Yeah, the OSM PBF format is really efficient in size, but it can be quite a pain to work with those files. Every way consists of nodes, which are usually at the very beginning of the file (first nodes, then ways, then relations). So in order to work efficiently, you'll need a node cache (which at the moment is more than 40G). Seeking is not viable because there is no indication in which block the node will be. On-disk caching is an option, but it will slow you down substantially.

So unless you are doing meta-analysis of the raw OSM data, without assembling geometries, hand-rolling an OSMPBF reader is viable, otherwise I would suggest either using something pre-processed (extracts in real geodata formats), an established parser like osmium or rather import the data into e.g. postgres and do some querying there.