Comment by joe_the_user | Hacker News Reader

Comment by joe_the_user | Hacker News Reader

The theoretical properties of hash table always seemed so impressive to me that they bordered on magic (and this just extends them). What seemed crazy was how they could be so much better than trees, which to me were intuitively the most efficient way to store data.

What I realized is that the theory of hash tables involves a fixed-sized collection of objects. For this fixed collection, you create a hash-function and used that like a vector-index and store the collection in a (pre-allocated) vector. This gives a (fuzzy-lens'd) recipe for O(1) time insert, deletion and look-up. (The various tree structures, in contrast, don't assume a particular size).

The two problems are you have to decide size beforehand and if your vector gets close to full, you insert etc processes might bog-down. So scanning the article, it seems this is a solution to the bogging down part - it allows quick insertion to a nearly-full table. It seems interesting and clever but actually not a great practical advance. In practice, rather than worrying a clever way to fill the table, I'd assume you just increase your assumed size.

Edit: I'm posting partly to test my understanding, so feel to correct me if I'm not getting something.

Proofs of constant time operations include time taken to resize the table. This takes much more time (linear in the size of the table), on insertions when the table is resized, but that time is amortized over all the insertions already done. It still works out to constant average time if you grow the table enough each time (once it starts to get too full) so it happens with decreasing frequency.

>What seemed crazy was how they could be so much better than trees, which to me were intuitively the most efficient way to store data.

Trees (sorted) are good at finding subsets and ranges "scanning" or "searching" but hashmaps are better at "seeking" like a key-value lookup

I think this is only true in the imperative world, where mutation is used. For the functional world it is probably still trees.

It looks to me like the idea is, as you generally describe, that you segment your table into a 2d structure (well conceptually) and proceed to fill one ‘row’ at a time until it’s about 75% full, at which point you move on to the next one.

I don’t have time to fully grok the paper, but they claim this makes insertion consistently fast (I believe this until we’re at 75% of total capacity, but maybe they have some other mode for filling when they’re at 75% in every row?). They also claim retrieval is fast, and I didn’t read enough to understand how even retrieval works, or why it is faster.

I’ll put out that there a lot of times that it would be really nice to have a nearly full hash table still, you know, work. You can’t always change the size of one during execution of a program. And, in some environments memory counts a lot. That said, I would like to see and play with an implementation — I’m not sure this is ‘worth it’ in the general case.

It is also probably cache inefficient, as are most things about hash tables, with the exception of linear probing for reading out of a fairly full one, in which case, you get to just keep pulling stuff directly out of memory to check it. So, it’s not clear to me that this is performance wise worth it. Anyway, I’d like to fully understand it, it seems like an interesting new idea.

Reply on news.ycombinator.com