Re: O/S Search Comparisons
My testing experience has shown around 100 to be good for things like Wikipedia, etc. That is an interesting point to think about in regards to paying the cost once optimize is undertaken and may be worth exploring more. I also wonder how partial optimizes may help. The Javadocs say: Determines how often segment indices are merged by addDocument(). With * smaller values, less RAM is used while indexing, and searches on * unoptimized indices are faster, but indexing speed is slower. With larger * values, more RAM is used during indexing, and while searches on unoptimized * indices are slower, indexing is faster. Thus larger values ( 10) are best * for batch index creation, and smaller values ( 10) for indices that are * interactively maintained. * * pNote that this method is a convenience method: it * just calls mergePolicy.setMergeFactor as long as * mergePolicy is an instance of [EMAIL PROTECTED] LogMergePolicy}. * Otherwise an IllegalArgumentException is thrown./p * * pThis must never be less than 2. The default value is 10. I'd like to append to the last line to say something like: Empirical testing suggests a maximum value around 100, but this depends on the collection. Really large values ( 100) are discouraged. On Dec 18, 2007, at 12:10 AM, Doron Cohen wrote: On Dec 18, 2007 2:38 AM, Mark Miller [EMAIL PROTECTED] wrote: For the data that I normally work with (short articles), I found that the sweet spot was around 80-120. I actually saw a slight decrease going above that...not sure if that held forever though. That was testing on an earlier release (I think 2.1?). However, if you want to test searching it would seem that you are going to want to optimize the index. I have always found that whatever I save by changing the merge factor is paid back when you optimize. I have not scientifically tested this, but found it to be the case in every speed test I ran. This is an interesting thing to me for this test. Do you test with a full optimize for indexing? If you don't, can you really test the search performance with the advantage of a full optimize? So, if you are going to optimize, why mess with the merge factor? It may still play a small role, but at best I think its a pretty weak lever. I had similar experience - set merge factor to ~maxint and optimized at the end, and felt like it was the same (never meassured though). In fact, with the new concurrent merges, I think it should be faster to merge on the fly? (One comment - it is important to set back merge factor to a reasonable number before the final optimize, otherwise you hit OutOfMem due to so many segments being merged at once.) - Mark Grant Ingersoll wrote: I did hear back from the authors. Some of the issues were based on values chosen for mergeFactor (10,000) I think, but there also seemed to be some questions about parsing the TREC collection. It was split out into individual files, as opposed to trying to stream in the documents like we do with Wikipedia, so I/O overhead may be an issue. At the time, 1.9.1 did not have much TREC support, so splitting files is probably the easiest way to do it. There indexing code was based off the demo and some LIA reading. They thought they would try Lucene again when 2.3 comes out. From our end, I think we need to improve the docs around mergeFactor. We generally just say bigger is better, but my understanding is there is definitely a limit to this (100?? Maybe 1000) so we should probably suggest that in the docs. And, of course, I think the new contrib/benchmark has support for reading TREC (although I don't know if it handles streaming it) such that I think it shouldn't be a problem this time around. Yes it does streaming - TREC compressed files are read with GZIPInputStream on demand - next doc's text is read/parsed only when the indexer requests it, and the indexable document is created, no doc files are created on disk. At any rate, I think we are for the most part doing the right things. Anyone have any thoughts on advice about an upper bound for mergeFactor? Cheers, Grant On Dec 10, 2007, at 2:54 PM, Mike Klaas wrote: On 8-Dec-07, at 10:04 PM, Doron Cohen wrote: +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Solr has well-tested code that provides all this functionality and more (except for automatically spawning extra indexing threads, which I agree would be a useful addition). It does heavily depend on 1.5's java.util.concurrent package, though. Many people seem like using Solr as an embedded library layer on top of Lucene to do it all in-process, as
Re: O/S Search Comparisons
