When we add items to the similarity.similarity table, we add all columns at once. If we want to add a new type of similarity that requires a new column in this table, we'll have to decide if we want to drop the table, recreate it with all columns and then re-import again, or add a new column and backfill it.

With 24 million rows, it takes about 16 hours to create this table, so it's definitely possible to do so multiple times, but we should check if we can make it significantly faster by just adding a column. AB is currently using PG 10, so we cannot create a non-null column without locking the table. As this table is only used when rebuilding the index files this may not be a huge problem.