Good to hear that caching is being brought up as a priority. The lack of a timeline cache is the number one issue I have, especially for any comp that has animated values, such as for mograph.
There is also a great need to be able to preview at a lower resoultion/quality than full quality. Other compositing software will allow you to preview at 1/2, 1/4, 1/8 etc res quality, but as far as I know this is not possible in Blender without creating your own proxies manually.
I have also noticed that there is a massive speed difference between a video file and an image sequence when using the viewport compositor.
You can get realtime or close to with a video file, but drastically slower with an image sequence. Even more so if it is a 16bit image sequence. A highly compressed video is much faster for playback than an 8bit png sequence with no compression, which is surprising. Maybe this is a known issue?