I did hear back from the authors. Some of the issues were based on values chosen for mergeFactor (10,000) I think, but there also seemed to be some questions about parsing the TREC collection. It was split out into individual files, as opposed to trying to stream in the documents like we do with Wikipedia, so I/O overhead may be an issue. At the time, 1.9.1 did not have much TREC support, so splitting files is probably the easiest way to do it. There indexing code was based off the demo and some LIA reading. They thought they would try Lucene again when 2.3 comes out. From our end, I think we need to improve the docs around mergeFactor. We generally just say bigger is better, but my understanding is there is definitely a limit to this (100?? Maybe 1000) so we should probably suggest that in the docs. And, of course, I think the new contrib/ benchmark has support for reading TREC (although I don't know if it handles streaming it) such that I think it shouldn't be a problem this time around. At any rate, I think we are for the most part doing the right things. Anyone have any thoughts on advice about an upper bound for mergeFactor? Cheers, Grant On Dec 10, 2007, at 2:54 PM, Mike Klaas wrote: On 8-Dec-07, at 10:04 PM, Doron Cohen wrote: +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Solr has well-tested code that provides all this functionality and more (except for automatically spawning extra indexing threads, which I agree would be a useful addition). It does heavily depend on 1.5's java.util.concurrent package, though. Many people seem like using Solr as an embedded library layer on top of Lucene to do it all in-process, as well. -Mike - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] -- Grant Ingersoll http://lucene.grantingersoll.com Lucene Helpful Hints: http://wiki.apache.org/lucene-java/BasicsOfPerformance http://wiki.apache.org/lucene-java/LuceneFAQ - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
For the data that I normally work with (short articles), I found that the sweet spot was around 80-120. I actually saw a slight decrease going above that...not sure if that held forever though. That was testing on an earlier release (I think 2.1?). However, if you want to test searching it would seem that you are going to want to optimize the index. I have always found that whatever I save by changing the merge factor is paid back when you optimize. I have not scientifically tested this, but found it to be the case in every speed test I ran. This is an interesting thing to me for this test. Do you test with a full optimize for indexing? If you don't, can you really test the search performance with the advantage of a full optimize? So, if you are going to optimize, why mess with the merge factor? It may still play a small role, but at best I think its a pretty weak lever. - Mark Grant Ingersoll wrote: I did hear back from the authors. Some of the issues were based on values chosen for mergeFactor (10,000) I think, but there also seemed to be some questions about parsing the TREC collection. It was split out into individual files, as opposed to trying to stream in the documents like we do with Wikipedia, so I/O overhead may be an issue. At the time, 1.9.1 did not have much TREC support, so splitting files is probably the easiest way to do it. There indexing code was based off the demo and some LIA reading. They thought they would try Lucene again when 2.3 comes out. From our end, I think we need to improve the docs around mergeFactor. We generally just say bigger is better, but my understanding is there is definitely a limit to this (100?? Maybe 1000) so we should probably suggest that in the docs. And, of course, I think the new contrib/benchmark has support for reading TREC (although I don't know if it handles streaming it) such that I think it shouldn't be a problem this time around. At any rate, I think we are for the most part doing the right things. Anyone have any thoughts on advice about an upper bound for mergeFactor? Cheers, Grant On Dec 10, 2007, at 2:54 PM, Mike Klaas wrote: On 8-Dec-07, at 10:04 PM, Doron Cohen wrote: +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Solr has well-tested code that provides all this functionality and more (except for automatically spawning extra indexing threads, which I agree would be a useful addition). It does heavily depend on 1.5's java.util.concurrent package, though. Many people seem like using Solr as an embedded library layer on top of Lucene to do it all in-process, as well. -Mike - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] -- Grant Ingersoll http://lucene.grantingersoll.com Lucene Helpful Hints: http://wiki.apache.org/lucene-java/BasicsOfPerformance http://wiki.apache.org/lucene-java/LuceneFAQ - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
On Dec 18, 2007 2:38 AM, Mark Miller [EMAIL PROTECTED] wrote: For the data that I normally work with (short articles), I found that the sweet spot was around 80-120. I actually saw a slight decrease going above that...not sure if that held forever though. That was testing on an earlier release (I think 2.1?). However, if you want to test searching it would seem that you are going to want to optimize the index. I have always found that whatever I save by changing the merge factor is paid back when you optimize. I have not scientifically tested this, but found it to be the case in every speed test I ran. This is an interesting thing to me for this test. Do you test with a full optimize for indexing? If you don't, can you really test the search performance with the advantage of a full optimize? So, if you are going to optimize, why mess with the merge factor? It may still play a small role, but at best I think its a pretty weak lever. I had similar experience - set merge factor to ~maxint and optimized at the end, and felt like it was the same (never meassured though). In fact, with the new concurrent merges, I think it should be faster to merge on the fly? (One comment - it is important to set back merge factor to a reasonable number before