Perform on-the-fly generation of tiles

[ItsLastDay]

The OpenStreetMap project gives statistics about their tile view count here. They have observed that only a small portion of all tiles are ever viewed for large zoom values. So they pre-render some tiles, and then generate the rest on-the-fly.
Another idea from the same source is using meta tiles. These are chunks of 64 tiles, computed as a single command. Later they are served separately. It is very clever, because neighbouring tiles are viewed simultaneously in most cases.

We can benefit from both ideas. Approximate list of steps needed:

• implement a server (with tornado?) that performs serving tiles (either pre-rendered or on-the-fly) to user;
• find a memory leak in current implementation of tile rendering (the script is consuming more and more memory during the run);
• explore a way to compute meta-tiles, but server them as usual tiles

[ItsLastDay]

There are several OSM solutions for this:

• mod_tile Apache module, which is responsible for serving tiles (caching, serving multiple responses, ...)
• Tirex - a backend for rendering tiles. An excerpt code, showing how Tirex handles metatiles.

[ItsLastDay]

Scripts that seem to be useful:

• livetiles.wsgi shows how to serve images from python directly with

  response_headers = [('Content-type', 'image/png'),
                    ('Content-Length', str(len(output)))]
  start_response(status, response_headers)
  return [output]

• polytiles.py - a monstrous script for serving tiles. Uses a database for storing tiles, multithreading for generating them.

[ItsLastDay]

Feature implemented in thefly branch. I'll merge to master one of these days.

[ItsLastDay]

It turned out that performance issues arise when implementing this feature absent-mindedly. Postponed.