lucene-java-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Jay Yu ...@AI.SRI.COM>
Subject Re: thread safe shared IndexSearcher
Date Tue, 25 Sep 2007 18:55:10 GMT
I agree with you on the compromise aspect of the design.
In particular, I think it's hard to preload all the index accessors in 
the static init while allowing users specify the analyzer for each dir 
without requiring complicated config file ans using reflection.
So a good compromise might be abandon preload the accessors. After all, 
the accessors are cached and not created often.

Thanks!

Jay


Mark Miller wrote:
> I think its just a compromise in the design, though it could be 
> improved. You only ever want a single Writer at a time on the index. 
> Those two flags are really just hints for when a Writer is first 
> opened...should it auto-commit and should it overwrite/create...if a 
> thread tries to writer concurrently with another thread, they will 
> briefly share a Writer, but generally a new Writer is created fairly often.
> 
> The general strategy should be to pick constant values and always pass 
> them. There is an opening for the issue that you have a Writer and are 
> adding a doc, and then before releasing that Writer, another Writer from 
> another thread tries to clear the index with a create=true, and it won't 
> work. That's not a big concern though.
> 
> So the problem really is that these params control what happens when a 
> new writer is created, but your not guaranteed to be creating a Writer, 
> it may be cached. You really should pass the same autocommit flag , 
> though its not necessary. I am open to suggestions for a more coherent 
> design, but functionally, it does work. I am also thinking about how to 
> handle the Analyzer, and I think the solution (the need to init some 
> indexaccessor params) might involve all these issues.
> 
> - Mark
> 
> Jay Yu wrote:
>> Mark,
>>
>> Looking at your implementation of the DefaultIndexAccessor regarding 
>> the writer, I think there could be a problem: you have only one cached 
>> writer but the getWriter(boolean, boolean) allows 2 booleans, so 
>> ideally, you need 4 cached writer. Otherwise if one starts with a 
>> writer that over writes the existing index, then later he cannot 
>> append docs to the index.
>> Do I miss sth here or you have not finished the implementation of 
>> getWriter yet?
>>
>> Thanks!
>>
>> Jay
>>
>> Mark Miller wrote:
>>> Ah, thanks for catching that. One of the pieces I did not 
>>> finish...the keyword analyzer was placeholder code.
>>>
>>> I will take your comments into account and update the code.
>>>
>>> I have some other pieces to polish as well. Previously, I extended 
>>> and built upon the original code, but I can't give it away, so this 
>>> is my attempt at something lessor, but cleaner.
>>>
>>> Jay Yu wrote:
>>>> Thanks for the tip.
>>>> One small improvement on the IndexAccessorFactory might be to allow 
>>>> user to specify the Analyzer instead of using a default 
>>>> KeywordAnalyzer, which of course will make your static init of the 
>>>> cached accessors difficult unless you add more interfaces to the 
>>>> accessor to allow reset analyzer/Dir as in my own version.
>>>>
>>>>
>>>>
>>>>
>>>> Jay
>>>>
>>>> Mark Miller wrote:
>>>>> One final note....if you are using the IndexAccessor and you are 
>>>>> only accessing the index from one JVM, you can use the 
>>>>> NoLockFactory and save some sync cost there.
>>>>>
>>>>> Jay Yu wrote:
>>>>>> Mark,
>>>>>>
>>>>>> Great effort getting the original lucene index accessor package in

>>>>>> this shape. I am sure this will benefit a lot of people using 
>>>>>> Lucene in a multithread env.
>>>>>> I have a quick question to ask you:
>>>>>> Do you have to use the core Lucene 2.3-dev in order to use the 
>>>>>> accessor?
>>>>>>
>>>>>> I will take a look at your codes to see if I could help. I used a

>>>>>> slightly modified version of the original package in my project 
>>>>>> but it breaks some of my tests. I hope your version works better.
>>>>>>
>>>>>> Thanks a lot!
>>>>>>
>>>>>> Jay
>>>>>>
>>>>>>
>>>>>> Mark Miller wrote:
>>>>>>> I have sat down and rewrote IndexAccessor from scratch. I copied