the final optimize, otherwise you hit OutOfMem due to so many segments being merged at once.) - Mark Grant Ingersoll wrote: I did hear back from the authors. Some of the issues were based on values chosen for mergeFactor (10,000) I think, but there also seemed to be some questions about parsing the TREC collection. It was split out into individual files, as opposed to trying to stream in the documents like we do with Wikipedia, so I/O overhead may be an issue. At the time, 1.9.1 did not have much TREC support, so splitting files is probably the easiest way to do it. There indexing code was based off the demo and some LIA reading. They thought they would try Lucene again when 2.3 comes out. From our end, I think we need to improve the docs around mergeFactor. We generally just say bigger is better, but my understanding is there is definitely a limit to this (100?? Maybe 1000) so we should probably suggest that in the docs. And, of course, I think the new contrib/benchmark has support for reading TREC (although I don't know if it handles streaming it) such that I think it shouldn't be a problem this time around. Yes it does streaming - TREC compressed files are read with GZIPInputStream on demand - next doc's text is read/parsed only when the indexer requests it, and the indexable document is created, no doc files are created on disk. At any rate, I think we are for the most part doing the right things. Anyone have any thoughts on advice about an upper bound for mergeFactor? Cheers, Grant On Dec 10, 2007, at 2:54 PM, Mike Klaas wrote: On 8-Dec-07, at 10:04 PM, Doron Cohen wrote: +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Solr has well-tested code that provides all this functionality and more (except for automatically spawning extra indexing threads, which I agree would be a useful addition). It does heavily depend on 1.5's java.util.concurrent package, though. Many people seem like using Solr as an embedded library layer on top of Lucene to do it all in-process, as well. -Mike - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] -- Grant Ingersoll http://lucene.grantingersoll.com Lucene Helpful Hints: http://wiki.apache.org/lucene-java/BasicsOfPerformance http://wiki.apache.org/lucene-java/LuceneFAQ - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
On Dec 7, 2007, at 3:01 PM, Mark Miller wrote: Yes, and even if they did not use the stock defaults, I would bet there would be complaints about what was done wrong at every turn. This seems like a very difficult thing to do. How long does it take to fully learn how to correctly utilize each search engine for the task at hand? I am sure longer than these busy men could possibly take. It seems that such a comparison could only be done legitimately if experts for each search engine set up the indexing/ searching processes. Even then the results seem like they could be difficult to measure...eg was each search engine configured so that they would only break on spaces for indexing and do nothing else special at all? So many small settings and knowledge need to ensure each engine is on level ground... This is why I have called on NIST/TREC to open source their collections. Until then, Lucene and the other O/S search engines will be reliant on those contributors who have access to them, which is spotty at best. (And, yes, I know, TREC is not the be all, end all of IR evaluations, but it is a common ground for doing research) See http://www.gossamer-threads.com/lists/lucene/java-dev/52022?search_string=TREC;#52022 -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
On 8-Dec-07, at 10:04 PM, Doron Cohen wrote: +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Solr has well-tested code that provides all this functionality and more (except for automatically spawning extra indexing threads, which I agree would be a useful addition). It does heavily depend on 1.5's java.util.concurrent package, though. Many people seem like using Solr as an embedded library layer on top of Lucene to do it all in- process, as well. -Mike - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
Well, at some point the answer is use Solr. I think Lucene should stay focused on being a good search library/component, and server level capabilities should be handled by Solr or the application layer on top of Lucene. That said, I still think there is a need for a layer that handles/ hides threading, does search refreshing/warming, etc., on the Lucene side. Actually, LuceneIndexAccessor (LUCENE-390) is already a step in that direction. Mike On Dec 9, 2007, at 1:29 AM, robert engels wrote: This is along the lines of what I have tried to get the Lucene community to adopt for a long time. If you want to take Lucene to the next level, it needs a server implementation. Only with this can you get efficient locks, caching, transactions, which leads to more efficient indexing and searching. IMO, the shared storage nature of Lucene is its biggest weakness. A lot of changes have been made to improve this, when it probably just needs to be dropped. If you have a network, it is really no different to communicate with processes rather than storage. On Dec 9, 2007, at 12:04 AM, Doron Cohen wrote: Grant Ingersoll [EMAIL PROTECTED] wrote on 08/12/2007 16:02:31: On Dec 8, 2007, at 4:51 AM, Michael McCandless