PostGIS Raster driver: Block caching

[jorgearevalo]

Is necessary a caching method to avoid fetching data from database each time IReadBlock is called.

[jorgearevalo]

Implementing IRasterIO (at dataset level) we avoid too many server rounds by fetching all the required tiles in one query. The problem is not all tiles has to have the same dimensions. This complicates the use of GDAL cache. We can try to solve it by divinding all the fetched data in equally sized blocks, and manually storing them on the cache.

[jorgearevalo]

Basic caching on the easiest case (regulary blocked raster, region size equal to buffer size) implemented in r23098. Working on known bugs and limitations.

[jorgearevalo]

Bug related with non-square tiles. 2 use cases cause errors (image looks like a puzzle with the pieces in wrong places):

• Image divided into non-square tiles
  
• Non-square image divided into square tiles (or non-square tiles, it fails anyway)

[jorgearevalo]

The bug wasn't related with non-square tiles. The problem was PostGISRasterDataset::IRasterIO puts all the blocks in the output buffer in the same order they were fetched from database, and this order isn't necessarily be the 'right' order, in terms of position inside the raster.

The temporary solution is to force a order by rid in the query that fetches the blocks from database. This is not the perfect solution, because the blocks are not stored in a particular order (the rid order normally matches the position of each block inside the image, but this is not a constraint). This temporary solution has been applied at r23180

[jorgearevalo]

With r23218, the tiles are fetched following a upperlefty, upperleftx order. Now, it's necessary to ensure a good block size, that works in any situation (regular and irregular blocking).

[jorgearevalo]

With r23219, we check that all the blocks have the same size, snap to the same grid and have the same srid. If one of these fail, we raise and error. Defined attributes to allow user to specify these constraints are true by default, and avoid the checkings.

[warmerdam]

removing milestone.

[jorgearevalo]

The problem is avoided using MEM/VRT drivers to take care of the data. Done in r24815