[joepie91_]

Worse, even if they were all fetched in parallel, you would still see terrible loading times simply because a dependency graph can only be traversed depth-wise serially. Doesn't matter what network protocol you use or how parallel it is.

[spankalee]

This is not true at all. For every module you can parse out and load the module's imports in parallel. As you traverse the graph the known and loadable module frontier can grow much wider.

The only way it would be serial is if every module only imported one other module.

[joepie91_]

Note how I was specifically talking about depth-wise.

[spankalee]

Fair, though that wasn't clear.

So then even without modulepreload or bundling, it's not always going to be true that the longest import depth is the limiting factor. Earlier loaded modules can still be parsed in parallel while dependencies are fetched, and completed subtrees can be linked and evaluated. Given that modules are deferred and can be imported early, there's often time for parallel work.

[joepie91_]

I'm not following. Dependencies can be nested, so you cannot assume that a dependency by a given name can be satisfied by a previously loaded dependency by the same name. Which means you still need as many serial roundtrips as there are depth levels, whether you parse in parallel or not.

Sure, you can cut down on the delay introduced by the parsing, but 10 depth levels over a 100ms connection is still going to take a second to fetch, because you simply can't know the N+1th dependency until you have at least completed the Nth roundtrip.