Yeah, I've seen that. If the blob actually contains tile files verbatim, I think it should be easy to hack together a little Python program from looking at how you store tiles, or even to modify generate_tiles.py from Mapnik to create the SQLite databases right from the start.

Some pitfalls:

1) when populating the data for the first time, I forgot to enable hstore and use -k for osm2pgsql, and was badly screwed in terms of number of retries.

2) if you want to have not one country, but, say, two bordering countries, and hope to reduce processing time by downloading corresponding datasets from, say, download-fabrik.de — think again. Right at the border, there will be duplicate nodes on those maps. They all also contain some duplicate nodes needed to create overviews, at least capitals of other countries are mapped.

This means that importing countries one by one will fail on the second country. I don't know how to resolve this, but the easier for a human (but not for the machine) way is to download the whole continent, import it and render the bounding boxes as you like.

3) the rendering time is, too, rather slow. I wonder if using PyPy would speed it up a bit. UPD It won't, unless by some miracle PyPy supports Boost bindings.