wrote: Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) +1 There is still something for us to learn improve in Lucene, even if the comparison is necessarily apples/oranges or unfair. Lucene was listed as not having Result Excerpt which isn't really fair, though it is true you have to pull in contrib/highlighter to enable it. Yeah, I noted that mentally, but didn't think it was a big deal since not everyone wants it. The other thing is, some of it comes down to how you structure your content. I think a lot of people use metadata fields to provide enough summary info about a document. Did it crash on the 10 GB? I thought it said that it just took way to long (7 times the best or something). Frankly, either case is suspect. Last summer I indexed about 5 million docs with a total size at the *very* least of 10 GB on my 3 year old desktop. It didn't take much more than 8 hours to index and searches where still lightning fast. Maybe they forgot to give the JVM more than the default amount of RAM g The paper just said ht://Dig and Lucene degraded considerably their indexing time, and we excluded them from the final comparison. Maybe Lucene just hit a very large segment merge and the author incorrectly thought something had gone wrong since the addDocument call was taking incredibly long? In which case the new default ConcurrentMergeScheduler should improve that. I would expect Lucene 2.3 to now have an advantage in that it makes use of concurrency in the hardware, out of the box, whereas likely other older engines are single threaded. Yep. I've also thought about creating a simple optional threaded layer on top of IndexWriter which uses multiple threads to add documents, under the hood. Such a class would expose all of the methods of IndexWriter (would feel just like IndexWriter), except calls to add/ updateDocument would drop into a queue which multiple threads (maintained by this class) would pull from and execute. This would then let Lucene make use of even more concurrency ... and saves the complexity of application writers having to manage threads above Lucene. +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Here's some of my wishes: 1. Better Demo 2. Alternate scoring algorithms (which implies indexing too) that perform at or near the same level as the current ones +1 3. A way of announcing improvements to Interfaces such that we have better ability to add methods to interfaces, knowing full well it will break some people. Same goes for deprecated. In this day and age of agile programming, it seems a bit restrictive to me that we wait 1+ years (the average time between major releases) to remove what we consider to be cruft in our code or add new capabilities to interfaces. I would suggest we announce a deprecated method, version it, mark it to when it is going away (i.e. This will be removed in version 2.6) and then do so in that version. So, if we deprecate something in 2.3, we could, assuming consecutive numbered releases, remove it in 2.5. This would
Re: O/S Search Comparisons
On Dec 8, 2007, at 4:51 AM, Michael McCandless wrote: Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) +1 There is still something for us to learn improve in Lucene, even if the comparison is necessarily apples/oranges or unfair. Lucene was listed as not having Result Excerpt which isn't really fair, though it is true you have to pull in contrib/highlighter to enable it. Yeah, I noted that mentally, but didn't think it was a big deal since not everyone wants it. The other thing is, some of it comes down to how you structure your content. I think a lot of people use metadata fields to provide enough summary info about a document. Did it crash on the 10 GB? I thought it said that it just took way to long (7 times the best or something). Frankly, either case is suspect. Last summer I indexed about 5 million docs with a total size at the *very* least of 10 GB on my 3 year old desktop. It didn't take much more than 8 hours to index and searches where still lightning fast. Maybe they forgot to give the JVM more than the default amount of RAM g The paper just said ht://Dig and Lucene degraded considerably their indexing time, and we excluded them from the final comparison. Maybe Lucene just hit a very large segment merge and the author incorrectly thought something had gone wrong since the addDocument call was taking incredibly long? In which case the new default ConcurrentMergeScheduler should improve that. I would expect Lucene 2.3 to now have an advantage in that it makes use of concurrency in the hardware, out of the box, whereas likely other older engines are single threaded. Yep. I've also thought about creating a simple optional threaded layer on top of IndexWriter which uses multiple threads to add documents, under the hood. Such a class would expose all of the methods of IndexWriter (would feel just like IndexWriter), except calls to add/ updateDocument would drop into a queue which multiple threads (maintained by this class) would pull from and execute. This would then let Lucene make use of even more concurrency ... and saves the complexity of application writers having to manage threads above Lucene. +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Here's some of my wishes: 1. Better Demo 2. Alternate scoring algorithms (which implies indexing too) that perform at or