lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Michael McCandless <luc...@mikemccandless.com>
Subject Re: Faceting : what are the limitations of Taxonomy (Separate index and hierarchical facets) and SortedSetDocValuesFacetField ( flat facets and no sidecar index) ?
Date Tue, 29 Nov 2016 22:21:14 GMT
Doc values fields are never loaded into memory; at most some small
index structures are.

When you use those fields, the bytes (for just the one doc values
field you are using) are pulled from disk, and the OS will cache them
in memory if available.

Mike McCandless

http://blog.mikemccandless.com


On Mon, Nov 28, 2016 at 6:01 AM, Chitra R <chithu.r111@gmail.com> wrote:
> Hi,
>          When opening SortedSetDocValuesReaderState at search time, whether
> the whole doc value files (.dvd & .dvm) information are loaded in memory or
> specified field information(say $facets field) alone load in memory?
>
>
>
>
> Any help is much appreciated.
>
>
> Regards,
> Chitra
>
> On Tue, Nov 22, 2016 at 5:47 PM, Chitra R <chithu.r111@gmail.com> wrote:
>>
>>
>> Kindly post your suggestions.
>>
>> Regards,
>> Chitra
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>>
>> On Sat, Nov 19, 2016 at 1:38 PM, Chitra R <chithu.r111@gmail.com> wrote:
>>>
>>> Hey, I got it clearly. Thank you so much. Could you please help us to
>>> implement it in our use case?
>>>
>>>
>>> In our case, we are having dynamic index and it is variable depth too. So
>>> flat facet is enough.No need of hierarchical facets.
>>>
>>> What I think is,
>>>
>>> Index my facet field as normal doc value field, so that no special
>>> operation (like taxonomy and sorted set doc values facet field) will be done
>>> at index time and only doc value field stores its ordinals in their
>>> respective field.
>>> At search time, I will pass query (user search query) , filter (path
>>> traversed list)  and collect the matching documents in Facetscollector.
>>> To compute facet count for the specific field, I will gather those
>>> resulted docs, then move through each segment for collecting the matching
>>> ordinals using AtomicReader.
>>>
>>>
>>> And know when I use this means, can't calculate facet count for more than
>>> one field(facet) in a search.
>>>
>>> Instead of loading all the dimensions in DocValuesReaderState (will take
>>> more time and memory) at search time, loading specific fields will take less
>>> time and memory, hope so. Kindly help to solve.
>>>
>>>
>>> It will do it in a minimal index and search cost, I think. And hope this
>>> won't put overload at index time, also at search time this will be better.
>>>
>>>
>>> Kindly post your suggestions.
>>>
>>>
>>> Regards,
>>> Chitra
>>>
>>>
>>>
>>>
>>> On Fri, Nov 18, 2016 at 7:15 PM, Michael McCandless
>>> <lucene@mikemccandless.com> wrote:
>>>>
>>>> I think you've summed up exactly the differences!
>>>>
>>>> And, yes, it would be possible to emulate hierarchical facets on top
>>>> of flat facets, if the hierarchy is fixed depth like year/month/day.
>>>>
>>>> But if it's variable depth, it's trickier (but I think still
>>>> possible).  See e.g. the Committed Paths drill-down on the left, on
>>>> our dog-food server
>>>> http://jirasearch.mikemccandless.com/search.py?index=jira
>>>>
>>>> Mike McCandless
>>>>
>>>> http://blog.mikemccandless.com
>>>>
>>>>
