The current broadcasting strategy is to just write into etcd and then read it back from all nodes. Etcd isn't designed for this and this is probably bad for anything over a megabyte in size or so.
Strategies to introduce:
Spark filesystem (kinda like how the observations and user weights reading & writing currently work)
Torrent
This issue also involves decoupling the Model code from the exact type of broadcast being provided.
The current broadcasting strategy is to just write into etcd and then read it back from all nodes. Etcd isn't designed for this and this is probably bad for anything over a megabyte in size or so.
Strategies to introduce:
This issue also involves decoupling the Model code from the exact type of broadcast being provided.