near the same level as the current ones 3. A way of announcing improvements to Interfaces such that we have better ability to add methods to interfaces, knowing full well it will break some people. Same goes for deprecated. In this day and age of agile programming, it seems a bit restrictive to me that we wait 1+ years (the average time between major releases) to remove what we consider to be cruft in our code or add new capabilities to interfaces. I would suggest we announce a deprecated method, version it, mark it to when it is going away (i.e. This will be removed in version 2.6) and then do so in that version. So, if we deprecate something in 2.3, we could, assuming consecutive numbered releases, remove it in 2.5. This would presumably move things up a bit to about the 6 mos. time range. Just a thought... :-) -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
Grant Ingersoll [EMAIL PROTECTED] wrote on 08/12/2007 16:02:31: On Dec 8, 2007, at 4:51 AM, Michael McCandless wrote: Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) +1 There is still something for us to learn improve in Lucene, even if the comparison is necessarily apples/oranges or unfair. Lucene was listed as not having Result Excerpt which isn't really fair, though it is true you have to pull in contrib/highlighter to enable it. Yeah, I noted that mentally, but didn't think it was a big deal since not everyone wants it. The other thing is, some of it comes down to how you structure your content. I think a lot of people use metadata fields to provide enough summary info about a document. Did it crash on the 10 GB? I thought it said that it just took way to long (7 times the best or something). Frankly, either case is suspect. Last summer I indexed about 5 million docs with a total size at the *very* least of 10 GB on my 3 year old desktop. It didn't take much more than 8 hours to index and searches where still lightning fast. Maybe they forgot to give the JVM more than the default amount of RAM g The paper just said ht://Dig and Lucene degraded considerably their indexing time, and we excluded them from the final comparison. Maybe Lucene just hit a very large segment merge and the author incorrectly thought something had gone wrong since the addDocument call was taking incredibly long? In which case the new default ConcurrentMergeScheduler should improve that. I would expect Lucene 2.3 to now have an advantage in that it makes use of concurrency in the hardware, out of the box, whereas likely other older engines are single threaded. Yep. I've also thought about creating a simple optional threaded layer on top of IndexWriter which uses multiple threads to add documents, under the hood. Such a class would expose all of the methods of IndexWriter (would feel just like IndexWriter), except calls to add/ updateDocument would drop into a queue which multiple threads (maintained by this class) would pull from and execute. This would then let Lucene make use of even more concurrency ... and saves the complexity of application writers having to manage threads above Lucene. +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Here's some of my wishes: 1. Better Demo 2. Alternate scoring algorithms (which implies indexing too) that perform at or near the same level as the current ones +1 3. A way of announcing improvements to Interfaces such that we have better ability to add methods to interfaces, knowing full well it will break some people. Same goes for deprecated. In this day and age of agile programming, it seems a bit restrictive to me that we wait 1+ years (the average time between major releases) to remove what we consider to be cruft in our code or add new capabilities to interfaces. I would suggest we announce a deprecated method, version it, mark it to when it is going away (i.e. This will be removed in version 2.6) and then do so in that version. So, if we deprecate something in 2.3, we could, assuming consecutive numbered releases, remove it in 2.5. This would presumably move things up a bit to about the 6 mos. time range. Just a thought... :-) -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
This is along the lines of what I have tried to get the Lucene community to adopt for a long time. If you want to take Lucene to the next level, it needs a server implementation. Only with this can you get efficient locks, caching, transactions, which leads to more efficient indexing and searching. IMO, the shared storage nature of Lucene is its biggest weakness. A lot of changes have been made to improve this, when it probably just needs to be dropped. If you have a network, it is really no different to communicate with processes rather than storage. On Dec 9, 2007, at 12:04 AM, Doron Cohen wrote: Grant Ingersoll [EMAIL PROTECTED] wrote on 08/12/2007 16:02:31: On Dec 8, 2007, at 4:51 AM, Michael McCandless wrote: Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) +1 There is still something for us to learn improve in Lucene, even if the comparison is necessarily apples/oranges or unfair. Lucene was listed as not having Result Excerpt which isn't really fair, though it is true you have to pull in contrib/highlighter to enable it. Yeah, I noted that mentally, but didn't think it was a big deal since not everyone wants it. The other thing is, some of it comes down to how you