Hacker News new | ask | show | jobs
by leif 4687 days ago
Actually, you can beat B-trees pretty handily across the board in exactly the scenario you described (a, b, c).

The log(N) performance of a B-tree is not just extremely hard to improve on (for searches), it's impossible to improve on. The lower bound for searching (in the DAM model) is log(N)/log(B), and B-trees meet that.

But B-trees are also log(N)/log(B) for insertions, which is, it turns out, pretty damn slow. There's an optimal trade-off curve between insertions and queries (the fastest data structure for insertions is to just log them, but that forces all queries to read all the data), and B-trees are on it, but there are many more interesting points on that curve.

There is a family of data structures that matches B-trees' performance on queries while blowing it completely out of the water for insertions. The COLA and Cache-Oblivious Streaming B-tree are where you'll find them in academia, and the implementation I work on has a very marketing-flavored name: the Fractal Tree. All that means is that we took the theory and started implementing it and came up with enough innovations in the implementation that it sort of needed a new name, but it's spiritually the same concept.

I've written a fair amount about this that I'll link to at the end, but here's a brief description so you don't think I'm making this all up.

Basically what we do is we take a B-tree and, on all the internal nodes, we stick large buffers that accumulate messages. Want to insert something? Just stick it in the root's buffer, you don't need to do any I/O to find the proper leaf node it needs to go in (yet). If the buffer's full, flush it down by taking all its messages, sorting them between the root's children, and putting messages in the buffers in the children. This flushing can cascade as you'd imagine, and splits and merges work about the same as in a B-tree.

So now what's the analysis?

Well, the tree has the same shape as a B-tree, so searches have to look at the same log(N) number of nodes (which in practice is almost always just 1 for the leaf node after cache hits on the internal nodes, both for B-trees and for Fractal Trees, but the asymptotic analysis also tells you they have the same query cost).

For insertions though, let's walk through the process. The tree has height log(N), which means that, for an insertion to get "fully inserted", meaning it reaches the leaf node and won't be flushed any more, it has to get flushed down log(N) times. And what's the cost to flush a buffer full of messages? That's just O(1) I/Os, for the parent and children (and in practice it's actually only 2 I/Os because you can just flush to a single child, not all of them). But a buffer flush does work for O(B) messages at once, so the amortized cost to flush a single message down one level is just O(1/B). Do that log(N) times, and the insertion cost is O(log(N)/B), which is in practice around 100x less expensive than the B-tree's O(log(N)/log(B)) insertion cost.

On top of this, while for B-trees you want small leaf nodes because you're going to be reading and writing them all the time, for Fractal Trees, since the goal is to get a lot done with each I/O, you actually want large leaves, on the order of a few megabytes each. This has two nice effects:

1) While range queries on a B-tree can slow down as the tree ages and the leaves start to get randomly placed on disk, range queries on a Fractal Tree stay fast because each time you do a disk seek, you get to read and report a few megabytes' worth of data. This basically solves the "B-tree fragmentation" problem that makes database users run optimize table, or vacuum, or reIndex() or compact() operations like madmen.

2) Compression algorithms (like zlib, our default) can compress large blocks of data much more effectively than they can compress small blocks. So InnoDB, which has small blocks like most B-trees, if you turn on compression, apart from eating CPU as it tries and fails and re-tries to compress your data to fit it into its block size, it'll only get at most about 4x compression. In contrast, TokuDB (our MySQL storage engine using Fractal Trees [4]) routinely gets 10-20x compression without breaking a sweat.

I have some blog posts about this [1] and [2], and our benchmarks page is [3]. We also have a version of MongoDB in which we've replaced all the storage code with Fractal Trees, we call it TokuMX [5]. Don't mind the marketing haze, it's all serious tech under the hood.

[1]: http://www.tokutek.com/2011/09/write-optimization-myths-comp...

[2]: http://www.tokutek.com/2011/10/write-optimization-myths-comp...

[3]: http://www.tokutek.com/resources/benchmarks/

[4]: http://www.tokutek.com/products/tokudb-for-mysql/

[5]: http://www.tokutek.com/products/tokumx-for-mongodb/

3 comments

That looks extremely interesting! The idea of amortizing the cost of inserts is fascinating. Looking at the design you sketched a few questions come to mind:

1) Multi-threading: suppose I seek down the B-tree for key K. Most B-tree implementations use the latch on the node containing K as the final arbiter of concurrency. For example, if I'm looking at K and then I want the next row (perhaps because I'm using the new Index Condition Pushdown optimization in MySQL 5.6, or I'm doing an online index build and need to scan all the rows) then I can simply look at the next row on the page I currently have (read) latched. With a fractal tree it looks like I have to worry about someone inserting a row immediately after the current row because that insert could have been cached at a higher level. Does this mean I need to keep some sort of latch/lock on the entire b-tree path down to the page I'm reading, instead of using latch coupling to work my way down? Alternatively do I have to work my way down from the top of the tree every time I want the next key?

