need to be retrieved.
When a associatedStreet relation has multiple street members
always take the closest one. Avoid geometry operations for
the frequent case that there is only one street.
This makes Nominatim compatible with osm2pgsql's default update
modus operandi of deleting and reinserting data. Deletes are diverted
into a TODO table instead of executing them. When data is reinserted,
the corresponding entry in the TODO table is deleted. After updates are
finished, the remaining entries in the TODO table are executed, doing
the same work as the delete trigger did before.
The new behaviour also works against the gazetteer output with its
insert-only mechanism.
The values in the raster are already normalized between 0 and 2**16,
so a simple conversion to [0, 1] will do.
Check for existance of secondary_importance table statically when
creating the SQL function. For that to work importance tables need
to be created before the functions.
