Optimization list

[sparshev]

We can optimize the processes to make them much more usable, and that's how:

• Preallocate the resources - have the started up label just waiting for the metadata from requestor. Need to calculate how often the cluster uses this label and decide if the preallocation worth it for specific node.
• Predownload the label images - if label was updated, it's a good signal to download the images if they was changed and save time on downloading
• p2p for images distribution - to remove some load from the artifact storage where all the images are placed we can use p2p to distribute the images across the cluster

... - there is much more, so let's keep them here

[solvingj]

Heuristics and smart behavior is all good and valuable. I would probably first look for configurable override. The ability to specify some list of images/labels/versions which get pre-populated, but also with some control over which nodes get which images. On second thought, I bet providing such configurability in a way which is ACTUALLY usable and ergonomic is actually much harder than it seems. Probably naive strategies won't work out in practice.