structure your content. I think a lot of people use metadata fields to provide enough summary info about a document. Did it crash on the 10 GB? I thought it said that it just took way to long (7 times the best or something). Frankly, either case is suspect. Last summer I indexed about 5 million docs with a total size at the *very* least of 10 GB on my 3 year old desktop. It didn't take much more than 8 hours to index and searches where still lightning fast. Maybe they forgot to give the JVM more than the default amount of RAM g The paper just said ht://Dig and Lucene degraded considerably their indexing time, and we excluded them from the final comparison. Maybe Lucene just hit a very large segment merge and the author incorrectly thought something had gone wrong since the addDocument call was taking incredibly long? In which case the new default ConcurrentMergeScheduler should improve that. I would expect Lucene 2.3 to now have an advantage in that it makes use of concurrency in the hardware, out of the box, whereas likely other older engines are single threaded. Yep. I've also thought about creating a simple optional threaded layer on top of IndexWriter which uses multiple threads to add documents, under the hood. Such a class would expose all of the methods of IndexWriter (would feel just like IndexWriter), except calls to add/ updateDocument would drop into a queue which multiple threads (maintained by this class) would pull from and execute. This would then let Lucene make use of even more concurrency ... and saves the complexity of application writers having to manage threads above Lucene. +1 I have been thinking about this too. Solr clearly demonstrates the benefits of this kind of approach, although even it doesn't make it seamless for users in the sense that they still need to divvy up the docs on the app side. Would be nice if this layer also took care of searchers/readers refreshing warming. Here's some of my wishes: 1. Better Demo 2. Alternate scoring algorithms (which implies indexing too) that perform at or near the same level as the current ones +1 3. A way of announcing improvements to Interfaces such that we have better ability to add methods to interfaces, knowing full well it will break some people. Same goes for deprecated. In this day and age of agile programming, it seems a bit restrictive to me that we wait 1+ years (the average time between major releases) to remove what we consider to be cruft in our code or add new capabilities to interfaces. I would suggest we announce a deprecated method, version it, mark it to when it is going away (i.e. This will be removed in version 2.6) and then do so in that version. So, if we deprecate something in 2.3, we could, assuming consecutive numbered releases, remove it in 2.5. This would presumably move things up a bit to about the 6 mos. time range. Just a thought... :-) -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
Yeah, I wasn't too excited over it and I certainly didn't lose any sleep over it, but there are some interesting things of note in there concerning Lucene, including the claim that it fell over on indexing WT10g docs (page 40) and I am always looking for ways to improve things. Overall, I think Lucene held up pretty well in the evaluation, and I know how suspect _any_ evaluation is given the myriad ways of doing search. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, I don't think we can dismiss them out of hand. So regardless of the tests being right or wrong, they are worth either addressing the failures in Lucene or the failures in the test such that we make sure we are properly educating our users on how best to use Lucene. I emailed the authors asking for information on how the test was run etc., so we'll see if anything comes of it. On Dec 7, 2007, at 12:04 PM, robert engels wrote: I wouldn't get too excited over this. Once again, it does not seem the evaluator understands the nature of GC based systems, and the memory statistics are quite out of whack. But it is hard to tell because there is no data on how memory consumption was actually measured. A far better way of measuring memory consumption is to cap the process at different levels (max ram sizes), and compare the performance at each level. There is also fact that a process takes memory from disk cache, and visa versa, that heavily affects search performance, etc. Since there is no detailed data (that I could find) about system configuration, etc. the results are highly suspect. There is also no mention of performance on multi-processor systems. Some systems (like Lucene) pay a penalty to support multi-processing (both in Java and Lucene), and only realize this benefit when operating in a multi-processor environment. Based on the shear speed of XMLSearch and Zettair those seem likely candidates to inspect their design. On Dec 7, 2007, at 7:03 AM, Grant Ingersoll wrote: Was wondering if people have seen http://wrg.upf.edu/WRG/dctos/Middleton-Baeza.pdf Has some interesting comparisons. Obviously, the comparison of Lucene indexing is done w/ 1.9 so it probably needs to be done again. Just wondering if people see any opportunities to improve Lucene from it.I am going to try and contact the authors to see if I can get what there setup values were (mergeFactor, Analyzer, etc.) as I think it would be interesting to run the tests again on 2.3. -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
