hash() yields different results for different platforms

[Qiangning Hong]

I'm writing a spider. I have millions of urls in a table (mysql) to
check if a url has already been fetched. To check fast, I am
considering to add a "hash" column in the table, make it a unique key,
and use the following sql statement:
insert ignore into urls (url, hash) values (newurl, hash_of_newurl)
to add new url.

I believe this will be faster than making the "url" column unique key
and doing string comparation. Right?

However, when I come to Python's builtin hash() function, I found it
produces different values in my two computers! In a pentium4,
hash('a') -> -468864544; in a amd64, hash('a') -> 12416037344. Does
hash function depend on machine's word length?

If it does, I must consider another hash algorithm because the spider
will run concurrently in several computers, some are 32-bit, some are
64-bit. Is md5 a good choice? Will it be too slow that I have no
performance gain than using the "url" column directly as the unique
key?

I will do some benchmarking to find it out. But while making my hands
dirty, I would like to hear some advice from experts here.

 

[Paul Rubin]

However, when I come to Python's builtin hash() function, I found it
produces different values in my two computers! In a pentium4,
hash('a') -> -468864544; in a amd64, hash('a') -> 12416037344. Does
hash function depend on machine's word length?

The hash function is unspecified and can depend on anything the
implementers feel like. It may(?) even be permitted to differ from
one run of the interpreter to another (I haven't checked the spec for
this). Don't count on it being consistent from machine to machine.

If it does, I must consider another hash algorithm because the spider
will run concurrently in several computers, some are 32-bit, some are
64-bit. Is md5 a good choice? Will it be too slow that I have no
performance gain than using the "url" column directly as the unique key?

If you're going to accept the overhead of an SQL database you might as
well enjoy the use of the abstraction it gives you, instead of trying
to implement what amounts to your own form of indexing instead of
letting the db take care of it. But md5(url) is certainly very fast
compared with processing the outgoing http connection that you
presumably plan to open for each url.

I will do some benchmarking to find it out.

That's the right way to answer questions like this.

 

[Grant Edwards]

I'm writing a spider. I have millions of urls in a table (mysql) to
check if a url has already been fetched. To check fast, I am
considering to add a "hash" column in the table, make it a unique key,
and use the following sql statement:
insert ignore into urls (url, hash) values (newurl, hash_of_newurl)
to add new url.

I believe this will be faster than making the "url" column unique key
and doing string comparation. Right?

I doubt it will be significantly faster. Comparing two strings
and hashing a string are both O(N).

However, when I come to Python's builtin hash() function, I
found it produces different values in my two computers! In a
pentium4, hash('a') -> -468864544; in a amd64, hash('a') ->
12416037344. Does hash function depend on machine's word
length?

Apparently.

The low 32 bits match, so perhaps you should just use that
portion of the returned hash?

'0xFFFFFFFFE40DB1E0L'

'0xE40DB1E0L'

 

[Carl Banks]

I doubt it will be significantly faster. Comparing two strings
and hashing a string are both O(N).

Playing Devil's Advocate: The hash would be a one-time operation during
database insertion, whereas string comparison would happen every
search. Conceivably, comparing hash strings (which is O(1)) could
result in a big savings compared to comparing regular strings; but I
expect most decent sql implementations already hash data internally, so
rolling your own hash would be useless at best.

If the OP's database is lacking, md5 is probably fine. Perhaps using a
subset of the md5 (the low 32 bits, say) could speed up comparisons at
risk of more collisions. Probably a good trade off unless the DB is
humungous.


Carl Banks

 

[Dennis Lee Bieber]

I believe this will be faster than making the "url" column unique key
and doing string comparation. Right?

Given that the queries pass the data as strings, you're trading an
integer->string, string->integer conversion pair on each search,
followed by an index file look-up, for a straight index file look-up.

Granted, my college days were a quarter century ago, but at the
time, the data structures class required implementing a hash table
look-up program (hashed-head, multiple-linked list). Hashes weren't
guaranteed to be unique -- the hash only gave an entry point from which
one then performed string comparisons. *

You could maybe use MD5 encoding if you don't want the pure URL in
the look-up, and MySQL, at least, has a built-in for MD5 so you wouldn't
even have to worry about "differences".


