Assuming that the graph is not too complex locally (in a 256 bp window), ~2 billion initial kmers in a single graph file should require 100-200 GB memory and 200-300 GB disk space in $TMPDIR.

GCSA construction uses a semi-external algorithm that works best when the graph is partitioned (e.g. by chromosome) into multiple .vg files. It can then reduce the memory usage significantly by loading kmers from one graph file at a time.

Excellent, thank you! The first graph I generated, using a minigraph gfa, was incompatible due to missing fields (P). I'm roughly following the pggb pipeline (minimap2 -> seqwish -> vg) now so hopeful this will work.

For the first attempt I used vg prune according to your index construction on the vg wiki. So, with this new, and probably larger graph, should I construct using the -r argument, and then prune?

I'm also unsure how to break up the large graph by chromosome as 1) I'm not working with vcf's rather gfa and 2) they're of varying contiguity/quality with different scaffold names.

Will the vg construct find the reference scaffolds with the -r fasta file or do I have to input them as a variable like you did in your example (e.g seq 1 22) and will this work on a gfa? Thank you!