PROs:

• its easier/faster to check&skip already existing results w.r.t. walking the filesystem
• no additional deps (built-in in python)
• still a multi-language self-contained file-based storage
• we can implement several logic on SQL (object filtering and counting, conditional indexing, etc.)

CONs:

• SQL-like management, we'll have to cope with migrations
• probably will use more disk space
• embeddings as BLOBs. Save and load are pretty fast though, on SSD:

  Bulk insertion of 100k 1024-d vectors took 5.6198 seconds.
  Reading out 100k 1024-d vectors took 2.3810 seconds.