* The Amiga file-system was a hashed-head, multiple-linked list -- the
only place I'd ever encountered one outside of that 3rd year college
course. File-names were hashed into an index into a directory block; the
directory block had pointers to file-header/subdirectory blocks... those
blocks then had a linked list to other fh/sd blocks that shared the hash
value.
--
Wulfraed Dennis Lee Bieber KD6MOG
(e-mail address removed) (e-mail address removed)
HTTP://wlfraed.home.netcom.com/
(Bestiaria Support Staff: (e-mail address removed))
HTTP://www.bestiaria.com/

 

[Dennis Lee Bieber]

If the OP's database is lacking, md5 is probably fine. Perhaps using a
subset of the md5 (the low 32 bits, say) could speed up comparisons at
risk of more collisions. Probably a good trade off unless the DB is
humungous.

And if there are collisions, a unique index on the hash is not
usable...
--
Wulfraed Dennis Lee Bieber KD6MOG
(e-mail address removed) (e-mail address removed)
HTTP://wlfraed.home.netcom.com/
(Bestiaria Support Staff: (e-mail address removed))
HTTP://www.bestiaria.com/

 

[Qiangning Hong]

Apparently.

The low 32 bits match, so perhaps you should just use that
portion of the returned hash?

'0xE40DB1E0L'

Is this relationship (same low 32 bits) guaranteed? Will it change in
the future version?

 

[Tim Peters]

[Grant Edwards]
[Qiangning Hong]

Is this relationship (same low 32 bits) guaranteed?

No. Nothing about hashes is guaranteed, except that when x and y are
of a hashable type, and x == y, then hash(x) == hash(y) too.

Will it change in the future version?

That's possible, but not planned. Note that the guts of string
hashing in CPython today is implemented via

while (--len >= 0)
x = (1000003*x) ^ *p++;

where x is C type "long", and the C language doesn't even define what
that does (behavior when signed multiplication overflows isn't defined
in C).

 

[Fredrik Lundh]

/.../ add a "hash" column in the table, make it a unique key

at this point, you should have slapped yourself on the forehead, and gone
back to the drawing board.

</F>

 

[Nick Vatamaniuc]

Using Python's hash as column in the table might not be a good idea.
You just found out why. So you could instead just use the base url and
create an index based on that so next time just quickly get all urls
from same base address then do a linear search for a specific one, or
even easier, implement your own hashes without using any of the
Python's built-in hash() functions. For example, transform each
character to an int and multply them all mod 2^32-1 or something like
that. Even better I think someone already posted the Python's way of
generating hashes for string, well, just re-implement it in Python such
that your version will yield the same hash on any platform.

Hope this helps,
Nick V.

 

[Piet van Oostrum]

GE> The low 32 bits match, so perhaps you should just use that
GE> portion of the returned hash?

If the hashed should be unique, 32 bits is much too low if you have
millions of entries.

 

[Grant Edwards]

Playing Devil's Advocate: The hash would be a one-time operation during
database insertion, whereas string comparison would happen every
search.

Good point.

Conceivably, comparing hash strings (which is O(1)) could
result in a big savings compared to comparing regular strings;

Still, I doubt that the URLs are long enough so that there's a
significant difference.

but I expect most decent sql implementations already hash data
internally, so rolling your own hash would be useless at best.

Precisely. DB designers and implementers have been working on
this problem for 30 years. I doubt the OP is going to be able
to best them with a few minutes work.

If the OP's database is lacking, md5 is probably fine. Perhaps
using a subset of the md5 (the low 32 bits, say) could speed
up comparisons at risk of more collisions. Probably a good
trade off unless the DB is humungous.

My advice: do it the simple way first (let the DB handle it).
Don't try to fix a problem until you know it exists.

Premature optimization....

 

[Grant Edwards]

Is this relationship (same low 32 bits) guaranteed?

No, I don't believe so.

Will it change in the future version?

It may.