* All functions that access database must be declared as `STABLE`, not `IMMUTABLE` -- because database can change at any moment, e.g. during an update
* there are a few functions that could be made `STRICT` -- passing `NULL` as a parameter will always result in a `NULL`, but for some reason that causes a significant decrease in perf.
* tagged one function as parallel safe
NOTE: somehow `ST_AsMVT()` method of tile generation is showing 70-90% slowdown with this patch. I am not sure of why this is happening. If the reason is the `IMMUTABLE` -> `STABLE` change, we may have to dig deeper into PG optimization
I would like to reformat all of our SQL to have a concise coding style.
This makes it far easier to understand the code for a casual contributor,
and lets us spot errors more easily.
Most importantly, it makes it much easier to grep (search) the code because it is more likely to be in the same syntax
Some key changes:
* SQL keywords are always UPPERCASE, e.g. `SELECT WHEN AS END ...`
* types, variables, aliases, and field names (identifiers) are always lower case
* `LANGUAGE 'plpgsql'` is now `LANGUAGE plpgsql` (no quotes)
* a few minor spacing/semicolon cleanups
P.S. Per @TomPohys request, `TABLE` is spelled using upper case despite being a type for consistency with PG Docs. Same for `LANGUAGE SQL` vs `LANGUAGE plpgsql`.
This PR allows queries to be parallelized on recent versions of Postgres. The `PARALLEL SAFE` modifier has been added to the layer functions and a PLPGSQL function to convert strings into number has been replaced.
`PARALLEL SAFE` is a modifier for `CREATE FUNCTION` available since Postgres 9.6, so this change does not break current OpenMapTiles supported database version. More details about this topic [here](https://www.postgresql.org/docs/current/parallel-safety.html) and at the reference documentation for [`CREATE FUNCTION`](https://www.postgresql.org/docs/current/sql-createfunction.html).
### Testing procedure
The procedure to test this was:
* Imported `spain.pbf` in a clean environment
* Dumped the OpenMapTiles database from the Postgres Docker image
* Created a clean Postgres 12 database using the default Docker image
* Installed `postgis` 3 from the default Debian package and `osml10n` 2.5.8 from the repository (`make`, etc.)
* Restored the dump
* Lowered the postgres planner parameters for triggering parallel plans:
```sql
set parallel_setup_cost = 5;
set parallel_tuple_cost = 0.005;
```
* Manually added the `PARALLEL SAFE` modifier to each function involved in layer queries (not on updates or inserting functions).
* For each layer, run a testing query to confirm parallel workers were created, something like this:
```sql
explain analyze
select * from layer_aerodrome_label(tilebbox(8,128,95),10,null)
union all
select * from layer_aerodrome_label(tilebbox(8,128,97),10,null);
```
* After all the layers were processed and confirmed to start parallel executions, a more complete example was run. This example just retrieves the geometries for all the layers from the same tile but without using any MVT related function.
<details><summary>Testing query</summary>
```sql
-- Using the function layer_landuse
explain analyze
select geometry from layer_water(tilebbox(14,8020,6178),14)
union all
select geometry from layer_waterway(tilebbox(14,8020,6178),14)
union all
select geometry from layer_landcover(tilebbox(14,8020,6178),14)
union all
select geometry from layer_landuse(tilebbox(14,8020,6178),14)
union all
select geometry from layer_mountain_peak(tilebbox(14,8020,6178),14)
union all
select geometry from layer_park(tilebbox(14,8020,6178),14)
union all
select geometry from layer_boundary(tilebbox(14,8020,6178),14)
union all
select geometry from layer_aeroway(tilebbox(14,8020,6178),14)
union all
select geometry from layer_transportation(tilebbox(14,8020,6178),14)
union all
select geometry from layer_building(tilebbox(14,8020,6178),14)
union all
select geometry from layer_water_name(tilebbox(14,8020,6178),14)
union all
select geometry from layer_transportation_name(tilebbox(14,8020,6178),14)
union all
select geometry from layer_place(tilebbox(14,8020,6178),14)
union all
select geometry from layer_housenumber(tilebbox(14,8020,6178),14)
union all
select geometry from layer_poi(tilebbox(14,8020,6178),14)
union all
select geometry from layer_aerodrome_label(tilebbox(14,8020,6178),14);
```
</details>
You can inspect the execution plan and results on [this page](https://explain.dalibo.com/plan/3z). Also [attaching](https://github.com/openmaptiles/openmaptiles/files/3951822/explain-tile-simple.tar.gz) the query and JSON output for future reference. The website gives a ton of details, but you may want to search for nodes mentioning `workers` or `parallel` like in this area referring to `osm_border` or `osm_aeroway_linestring` entities
### Next steps
Since the execution plan is not showing a parallel append at the top level, meaning it's not running each layer individually, I want to continue experimenting with parameters and queries to see if it's possible to even parallelize more the request.
I will post my finding here, even no change in the code should happen.
cc. @nyurik
Co-authored-by: Yuri Astrakhan <yuriastrakhan@gmail.com>
Minor code cleanup:
SQL already returns NULL in the "WHEN" condition
if it is not matched by any of the cases.
Co-authored-by: Eva Jelinkova <evka.jelinkova@gmail.com>
Simplify some of the OSM->OMT field value mappings using declarative syntax.
This approach is not for all cases, but in many it removes
the need of storing the same field in both the .yaml and .sql files.
TODO: support more complex AND/OR cases