>>>> On Fri, Nov 18, 2016 at 1:43 AM, Chitra R <chithu.r111@gmail.com> wrote:
>>>> > case 1:
>>>> >         In taxonomy, for each indexed document, examines facet label
,
>>>> > computes their ordinals and mappings, and which will be stored in
>>>> > sidecar
>>>> > index at index time.
>>>> >
>>>> > case 2:
>>>> >         In doc values, these(ordinals) are computed at search time,
so
>>>> > there
>>>> > will be a time and memory trade-off between both cases, hope so.
>>>> >
>>>> >
>>>> > In taxonomy, building hierarchical facets at index time makes faceting
>>>> > cost
>>>> > minimal at search time than flat facets in doc values.
>>>> >
>>>> > Except (memory,time and NRT latency) , Is any another contrast between
>>>> > hierarchical and flat facets at search time?
>>>> >
>>>> >
>>>> > Kindly post your suggestions...
>>>> >
>>>> >
>>>> > Regards,
>>>> > Chitra
>>>> >
>>>> > On Thu, Nov 17, 2016 at 6:40 PM, Chitra R <chithu.r111@gmail.com>
>>>> > wrote:
>>>> >>
>>>> >> Okay. I agree with you, Taxonomy maintains and supports hierarchical
>>>> >> facets during indexing. Hope hierarchical in the sense, we might
>>>> >> index the
>>>> >> field Publish date : 2010/10/15 as Publish date: 2010 , Publish
date:
>>>> >> 2010/10 and Publish date: 2010/10/15 , their facet ordinals are
>>>> >> maintained
>>>> >> in sidecar index and it is mapped to the main index.
>>>> >>
>>>> >> For example:
>>>> >>
>>>> >>                 In search-lucene.com , I enter a term (say facet),
>>>> >> top
>>>> >> documents and their categories are displayed after performing the
>>>> >> search.
>>>> >> Say I drill down through Publish date/2010 to collect its child
>>>> >> counts and
>>>> >> after I will pass through publishdate/2010/10 to collect their child
>>>> >> counts.
>>>> >> And for each drill down, each search will be performed to collect
its
>>>> >> top
>>>> >> docs and categories.
>>>> >>
>>>> >>
>>>> >>                Even I can achieve this in flat facets by changing
the
>>>> >> drill down query.
>>>> >>
>>>> >> Am I right or missed anything? yet I don't know if I missed
>>>> >> anything...
>>>> >>
>>>> >> So What is the need of hierarchical facets? Could you please explain
>>>> >> it(hierarchical facets) in the real-world use case?
>>>> >>
>>>> >>
>>>> >> Regards,
>>>> >> Chitra
>>>> >>
>>>> >> On Wed, Nov 16, 2016 at 7:36 PM, Michael McCandless
>>>> >> <lucene@mikemccandless.com> wrote:
>>>> >>>
>>>> >>> You store dimension + string (a single value path, since it's
not
>>>> >>> hierarchical) into SSDVFF so that you can compute facet counts,
>>>> >>> either
>>>> >>> ordinary drill down counts or the drill sideways counts.
>>>> >>>
>>>> >>> You can see examples of drill sideways at
>>>> >>> http://jirasearch.mikemccandless.com, e.g. drill down on any
of
>>>> >>> those
>>>> >>> fields on the left and you don't lose the previous facet counts
for
>>>> >>> that field.
>>>> >>>
>>>> >>> Mike McCandless
>>>> >>>
>>>> >>> http://blog.mikemccandless.com
>>>> >>>
>>>> >>>
>>>> >>> On Wed, Nov 16, 2016 at 8:51 AM, Chitra R <chithu.r111@gmail.com>
>>>> >>> wrote:
>>>> >>> > Hi,
>>>> >>> >
>>>> >>> > Lucene-Drill sideways
>>>> >>> >
>>>> >>> > jira_issue:LUCENE-4748
>>>> >>> >
>>>> >>> >                                  Is this the reason( ie
Drill
>>>> >>> > sideways
>>>> >>> > makes
>>>> >>> > a very nice faceted search UI because we
