Aguri: Coolest Data Structure You've Never Heard Of

http://www.matasano.com/log/1009/aguri-coolest-data-structure-youve-never-heard-of/

by iamelgringo • 19 years ago

57 39 19 years ago

matasano.com

mojuba3• 19 years ago

Radix tries are nice. Use them when

(1) you need fast lookup and fast inserts

(2) you need to preserve the lexicograhpic order of your elements - something that a hash table won't give you

(3) you are afraid of DoS attacks or maybe just "bad" data - hash tables can be attacked easily if the implementation is known

(4) you don't care about memory usage that much.

I compared a simple 8-bit radix tree with some standard hash table implementation - the former took roughly ten times more memory. I then changed my radix to be based on 4 bits (each char is just split into 2 parts) and the memory usage improved twice. Now I'm wondering if radix tries have more room for improvement.

So this is something to consider.

And thanks for the article - never heard of Aguri, it's beautiful.

Upd. sorted lists basically give you same advantages except for fast inserts

tptacek10• 19 years ago

Compressed binary radix tries (also known as PATRICIA trees) give you excellent memory usage characteristics and pretty much optimal search time characteristics: both wins come from only storing the edges corresponding to significant bit differences.

Since hash tables by necessity have to allocate enough storage to keep table load down and minimize collisions, you could find a PATRICIA tree has even better memory footprint than whatever hash table you use.

Of course, all data structures share the DoS problem. There's an input to a PATRICIA tree that pessimizes lookup and memory usage: the one that creates a full 32 bits of fanout for every key. In all cases, the attack degrades performance to linear search.

Of all the "mainstream" data structures, balanced binary trees are probably the hardest to attack this way.

tptacek10• 19 years ago

I am, of course, an idiot; you can't make a PATRICIA lookup worse than O(k).

mojuba3• 19 years ago

both wins come from only storing the edges corresponding to significant bit differences.

Of course I used compressed suffixes, but it didn't help much as compared with hash tables.

you could find a PATRICIA tree has even better memory footprint than whatever hash table you use

I don't think you can demonstrate that Patricia has a better memory footprint than a good hashtable. In fact best radix trees are much worse.

Of course, all data structures share the DoS problem.

You can attack radix trees in terms of memory usage, while hash tables are prone to serious performance attacks.

In all cases, the attack degrades performance to linear search.

Not true. I don't think you understand these algorithms and O(n) complexity very well.

The question how a hash table degrades when attacked highly depends on how exactly conflict resolution is implemented. Among other methods is, for example, expansion and rehashing, and you can't tell for sure if it degrades to linear search or not.

And radix never degrades to linear search - never.

Of all the "mainstream" data structures, balanced binary trees are probably the hardest to attack this way.

They are prone to a very specific attack when you make the tree to re-balance often.

jwp• 19 years ago

Suffix array.

andrewf• 19 years ago

A couple of years back, googling for "suffix array" got you advertisements for jobs@google. I'm not sure whether Google HR ended the program, or just migrated to more obscure search terms :)

apathy• 19 years ago

Burrows-Wheeler Transform maybe?

One of the coolest compression algorithms evar...

dfranke• 19 years ago

Knuth has a pretty good treatment of tries in volume 3.

Xichekolas4• 19 years ago

He says binary trees are better default search table than hash table... binary tree search is O(lg(n)) while hash table is O(1)... so hash table is better for search, binary trees win in other ways (like ordered traversal if you need that), but no way you could say one is always better than the other.

Also confused how he proposes to build his Radix Trie in less than linear time... just seems like he ditches his original premise and launches into routing. There are selection algorithms that will find things in an unsorted array in O(lg(n))... but pretty sure building a tree of any sort is not part of it.

Radix Tries are sweet fun... so is just good old Radix Sort, but he starts with something that doesn't really make sense for the topic (the unsorted array he wants to pull three out of), then makes some weird claims, then jumps into routing with Radix Tries... not sure what is insightful here.

aston• 19 years ago

There aren't any algorithms to find things in an unsorted array in less than worst case linear time. I think you're confusing search (for some value) with select (a value with a given index in the sorted array).

Xichekolas4• 19 years ago

Yeah, yeah I was... was just thinking about that over breakfast.

eru2• 19 years ago

If you want to do several searches on the same data - that is initially unsorted - you can get better than O(n) per query. But like you said - not for the first one.

tptacek10• 19 years ago

Hash tables are NOT O(1). You just failed the job interview.

You're like the 4th person I've seen that read (some) of this article and pulled this tree vs. hash thing out of it. It's a throwaway point. Binary trees provide comparable performance to hash tables, but also provide efficient access to ranges of keys. Binary trees support more operations efficiently than hash tables.

The article isn't about some war I've declared on hash tables.

gojomo• 19 years ago

"Hash tables are NOT O(1)."

For average lookup time, which is a reasonable context to assume in the preceding comment, yes they are.

marketer• 19 years ago

What do you mean hash tables are not O(1)?

With a good hash function (collision resistant), the amortized cost of lookup is O(1), even if you need to resize the table.

Xichekolas4• 19 years ago

"tree vs. hash thing ... It's a throwaway point"

If you had read more than (some) of my comment you might have gathered that was my entire point. The first page of your post was a throwaway point... nothing to do with your actual point, which seemed to be Radix Tries. It would've been nice to dispense with the distraction was all I was saying.

I know that given a lot of collisions that hash tables degenerate into linear insertion and lookup... and I did point out that binary trees are better for several things, but I still maintain that you can't just outright claim one is better than the other for every situation.

Oh, and as for failing the job interview... not really sure why that was necessary. You don't make a stronger point by being a dick about it.

bayareaguy6• 19 years ago

Hash tables are NOT O(1). You just failed the job interview.

tptacek, what do you think they are?

Tichy• 19 years ago

For hash tables you need to compute the hash function and execute an equals operation. Computing the hash function might take longer than traversing a tree. For example, if you key is "sdolkfhasodfgkokjhdfgfnaryksfdksfdgkjgf" the hash function will probably take every single letter into account (>20 operations, didn't count), whereas a binary tree might only need to look at the first letter at most nodes. So at least for relatively few entries, trees might perform better than hashtables, O(1) nonwithstanding.

garret• 19 years ago

D. J. Bernstein uses these too.

http://cr.yp.to/critbit.html

codeslinger• 19 years ago

The point of the data structure in this case was not the radix trie alone but rather the combination of the use of tries in concert with the aggregation characteristics of IP traffic. This type of innovative combination is common in the network world due to the necessity of providing ever more introspection and functionality over ever higher throughputs and ever lower latencies. For some serious treatment of the subject, check out the book "Network Algorithmics" by George Varghese.

tptacek10• 19 years ago

This is an awesome, awesome, awesome book. Aguri isn't in it; aguri doesn't even rate for this book. That's how awesome this book is.

anupamkapoor• 19 years ago

network-algorithmics is just soooo cool. really, what you must be reading after stevens.

hisham• 19 years ago

Wow reading this post was like replaying one of my interviews with Google...

RajendranB• 19 years ago

The time space trade off between hashtables and binary trees are not touched upon in all the above writings. Hashtables will require continuous memory locations and hence from stack and hence won't work for large sets of data and binary search trees or their derivatives will allocate memory from heap and take more time (logn) for get and put than the constant time(O(1)) for hashtables.

bayareaguy6• 19 years ago

It sounds like you're making a lot of assumptions about the storage hierarchy and addressing scheme.

Aguri: Coolest Data Structure You've Never Heard Of

Related Stories