Mala Mahadevan continues a series on vector search:
At this point, the consideration becomes – ‘What if the graph and all of its vectors don’t fit in memory?’
If every step of graph traversal had to wait for an arbitrary SSD read, the navigation could become painfully slow.
This is the plus and the minus of an index like HNSW. If you can fit everything into memory, it’s faster than DiskANN. Once you start paging in and out of disk, DiskANN becomes considerably faster.