Wait, the proposed solution to a data format being slow to parse is to work around the bad performance by caching the already parsed representation? That seems like it has a clear flaw if you’re only accessing the data once…
Accessing any given data once. When you have a total dataset size in the 10-100s of gigabytes range, having to download any significant fraction of it to do data processing is really unfortunate.
But seriously what's up with this total disdain for anyone trying to build applications with OSM data? You don't seem to care whether parsing is near instant or as other commenters have mentioned, literally a majority of total processing time for certain compute jobs
And, you're suggesting that user-facing data should be harder to work with only to make it faster to parse?