>>>> >>> > don't "lose" the facet counts after drilling in) behind
storing
>>>> >>> > path
>>>> >>> > and
>>>> >>> > dimension for the given SSDVF field? Else anything?
>>>> >>> >
>>>> >>> > Regards,
>>>> >>> > Chitra
>>>> >>> >
>>>> >>> >
>>>> >>> >      Hey, thank you so much for the fast response, I agree
NRT
>>>> >>> > refresh
>>>> >>> > is
>>>> >>> > somewhat costly operations and this is the major pitfall,
suppose
>>>> >>> > we
>>>> >>> > use doc
>>>> >>> > value faceting.
>>>> >>> >
>>>> >>> >
>>>> >>> >                  While indexing SortedSetDocValuesFacetField
, it
>>>> >>> > stores
>>>> >>> > path and dimension of the given field internally. So Can
we
>>>> >>> > achieve
>>>> >>> > hierarchical facets using DrillDownQuery? Hope, purpose
of storing
>>>> >>> > path
>>>> >>> > and
>>>> >>> > dimension is to achieve hierarchical facets. If yes (ie
we can
>>>> >>> > achieve
>>>> >>> > hierarchy in SSDVFF) , so what is the need to move over
taxonomy?
>>>> >>> >  Else I missed anything?
>>>> >>> >
>>>> >>> >
>>>> >>> >                  What is the real purpose to store path
and
>>>> >>> > dimension
>>>> >>> > in
>>>> >>> > SSDVF field?
>>>> >>> >
>>>> >>> >
>>>> >>> > Kindly post your suggestions.
>>>> >>> >
>>>> >>> > Regards,
>>>> >>> > Chitra
>>>> >>> >
>>>> >>> >
>>>> >>> >
>>>> >>> > On Sat, Nov 12, 2016 at 4:03 AM, Michael McCandless
>>>> >>> > <lucene@mikemccandless.com> wrote:
>>>> >>> >>
>>>> >>> >> On Fri, Nov 11, 2016 at 5:21 AM, Chitra R <chithu.r111@gmail.com>
>>>> >>> >> wrote:
>>>> >>> >>
>>>> >>> >> >         i)Hope, when opening SortedSetDocValuesReaderState
, we
>>>> >>> >> > are
>>>> >>> >> > calculating ordinals( this will be used to calculate
facet
>>>> >>> >> > count )
>>>> >>> >> > for
>>>> >>> >> > doc
>>>> >>> >> > values field and this only made the state instance
somewhat
>>>> >>> >> > costly.
>>>> >>> >> >                       Am I right or any other
reason behind
>>>> >>> >> > that?
>>>> >>> >>
>>>> >>> >> That's correct.  It adds some latency to an NRT refresh,
and some
>>>> >>> >> heap
>>>> >>> >> used to hold the ordinal mappings.
>>>> >>> >>
>>>> >>> >> >          ii) During indexing, we are providing
facet ordinals
>>>> >>> >> > in
>>>> >>> >> > each
>>>> >>> >> > doc
>>>> >>> >> > and I think it will be useful in search side,
to calculate
>>>> >>> >> > facet
>>>> >>> >> > counts
>>>> >>> >> > only for matching docs.  otherwise, it carries
any other
>>>> >>> >> > benefits?
>>>> >>> >>
>>>> >>> >> Well, compared to the taxonomy facets, SSDV facets
don't require
>>>> >>> >> a
>>>> >>> >> separate index.
>>>> >>> >>
>>>> >>> >> But they add latency/heap usage, and they cannot do
hierarchical
>>>> >>> >> facets yet (though this could be fixed if someone just
built it).
>>>> >>> >>
>>>> >>> >> >          iii) Is SortedSetDocValuesReaderState
thread-safe (ie)
>>>> >>> >> > multiple
>>>> >>> >> > threads can call this method concurrently?
>>>> >>> >>
>>>> >>> >> Yes.
>>>> >>> >>
>>>> >>> >> Mike McCandless
>>>> >>> >>
>>>> >>> >> http://blog.mikemccandless.com
>>>> >>> >
>>>> >>> >
>>>> >>
>>>> >>
>>>> >
>>>
>>>
>>
>

---------------------------------------------------------------------
To unsubscribe, e-mail: java-user-unsubscribe@lucene.apache.org
For additional commands, e-mail: java-user-help@lucene.apache.org


Mime
View raw message