2) How can you check for uniqueness? Suppose I create a table like this:

CREATE TABLE t1 (id NUMBER PRIMARY KEY, val1 VARCHAR2(30));

Amortizing the inserts seems to imply that primary key uniqueness violations can't be discovered until the inserts are pushed all the way down to the leaf?! In general uniqueness is an important part of data normalization, a good input for query optimization and normally required for foreign key constraints...

It is highly concurrent but you need some tricks beyond that simple description. Here's how we did it: http://www.tokutek.com/2013/01/concurrency-improvements-in-t... http://www.tokutek.com/2013/02/concurrency-improvements-in-t...

Yes, unique checks are bad. They make it perform as badly as a B-tree for unique inserts. There are sometimes ways around that but at some level if you aren't reading in the leaf node, you're going to be at a loss for some information. B-trees seem to have spoiled users into thinking uniqueness checks don't make inserts any more expensive, when in fact that's just because B-tree inserts are already that slow.

So out of the main b-tree operations (Insert/Replace/Delete/Seek/Next/Prev) you make a convincing argument that Tokutek can be faster than b-trees for inserts, if you use non-unique indexes (which automatically disqualifies the primary index in most cases, and makes foreign keys hard).

That is good but not quite "beat B-trees pretty handily across the board" :-)

Many primary keys in the wild are sequential, which makes the unique check hit cache and not require any I/O.

Let me be clear: in the worst case with a unique index, Fractal Tree indexes are the same speed as B-tree indexes. In most cases, they're far better. When you add in compression and agility, it really is "beat B-trees across the board."

To be honest, I have no idea what "agility" is. Perhaps I am too cautious but I'll want to see a lot more data than some hand-picked benchmarks and a mystical "unique constraints aren't interesting" before I'm completely convinced.

[Random aside: I greatly dislike block (i.e. InnoDB-style) compression, which is what Tokutek seems to use. Decompressing the entire block just to retrieve one column of one row is crazily expensive, especially as the decompression cost goes up as the block size goes up. There is also the problem of deciding whether to store compressed blocks in the buffer cache (slow), decompressed blocks in the buffer cache (wastes ram) or to have a cache of uncompressed blocks (the InnoDB approach, which kind of combines both problems). I think that format-aware page or column compression (like Oracle, SQL Server or ESENT) is far more effective.]

I guess that this has gotten way off topic but I am glad there are people out there doing exciting new things in the B-tree space. I will keep a closer eye on Tokutek in the future.

Agility refers to our online schema change abilities in mysql (hot column add/delete, hot indexing).

Our compression technique is naive, you're right (but we don't decompress the whole 4MB for one single point query, as you may be thinking, we're a bit more sophisticated). We have some more ideas in the pipeline if we need them (basically a bunch of tricks you can play when you know the schema), but the fact remains that what we have right now is extremely effective and there isn't much incentive yet to finish off our smarter prototypes.

I now realize I didn't answer how we check uniqueness. We just do a query like any other normal point query, using a serializable transaction, and if we pass the check, then we do the insert with the same transaction (which took a row lock when we did the query because it was serializable, so nothing could have violated the uniqueness between the query and the insert).
Also, MVCC snapshots mean that if you're doing a range query, it doesn't matter if someone comes in and injects a message above you while you're scanning a node.
So why isn't everyone using this instead of InnoDB these days? Is it not production-ready yet?
It is production ready, and many people are using it, but as with any deep stack technology, growth is a slow process. From an engineering standpoint, I'd say the data structure wasn't fully mature (ready to replace nearly all uses of InnoDB) until probably the end of 2012. Most of this was issues with concurrency that we methodically picked apart from the 5.0 release through 6.6.

A few users still point out annoyances here and there (like the location of files on disk, how certain metadata is presented at the MySQL level, dealing with corner cases in the query optimizer) that don't have to do with the data structure, but with the integration as a MySQL concept, and though these complaints are rare, they will eventually need to be addressed, and it's hard to find the manpower to address them immediately.

Part of this problem is that it just takes a long time to sand down all the rough little edges of a product, even though the core data structure is mature. Another part is educating people, changing expectations (for example, teaching people that unique indexes do have a higher cost than non-unique indexes), and generating better documentation. It's a slow process but we're confident that it's progressing and will continue. Recall that InnoDB, while broadly accepted as superior to MyISAM for many years before, only became the actual default engine in MySQL 5.5.

> We also have a version of MongoDB in which we've replaced all the storage code with Fractal Trees, we call it TokuMX [5].

What's the deal with TokuKV? Is it a working drop-in replacement for BerkeleyDB and similar key-value stores?

It's not really a drop in replacement for BDB, it's more like a library whose API was inspired by BDB. You're welcome to use it directly, but it doesn't implement all of BDB, we have added a bunch of things (like db->update), and there may be some weird contractual things you need to get right that we haven't documented well. Contact us if you'd like to use it, we can help you.
Thanks! I'm certainly not hung up on the BDB API, so I will check it out.