Hive LIKE predicate. '_' wildcard decrease perfomance

[Igor Kuzmenko <f1...@gmail.com>]

I've got Hive Transactional table 'data_http' in ORC format, containing
around 100.000.000 rows.

When I execute query:

select * from data_http
where res_url like '%mts.ru%'

it completes in 10 seconds.

But executing query

select * from data_http
where res_url like '%mts_ru%'


takes more than 30 minutes.

Why '_' wildcard decrease perfomance?

Re: Hive LIKE predicate. '_' wildcard decrease perfomance

[Igor Kuzmenko <f1...@gmail.com>]

Thanks for reply, Gopal. Very helpful.

On Thu, Aug 4, 2016 at 10:15 PM, Gopal Vijayaraghavan <go...@apache.org>
wrote:

> > where res_url like '%mts.ru%'
> ...
> > where res_url like '%mts_ru%'
> ...
> > Why '_' wildcard decrease perfomance?
>
> Because it misses the fast path by just one "_".
>
> ORC vectorized reader has a zero-copy check for 3 patterns - prefix,
> suffix and middle.
>
> That means "https://%", "%.html", "%mts.ru%" will hit the fast path -
> which uses StringExpr::equal() which JITs into the following.
>
> https://issues.apache.org/jira/secure/attachment/
> 12748720/string-intrinsic-
> sse.png
>
>
> In Hive-2.0, you can mix these up too to get "https:%mts%.html" in a
> ChainedChecker.
>
>
> Anything other than these 3 cases becomes a Regex and takes the slow path.
>
> The pattern you mentioned gets rewritten into ".*mts.ru.*" and the inner
> loop has a new String() as the input to the matcher + matcher.matches() in
> it.
>
> I've put in some patches recently which rewrite it Lazy regexes like
> ".?*mts.ru.?*", so the regex DFA will be smaller (HIVE-13196).
>
> That improves the case where the pattern is found, but does nothing to
> improve the performance of the new String() GC garbage.
>
> Cheers,
> Gopal
>
>
>

Re: Hive LIKE predicate. '_' wildcard decrease perfomance

[Gopal Vijayaraghavan <go...@apache.org>]

> where res_url like '%mts.ru%'
...
> where res_url like '%mts_ru%'
...
> Why '_' wildcard decrease perfomance?

Because it misses the fast path by just one "_".

ORC vectorized reader has a zero-copy check for 3 patterns - prefix,
suffix and middle.

That means "https://%", "%.html", "%mts.ru%" will hit the fast path -
which uses StringExpr::equal() which JITs into the following.

https://issues.apache.org/jira/secure/attachment/12748720/string-intrinsic-
sse.png


In Hive-2.0, you can mix these up too to get "https:%mts%.html" in a
ChainedChecker.


Anything other than these 3 cases becomes a Regex and takes the slow path.

The pattern you mentioned gets rewritten into ".*mts.ru.*" and the inner
loop has a new String() as the input to the matcher + matcher.matches() in
it.

I've put in some patches recently which rewrite it Lazy regexes like
".?*mts.ru.?*", so the regex DFA will be smaller (HIVE-13196).

That improves the case where the pattern is found, but does nothing to
improve the performance of the new String() GC garbage.

Cheers,
Gopal