>>>>>>> in the same reference counting logic, pruned some things, and

>>>>>>> tried to make the whole package a bit simpler to use. I have
a 
>>>>>>> few things to do, but its pretty solid already. The only major

>>>>>>> thing I'd still like to do is add an option to warm searchers

>>>>>>> before putting them in the Searcher cache. Id like to writer
some 
>>>>>>> more tests as well. Any help greatly appreciated if your 
>>>>>>> interested in using the thing.
>>>>>>>
>>>>>>>
>>>>>>> http://myhardshadow.com/indexaccessor/trunk/src/test/com/mhs/indexaccessor/SimpleSearchServer.java

>>>>>>>
>>>>>>>
>>>>>>> Here is a an example of a class that can be instantiated in one

>>>>>>> of multiple threads and read /modify a single index without 
>>>>>>> worrying about what any
>>>>>>> of the other threads are doing to the index at any given time.

>>>>>>> This is a very simple example of how to use the IndexAccessor
and 
>>>>>>> not necessarily an
>>>>>>> example of best practices. The main idea is that you get your

>>>>>>> Writer, Searcher, or Reader, and then be sure to release it as

>>>>>>> soon as your done with it
>>>>>>> in a finally block. For loading, you will want to load many docs

>>>>>>> with a Writer (batch them) before releasing it, but remember
that 
>>>>>>> Readers will not get a new view
>>>>>>> of the index until you release all of the Writers. So beware

>>>>>>> hogging a Writer unless you thats what your intending.
>>>>>>>
>>>>>>> JavaDoc:
>>>>>>> http://myhardshadow.com/indexaccessorapi/
>>>>>>>
>>>>>>> Code:
>>>>>>> http://myhardshadow.com/indexaccessor/trunk/
>>>>>>>
>>>>>>> Jar:
>>>>>>> http://myhardshadow.com/indexaccessorreleases/indexaccessor.jar
>>>>>>>
>>>>>>>
>>>>>>> Your synchronized block concerns:
>>>>>>>
>>>>>>> The synchronized blocks that control accesss to the IndexAccessor

>>>>>>> do not have a huge impact on performance. Keep in mind that all

>>>>>>> of the work is not done in a synchonrized block, just the 
>>>>>>> retrieval of the Searcher, Writer, Reader. Even if the 
>>>>>>> synchronization makes the method twice as expensive, it is still

>>>>>>> overpowered by the cost of parsing queries and searching the

>>>>>>> index. This applies with or without contention. I wrote a simple

>>>>>>> test and included the output below. You might use the IBM Lock

>>>>>>> Analyzer for Java to further analyze these costs. Trust me, this

>>>>>>> thing is speedy. Its many times better than using IndexModifier.
>>>>>>>
>>>>>>> Without Contention
>>>>>>> Just retrieve and release Searcher 100000 times
>>>>>>> ----
>>>>>>> avg time:6.3E-4 ms
>>>>>>> total time:63 ms
>>>>>>>
>>>>>>> Parse query and search on 1 doc 100000 times
>>>>>>> ----
>>>>>>> avg time:0.03107 ms
>>>>>>> total time:3107 ms
>>>>>>>
>>>>>>>
>>>>>>> With Contention (40 other threads running 80000 searches)
>>>>>>> Just retrieve and release Searcher 100000 times
>>>>>>> ----
>>>>>>> avg time:0.04643 ms
>>>>>>> total time:4643 ms
>>>>>>>
>>>>>>> Parse query and search on 1 doc 100000 times
>>>>>>> ----
>>>>>>> avg time:0.64337 ms
>>>>>>> total time:64337 ms
>>>>>>>
>>>>>>>
>>>>>>> - Mark
>>>>>>>
>>>>>>> ---------------------------------------------------------------------

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

---------------------------------------------------------------------
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