Longest Prefix Matches for URLs

Question

I need information about any standard python package which can be used for \"longest prefix match\" on URLs. I have gone through the two standard packages http://packages.py

Answer 1

This example is good for small url lists but does not scale well.

def longest_prefix_match(search, urllist):
    matches = [url for url in urllist if url.startswith(search)]
    if matches:
        return max(matches, key=len)
    else:
        raise Exception("Not found")

An implementation using the trie module.

import trie


def longest_prefix_match(prefix_trie, search):
    # There may well be a more elegant way to do this without using
    # "hidden" method _getnode.
    try:
        return list(node.value for node in prefix_trie._getnode(search).walk())
    except KeyError:
        return list()

url_list = [ 
    'http://www.google.com/mail',
    'http://www.google.com/document',
    'http://www.facebook.com',
]

url_trie = trie.Trie()

for url in url_list:
    url_trie[url] = url 

searches = ("http", "http://www.go", "http://www.fa", "http://fail")

for search in searches:
    print "'%s' ->" % search, longest_prefix_match(url_trie, search)

Result:

'http' -> ['http://www.facebook.com', 'http://www.google.com/document', 'http://www.google.com/mail']
'http://www.go' -> ['http://www.google.com/document', 'http://www.google.com/mail']
'http://www.fa' -> ['http://www.facebook.com']
'http://fail' -> []

or using PyTrie which gives the same result but the lists are ordered differently.

from pytrie import StringTrie


url_list = [ 
    'http://www.google.com/mail',
    'http://www.google.com/document',
    'http://www.facebook.com',
]

url_trie = StringTrie()

for url in url_list:
    url_trie[url] = url 

searches = ("http", "http://www.go", "http://www.fa", "http://fail")

for search in searches:
    print "'%s' ->" % search, url_trie.values(prefix=search)

I'm beginning to think a radix tree / patricia tree would be better from a memory usage point of view. This is what the a radix tree would look like:

Whereas the trie looks more like:

Answer 2

The function below will return the index of the longest match. Other useful information can easily be extracted as well.

from os.path import commonprefix as oscp

def longest_prefix(s, slist):
    pfx_idx = ((oscp([s, url]), i) for i, url in enumerate(slist))
    len_pfx_idx = map(lambda t: (len(t[0]), t[0], t[1]), pfx_idx)
    length, pfx, idx = max(len_pfx_idx)
    return idx

slist = [
    'http://www.google.com/mail',
    'http://www.google.com/document',
    'http://www.facebook.com',
]

print(longest_prefix('http://www.google.com/doc', slist))
print(longest_prefix('http://www.face', slist))

Answer 3

Performance comparison

suffixtree vs. pytrie vs. trie vs. datrie vs. startswith -functions

Setup

The recorded time is a minimum time among 3 repetitions of 1000 searches. A trie construction time is included and spread among all searches. The search is performed on collections of hostnames from 1 to 1000000 items.

Three types of a search string:

• non_existent_key - there is no match for the string
• rare_key - around 20 in a million
• frequent_key - number of occurrences is comparable to the collection size

Results

Maximum memory consumption for a million urls:

| function    | memory, | ratio |
|             |     GiB |       |
|-------------+---------+-------|
| suffix_tree |   0.853 |   1.0 |
| pytrie      |   3.383 |   4.0 |
| trie        |   3.803 |   4.5 |
| datrie      |   0.194 |   0.2 |
| startswith  |   0.069 |   0.1 |
#+TBLFM: $3=$2/@3$2;%.1f

To reproduce the results, run the trie benchmark code.

• rare_key/nonexistent_key case

  if number of urls is less than 10000 then datrie is the fastest, for N>10000 - suffixtree is faster, startwith is significally slower on average.

• axes:

  • vertical (time) scale is ~1 second (2**20 microseconds)
  • horizontal axis shows total number of urls in each case: N= 1, 10, 100, 1000, 10000, 100000, and 1000000 (a million).

• frequent_key

  Upto N=100000 datrie is the fastest (for a million urls the time is dominated by the trie construction time).

  The most time is taken by finding the longest match among found matches. So all functions behave similar as expected.

startswith - time performance is independent from type of key.

trie and pytrie behave similar to each other.

Performance without trie construction time

• datrie - the fastest, decent memory consumption

• startswith is even more at disadvantage here because other approaches are not penalized by the time it takes to build a trie.

• datrie, pytrie, trie - almost O(1) (constant time) for rare/non_existent key

Fitting (approximating) polynoms of known functions for comparison (same log/log scale as in figures):

| Fitting polynom              | Function          |
|------------------------------+-------------------|
| 0.15  log2(N)   +      1.583 | log2(N)           |
| 0.30  log2(N)   +      3.167 | log2(N)*log2(N)   |
| 0.50  log2(N)   +  1.111e-15 | sqrt(N)           |
| 0.80  log2(N)   +  7.943e-16 | N**0.8            |
| 1.00  log2(N)   +  2.223e-15 | N                 |
| 2.00  log2(N)   +  4.446e-15 | N*N               |

Answer 4

If you are willing to trade RAM for the time performance then SuffixTree might be useful. It has nice algorithmic properties such as it allows to solve the longest common substring problem in a linear time.

If you always search for a prefix rather than an arbitrary substring then you could add a unique prefix while populating SubstringDict():

from SuffixTree import SubstringDict

substr_dict = SubstringDict()
for url in URLS: # urls must be ascii (valid urls are)
    assert '\n' not in url
    substr_dict['\n'+url] = url #NOTE: assume that '\n' can't be in a url

def longest_match(url_prefix, _substr_dict=substr_dict):
    matches = _substr_dict['\n'+url_prefix]
    return max(matches, key=len) if matches else ''

Such usage of SuffixTree seems suboptimal but it is 20-150 times faster (without SubstringDict()'s construction time) than @StephenPaulger's solution [which is based on .startswith()] on the data I've tried and it could be good enough.

To install SuffixTree, run:

pip install SuffixTree -f https://hkn.eecs.berkeley.edu/~dyoo/python/suffix_trees