lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Mark Miller <markrmil...@gmail.com>
Subject Re: Solr updateRequestHandler and performance vs. atomicity
Date Tue, 25 May 2010 09:20:37 GMT
Okay, makes sense.

Perhaps one easier way to explore this is the aux index idea, but only
use stored fields - that gets us lucenes commit stuff for free,
without analysis.  Then there is just the more difficult part of
ensuring transfer from this mini index to the main index for indexing
on commit.

I'd def open a jira issue for this functionality. You will still pay
for committing so often (frequent fsync is costly, especially on some
fs) but I'm sure you can pay a lot less than currently.

On Tuesday, May 25, 2010,  <karl.wright@nokia.com> wrote:
> The reason for this is simple.  LCF keeps track of which documents it has handed off
to Solr, and has a fairly involved mechanism for making sure that every document LCF *thinks*
got there, actually does.  It even uses a mechanism akin to a 2-phase commit to make sure
that its internal records and those of the downstream index are never out of synch.
>
> Now, along comes Solr, and the system loses a good deal of its resilience, because there
is a chance that somebody or something will kick Solr after a document (or a set of documents)
has been transmitted to it, but LCF will have no awareness of this situation at all, and will
thus never try to fix the problem on the next job run (or whatever).  So instead of automatic
resilience, you get one of two possible solutions:
>
> (1) Manual intervention.  Somebody has to manually inform LCF of the Solr hiccup, and
LCF thus will have to invalidate all documents it ever sent to Solr (because it doesn't know
what documents could have been affected).
> (2) A solr commit on every post.  This slows down LCF significantly, because each document
post takes something like 10x as long to do.
>
> Does this help?
> Karl
>
> -----Original Message-----
> From: ext Mark Miller [mailto:markrmiller@gmail.com]
> Sent: Monday, May 24, 2010 4:40 PM
> To: dev@lucene.apache.org
> Subject: Re: Solr updateRequestHandler and performance vs. atomicity
>
> Indexing a doc won't be as fast as raw disk IO. But you won't be doing
> just raw disk IO to guarantee acceptance. And that will have a cost and
> complexity that really makes me wonder if its worth the speed advantage.
> For very large documents with complex analyzers...perhaps. But its not
> going to be an easily implementable feature (if its a true guarantee).
> And its still got to involve logs and/or fsync and all that.
>
> The reasoning for this is not ringing a bell - can you elaborate on the
> motivations?
>
> Is this so that you can commit on every doc? Every few docs?
>
> I can def see how this would be desirable in general, but just to be
> clear on your motivations.
>
>
> - Mark
>
> On 5/24/10 10:03 PM, karl.wright@nokia.com wrote:
>> Hi Mark,
>>
>> Unfortunately, indexing performance *is* of concern, otherwise I'd already be committing
on every post.
>>
>> If your guess is correct, you are basically saying that adding a document to an index
in Solr/Lucene is just as fast as writing that file directly to the disk.  Because, obviously,
if we want guaranteed delivery, that's what we'd have to do.  But I think this is worth the
experiment - Solr/Lucene may be fast, but I have doubts that it can perform as well as raw
disk I/O and still manage to do anything in the way of document analysis or (heaven forbid)
text extraction.
>>
>>
>>
>> -----Original Message-----
>> From: ext Mark Miller [mailto:markrmiller@gmail.com]
>> Sent: Monday, May 24, 2010 3:33 PM
>> To: dev@lucene.apache.org
>> Subject: Re: Solr updateRequestHandler and performance vs. atomicity
>>
>> On 5/24/10 3:10 PM, karl.wright@nokia.com wrote:
>>> Hi all,
>>> It seems to me that the "commit" logic in the Solr updateRequestHandler
>>> (or wherever the logic is actually located) conflates two different
>>> semantics. One semantic is what you need to do to make the index process
>>> perform well. The other semantic is guaranteed atomicity of document
>>> reception by Solr.
>>> In particular, it would be nice to be able to post documents in such a
>>> way that you can guarantee that the document is permanently in Solr's
>>> queue, safe in the event of a Solr restart, etc., even if the document
>>> has not yet been "committed".
>>> This issue came up in the LCF talk that I gave, and I initially thought
>>> that separating the two kinds of events would necessarily be an LCF
>>> change, but the more I thought about it the more I realized that other
>>> Solr indexing clients may also benefit from such a separation.
>>> Does anyone agree? Where should this logic properly live?
>>> Thanks,
>>> Karl
>>
>> Its an interesting idea - but I think you would likely pay a similar
>> cost to guarantee reception as you would to commit (also, I'm not sure
>> Lucene guarantees it - it works for consistency, but I'm not so sure it
>> achieves durability).
>>
>> I can think of two things offhand -
>>
>> Perhaps store the text and use fsync to quasi guarantee acceptance -
>> then index from the store on the commit.
>>
>> Another simpler idea if only the separation is important and not the
>> performance - index to another side index, taking advantage of Lucene's
>> current commit functionality, and then use addIndex to merge to the main
>> index on commit.
>>
>> Just spit balling though.
>>
>> I think this would obviously need to be an optional mode.
>>
>
>
> --
> - Mark
>
> http://www.lucidimagination.com
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
> For additional commands, e-mail: dev-help@lucene.apache.org
>
>

-- 
- Mark

http://www.lucidimagination.com

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


Mime
View raw message