NEED:
* How long does it take to render one (meta)tile (per zoom level)
* Expected latency for each tile load
* Tile byte size
* Tiles vs snapshot
snapshot might scale better, but everyone expects tiles, plus lets do a dark launch
* rsync low-zoom
* Hashing
- https://en.wikipedia.org/wiki/Rendezvous_hashing
- https://en.wikipedia.org/wiki/Consistent_hashing
* LVS -> NGINX -> varnish-front --> varnish-back -> mod_tile
== Rendevouz hashing algorithm ==
# serverId = first(order(map(list_of_healthy_backends, hash( backend_id + tile_id ))))
# If possible, modify above, to load balance between first N of the top list
== Varnish front ==
* no cache
* load-balance/sharding by tile id
* Hashing: backend = rendezvouz( tile_id + live-varnish_back-servers )
== Varnish BACK Hashing ==
* same machines as varnish front
* memory caching
* backend = rendezvouz( meta_tile_id + live-mod_tile-servers )
== mod_tile servers ==
* render meta tiles
== other ==
* low-zoom rsync; database, db update pull,, etc
==NOTES==
.png & .json for all tiles should go to the same servers