Hacker News Reader

Comment by raspasov | Hacker News Reader

CP database (in the CAP theorem sense), distributed transactions via Paxos. Most popular databases (Cassandra, Riak, et al) are AP* systems. CockroachDB is an interesting project in progress that aims to be a CP system from my understanding.

EDITED: seancribbs corrected my acronym spelling in the early morning; *CA - DOES NOT EXIST, what I wrote initially was completely wrong

View on HN

You mean AP, not CA.

CA doesn't exist.

Thanks man, good you caught it, I totally knew that, I need to be more careful with acronyms right after waking up :)

> CA doesn't exist.

I'm not sure that's a good way of putting it. You can have databases that offer both consistency and availability, but they need to fit on a single machine (no partition tolerance).

> I'm not sure that's a good way of putting it. You can have databases that offer both consistency and availability, but they need to fit on a single machine (no partition tolerance).

Garbage collection requires (a delay that is indistinguishable from) partition tolerance. It's pretty illusory on single machines as well. Especially if they're multi-core.

They might be related formally (though I'm not really sure). In practice, relational databases generally implement very special-purpose concurrent garbage collectors that will totally lock up only under pathological circumstances (though I won't pretend that it doesn't happen). If you're talking about GC in general, with special-purpose hardware and/or real time kernels, GC times can be bounded deterministically and are thus distinguishable from a partition (except in rare cases, as with all garbage collectors with cycle collection, where they must revert to stop-the-world because they can't reclaim memory fast enough). And, of course, it is possible to write actual real time systems that do not require garbage collection at all (though scant few of them have been formally verified to the extent that one might like).

CAP doesn't really apply to non-distributed systems. So I'm with the original sentiment. There is no CA.

It exists, it's just that assuming partitions never happen isn't very practical.

I may be incorrect, but my impression was that concurrent systems formalization (at least, in ways that are actually relevant on real hardware) was still an area of very active research. Is saying they exist meaningful even in a strictly theoretical sense?

Reply on news.ycombinator.com