Cdb: Add support for cdb64 | Hacker News Reader

Cdb: Add support for cdb64

cdb.cr.yp.to

86 pointskreco26 comments

CDB is an interesting format, optimized for read-heavy write-rarely[1] random lookups on slow media. This isn’t a very common requirement these days, but it’s convenient for very specific use cases.

[1] You “update” by overwriting the entire file. This is remarkably fast and means that there’s no overhead/tracking for empty space, but it does mean you probably want this to be a fairly rare operation.

I rolled my own cdb reader library for a project a few years ago, and wrote up my notes on the format and its internals here: https://search.feep.dev/blog/post/2022-12-03-cdb-file-format

Can't this be implemented as a PHF: https://en.wikipedia.org/wiki/Perfect_hash_function

GALACTIC SCALE QMAIL that can run efficiently on a 486 AND survive a supernova!

Haven't there been 64-bit ports of CDB for ages?

Yes, the modifications you need to support it are trivially obvious (literally just replace “4 bytes” with “8 bytes” everywhere in the spec) and have been implemented by a number of authors, some of which this page links to. I guess it’s nice that they’ve been “officially” acknowledged, though.

And update the hash algorithm, yes?

In answer to my own question, no, except for the trivial expansion to 64-bits. https://cdb.cr.yp.to/cdb-20251021/cdb_hash.c.html with constants at https://cdb.cr.yp.to/cdb-20251021/cdb.h.html .

What's Bernstein working on that would have demanded this?

Interesting, never heard of this before. I'm assuming the use case is when your data is too large to conveniently fit into memory?

It is a database for when you read a lot and don't write too often; when a write might be pretty big but not frequent; when you don't want to write a database engine yourself (I.e. figure out what to write and when). And, especially, when corrupting the data would be a big problem.

And it is especially good on copy-on-write filesystems, because it is CoW itself.

So it's not constant?

The lookups are ~O(1).

Nothing is truly constant lookup in number of elements in nature because we can’t pack it tighter than a sphere.

It's a database for strictly exact-match lookups for very read-intensive workloads; think systems where the database only changes when the configuration changes, like email alias or domain lookups. It's very simple (a first-level hash table chaining to a second-level open-addressed hash table) and easy to get your head around, but also very limiting; an otherwise strict K-V system that uses b-trees instead of hash tables can do range queries, which you can build a lot of other stuff out of.

Most people would use Redis or SQLite today for what CDB was intended for; CDB will be faster, but for a lot of applications that speed improvement will be sub-threshold for users.

For me this answer was helpful and succinct, thank you.

Great reply.

What comes to mind from my experience is storing full shipping rate tables for multiple shipping providers. Those change extremely rarely but are a high throughput exact lookup in a critical path (a checkout).

But we just implemented them in SQLite and deployed that file with the application. Simple clean, effective, and fast. Maybe shipping rate data is smaller than this is intended for, but I doubt using this instead would see a consequential perf increase. Seems niche, like the domain name lookup example.

I'm kind of surprised I hadn't heard of this, I could see this being something useful for a few projects. Historically for things in this space I've used RocksDB but RocksDB has given me headaches with unpredictable memory usage for large data sets.

In the low-level DB space there is also https://dbmx.net/tkrzw/, of the Tokyo Cabinet / Kyoto Cabinet lineage.

cdb is a fun format to implement! highly recommend it.

Now I’m curious about working around the writer limitations….

It's designed to rebuild the whole database with every write, and the format reflects that.

yes — I have a few thoughts in mind for how to build a usable database around static chunks of that sort but I don’t suppose I’ll ever get around to it nor do I feel the need to explain myself beyond this. :)

Title: cdb: Intro (please use the original title) https://news.ycombinator.com/newsguidelines.html

It'd never have crawled out of the new page with that title.

No, the title is much better as it is.

weak

Reply on news.ycombinator.com