jackrabbit-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Ard Schrijvers <a.schrijv...@onehippo.com>
Subject Re: dealing with large result sets
Date Tue, 10 Apr 2012 10:01:11 GMT
On Tue, Apr 10, 2012 at 11:55 AM, Christian Stocker
<christian.stocker@liip.ch> wrote:
>
>
> On 10.04.12 11:51, Ard Schrijvers wrote:
>> On Tue, Apr 10, 2012 at 11:42 AM, Christian Stocker
>> <christian.stocker@liip.ch> wrote:
>>>
>>>
>>> On 10.04.12 11:32, Ard Schrijvers wrote:
>>>> On Tue, Apr 10, 2012 at 11:21 AM, Lukas Kahwe Smith <mls@pooteeweet.org>
wrote:
>>>>> Hi,
>>>>>
>>>>> Currently I see some big issues with queries that return large result
sets. A lot of work is not done inside Lucene, which will probably not be fixed soon (or maybe
never inside 2.x). However I think its important to do some intermediate improvements.
>>>>>
>>>>> Here are some suggestions I have. I hope we can brainstorm together on
some ideas that are feasible to get implemented in a shorter time period than waiting for
Oak:
>>>>>
>>>>> 1) there should be a way to get a count
>>>>>
>>>>> This way if I need to do a query that needs to be ordered, I can first
check if the count is too high to determine if I should even bother running the search. Aka
in most cases a search leading to 100+ results means that who ever did the search needs to
further narrow it down.
>>>>
>>>> The cpu is not spend in ordering the results: That is done quite fast
>>>> in Lucene, unless you have millions of hits
>>>
>>> I read the code and also read this
>>> https://issues.apache.org/jira/browse/JCR-2959 and it looks to me that
>>> jackrabbit always sorts the result set by itself and not in lucene (or
>>> maybe additionally). This makes it slow even if you have a limit set,
>>> because it first sorts all nodes (fetching it from the PM if necessary),
>>> then does the limit. Maybe I have missed something but real life tests
>>> showed exactly this behaviour.
>>
>> Ah, I don't know about that part: We always sticked to xpath queries :
>> Sorting is done in Lucene (more precisely, in some Lucene exensions in
>> jr, but are equally fast) for at least xpath, I am quite sure
>
> Is the search part done differently in SQL2 and XPath? Can't remember ;)

I think in some areas, but, it should pretty much result in the same
Lucene queries. I've never looked into SQL2 : But, because you pointed
at [1] and it there explicitly mentioned SQL2, and since we don't have
this problem, I thought it would be SQL2 only

[1] https://issues.apache.org/jira/browse/JCR-2959

>
>>>> The problem with getting a correct count is authorization : This total
>>>> search index count should is fast (if you try to avoid some known slow
>>>> searches). However, authorizing for example 100k+ nodes if they are
>>>> not in the jackrabbit caches is very expensive.
>>>>
>>>> Either way: You get a correct count if you make sure that you include
>>>> in your (xpath) search at least an order by clause. Then, to avoid
>>>> 100k + hits, make sure you also set a limit. For example a limit of
>>>> 501 : You can then show 50 pages of 10 hits, and if the count is 501
>>>> you state that there are at least 500+ hits
>>>
>>> That's what we do now, but it doesn't help (as said above) if we have
>>> thousends of results which have to be ordered first.
>>
>> And the second sort is also slow? The first sort is also slow with
>> Lucene, as Lucene needs to load all terms to sort on from FS in
>> memory. However, consecutive searches are fast. We don't have problems
>> for resultsets sorting for a million hits
>
> It definitively loaded all nodes from the PM before sorting it. The
> lucene part itself was fast enough, that wasn't the issue.
>
>>
>>>
>>>>
>>>> We also wanted to get around this, thus in our api hooked in a
>>>> 'getTotalSize()' which returns the Lucene unauthorized count
>>>
>>> That would help us a lot, since we currently don't use the ACLs of
>>> Jackrabbit, so the lucene count would be pretty correct for our use case.
>>
>> Yes, however, you would have to hook into jr itself to get this done
>
> Yep, saw that, that's somewhere deep in the code. That's why I didn't
> try to adress that yet
>
> chregu
>
>>
>> Regards Ard
>>
>>>
>>> chregu
>>>
>>>>
>>>>>
>>>>> I guess the most sensible thing would be to simply offer a way to do
SELECT COUNT(*) FROM ..
>>>>>
>>>>> 2) a way to automatically stop long running queries
>>>>
>>>> It is not just about 'long' . Some queries easily blow up, and bring
>>>> you app to an OOM before they can be stopped. For example jcr:like is
>>>> such a thing. Or range queries on many unique values
>>>
>>>
>>>>
>>>> Regards Ard
>>>>
>>>>>
>>>>> It would be great if one could define a timeout for queries. If a query
takes longer than X, it should just fail. This should be a global setting, but ideally it
should be possible to override this on a per query basis.
>>>>>
>>>>> 3) .. ?
>>>>>
>>>>> regards,
>>>>> Lukas Kahwe Smith
>>>>> mls@pooteeweet.org
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>
>>> --
>>> Liip AG  //  Feldstrasse 133 //  CH-8004 Zurich
>>> Tel +41 43 500 39 81 // Mobile +41 76 561 88 60
>>> www.liip.ch // blog.liip.ch // GnuPG 0x0748D5FE
>>>
>>
>>
>>
>
> --
> Liip AG  //  Feldstrasse 133 //  CH-8004 Zurich
> Tel +41 43 500 39 81 // Mobile +41 76 561 88 60
> www.liip.ch // blog.liip.ch // GnuPG 0x0748D5FE
>



-- 
Amsterdam - Oosteinde 11, 1017 WT Amsterdam
Boston - 1 Broadway, Cambridge, MA 02142

US +1 877 414 4776 (toll free)
Europe +31(0)20 522 4466
www.onehippo.com

Mime
View raw message