Re: Dataimport performance

[Glen Newton <gl...@gmail.com>]

Hi,

LuSqlv2 beta comes out in the next few weeks, and is designed to
address this issue (among others).

LuSql original (http://lab.cisti-icist.nrc-cnrc.gc.ca/cistilabswiki/index.php/LuSql
now moved to: https://code.google.com/p/lusql/) is a JDBC-->Lucene
high performance loader.

You may have seen my posts on this list suggesting LuSql as high
performance alternative to DIH, for a subset of use cases.

LuSqlV2 has evolved into a full extract-transform-load (ETL) high
performance engine, focusing on many of the issues of interest to the
Lucene/SOLR community.
It has a pipelined, pluggable, multithreaded architecture.
It is basically: pluggable source --> 0 or more pluggable filters -->
pluggable sink

Source plugins implemented:
- JDBC, Lucene, SOLR (SolrJ), BDB, CSV, RMI, Java Serialization
Sink plugins implemented:
- JDBC, Lucene, SOLR (SolrJ), BDB, XML, RMI, Java Serialization, Tee,
NullSink [I am working on a memcached Sink]
A number of different filters implemented (i.e. get PDF file from
filesystem based on SQL field and convert & get test, etc) including:
BDBJoinFIlter, JDBCJoinFilter

--

This particular problem is one of the unit tests I have: given a
simple database of:
1- table Name
2- table City
3- table nameCityJoin
4- table Job
5- table nameJobJoin

run a JDBC-->BDB LuSql instance each for of City+nameCityJoin and
Job+nameJobJoin; then run a JDBC-->SolrJ on table Name, adding 2
BDBJoinFIlters, each which take the BDB generated earlier and do the
join (you just tell the filters which field from the JDBC-generated to
use against the BDB key).

So your use case use a larger example of this.

Also of interest:
- Java RMI (Remote Method Invocation): both an RMISink(Server) and
RMISource(Client) are implemented. This means you can set up N
machines which are doing something, and have one or more clients (on
their own machines) that are pulling this data and doing something
with it. For example, JDBC-->PDFToTextFilter-->RMI (converting PDF
files to text based on the contents of a SQL database, with text files
in the file system): basically doing some heavy lifting, and then
start up an RMI-->SolrJ (or Lucene) which is a client to the N PDF
converting machines, doing only the Lucene/SOLR indexing. The client
does a pull when it needs more data. You can have N servers x M
clients! Oh, string fields length > 1024 are automatically gzipped by
the RMI Sink(Server), to reduce network (at the cost of cpu:
selectable). I am looking into RMI alternatives, like Thrift, ProtoBuf
for my next Source/Sinks to implement. Another example is the reverse
use case: when the indexing is more expensive getting the data.
Example: One JDBC-->RMISink(Server) instance, N
RMISource(Client)-->Lucene instances; this allows multiple Lucenes to
be fed from a single JDBC source, across machines.

- TeeSink: the Tee sink hides N sinks, so you can split the pipeline
into multiple Sinks. I've used it to send the same content to Lucene
as well as BDB in one fell swoop. Can you say index and content store
in one step?

I am working on cleaning up the code, writing docs (I made the mistake
of making great docs for LusqlV1, so I have work to do...!), and
making a couple more tests.

I will announce the beta on this and the Lucene list.

If you have any questions, please contact me.

Thanks,
Glen Newton
http://zzzoot.blogspot.com

--> Old LuSql benchmarks:
http://zzzoot.blogspot.com/2008/11/lucene-231-vs-24-benchmarks-using-lusql.html

On Thu, Dec 16, 2010 at 12:04 PM, Dyer, James <Ja...@ingrambook.com> wrote:
> We have ~50 long-running SQL queries that need to be joined and denormalized.  Not all of the queries are to the same db, and some data comes from fixed-width data feeds.  Our current search engine (that we are converting to SOLR) has a fast disk-caching mechanism that lets you cache all of these data sources and then it will join them locally prior to indexing.
>
> I'm in the process of developing something similar for DIH that uses the Berkley db to do the same thing.  Its good enough that I can do nightly full re-indexes of all our data while developing the front-end, but it is still very rough.  Possibly I would like to get this refined enough to eventually submit as a jira ticket / patch as it seems this is a somewhat common problem that needs solving.
>
> Even with our current search engine, the join & denormalize step is always the longest-running part of the process.  However, I have it running fairly fast by partitioning the data by a modulus of the primary key and then running several jobs in parallel.  The trick is not to get I/O bound.  Things run fast if you can set it up to maximize CPU.
>
> James Dyer
> E-Commerce Systems
> Ingram Content Group
> (615) 213-4311
>
>
> -----Original Message-----
> From: Ephraim Ofir [mailto:EphraimO@icq.com]
> Sent: Thursday, December 16, 2010 3:04 AM
> To: solr-user@lucene.apache.org
> Subject: RE: Dataimport performance
>
> Check out http://mail-archives.apache.org/mod_mbox/lucene-solr-user/201008.mbox/%3C9F8B39CB3B7C6D4594293EA29CCF438B01702F22@ICQ-MAIL.icq.il.office.aol.com%3E
> This approach of not using sub entities really improved our load time.
>
> Ephraim Ofir
>
> -----Original Message-----
> From: Robert Gründler [mailto:robert@dubture.com]
> Sent: Wednesday, December 15, 2010 4:49 PM
> To: solr-user@lucene.apache.org
> Subject: Re: Dataimport performance
>
> i've benchmarked the import already with 500k records, one time without the artists subquery, and one time without the join in the main query:
>
>
> Without subquery: 500k in 3 min 30 sec
>
> Without join and without subquery: 500k in 2 min 30.
>
> With subquery and with left join:   320k in 6 Min 30
>
>
> so the joins / subqueries are definitely a bottleneck.
>
> How exactly did you implement the custom data import?
>
> In our case, we need to de-normalize the relations of the sql data for the index,
> so i fear i can't really get rid of the join / subquery.
>
>
> -robert
>
>
>
>
>
> On Dec 15, 2010, at 15:43 , Tim Heckman wrote:
>
>> 2010/12/15 Robert Gründler <ro...@dubture.com>:
>>> The data-config.xml looks like this (only 1 entity):
>>>
>>>      <entity name="track" query="select t.id as id, t.title as title, l.title as label from track t left join label l on (l.id = t.label_id) where t.deleted = 0" transformer="TemplateTransformer">
>>>        <field column="title" name="title_t" />
>>>        <field column="label" name="label_t" />
>>>        <field column="id" name="sf_meta_id" />
>>>        <field column="metaclass" template="Track" name="sf_meta_class"/>
>>>        <field column="metaid" template="${track.id}" name="sf_meta_id"/>
>>>        <field column="uniqueid" template="Track_${track.id}" name="sf_unique_id"/>
>>>
>>>        <entity name="artists" query="select a.name as artist from artist a left join track_artist ta on (ta.artist_id = a.id) where ta.track_id=${track.id}">
>>>          <field column="artist" name="artists_t" />
>>>        </entity>
>>>
>>>      </entity>
>>
>> So there's one track entity with an artist sub-entity. My (admittedly
>> rather limited) experience has been that sub-entities, where you have
>> to run a separate query for every row in the parent entity, really
>> slow down data import. For my own purposes, I wrote a custom data
>> import using SolrJ to improve the performance (from 3 hours to 10
>> minutes).
>>
>> Just as a test, how long does it take if you comment out the artists entity?
>
>



-- 

-

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