Did it crash on the 10 GB? I thought it said that it just took way to long (7 times the best or something). Frankly, either case is suspect. Last summer I indexed about 5 million docs with a total size at the *very* least of 10 GB on my 3 year old desktop. It didn't take much more than 8 hours to index and searches where still lightning fast. Maybe they forgot to give the JVM more than the default amount of RAM g - Mark Grant Ingersoll wrote: All true and good points. Lucene held up quite nicely in the search aspect (at least perf. wise) and I generally don't think making these kinds of comparisons are all that useful (we call it apple and oranges in English :-) ). What I am trying to get at is if this paper was just about Lucene and never mentioned a single other system, what, if anything, can we take from it that can help us make Lucene better. I know, for instance, from my own personal experience, that 2.3 is somewhere in the range of 3-5+ times faster than 2.2 (which I know is faster than 1.9). That being said, the paper clearly states that Lucene was not capable of doing the WT10g docs because performance degraded too much. Now, I know Lucene is pretty darn capable of a lot of things and people are using it to do web search, etc. at very large scales (I have personally talked w/ people doing it). So, what I worry about is that either we are: a) missing something in our defaults setup b) missing something in our docs and our education efforts, or c) we are missing some capability in our indexing such that it is crashing Now, what is to be done? It may well be nothing, but I just want to make sure we are comfortable with that decision or whether it is worth asking for a volunteer who has access to the WT10g docs to go have a look at it and see what happens. I personally don't have access to these docs, otherwise I would try it out. What we don't want to happen is for potential supporters/contributors to read that paper and say Lucene isn't for me because of this. Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) Cheers, Grant PS: Samir, any chance of contributing back your ranking algorithms? :-) On Dec 7, 2007, at 5:41 PM, Samir Abdou wrote: There is an expression in French that says comparer des pommes et des poires which literally means to compare apples and pears. That's what this paper is about. For my point of view, such a comparison would be interesting only if a cross analysis of different criterions (for example, retrieval effectiveness (aka search quality), search time, indexing time, index size, query language, index structure, and so on...) is done. Comparing different systems based only on one criterion is not well-grounded. There is always a kind of trade-off: for example, beside other parameters (ranking algorithm, frequencies statistics, document structure, etc.), indexing with zettair is much faster than indexing with lucene but if we consider searching time lucene is better than zettair. Why? Because of many reasons but probably zettair hasn't the complex document structure of lucene besides the ranking algorithm (Okapi BM25 vs. tf-idf). Some systems computes and stores the scores at indexing time which make them faster at searching time but less flexible if you want to change/implement a new ranking algorithm. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, If we consider the search quality, that's simply not true if we know how to implement in Lucene popular ranking algorithm such OkapiBM25 (at least). I've been working with Lucene for four years now, all experiments of my thesis have been done using Lucene (with many adaptations to implement the most recent ranking algorithm including different language model, divergence from randomness, etc.). I also participated to major IR campaigns (NTCIR, CLEF and TREC) and the results are not bad at all (see http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings5/data/CLIR/NTCIR5 -OV-CLIR-KishidaK.pdf for NTCIR-5 or http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings6/NTCIR/NTCIR6-OVE RVIEW.pdf for NTCIR-6, for CLEF have a look at http://www.clef-campaign.org/2006/working_notes/workingnotes2006/dinunzioOCL EF2006.pdf, ...) for other information search the web ;-) Samir -Message d'origine- De : Mark Miller [mailto:[EMAIL PROTECTED] Envoyé : vendredi 7 décembre 2007 21:01 À : java-dev@lucene.apache.org Objet : Re: O/S Search Comparisons Yes, and even if they did not use the stock defaults, I would bet
Re: O/S Search Comparisons
All true and good points. Lucene held up quite nicely in the search aspect (at least perf. wise) and I generally don't think making these kinds of comparisons are all that useful (we call it apple and oranges in English :-) ). What I am trying to get at is if this paper was just about Lucene and never mentioned a single other system, what, if anything, can we take from it that can help us make Lucene better. I know, for instance, from my own personal experience, that 2.3 is somewhere in the range of 3-5+ times faster than 2.2 (which I know is faster than 1.9). That being said, the paper clearly states that Lucene was not capable of doing the WT10g docs because performance degraded too much. Now, I know Lucene is pretty darn capable of a lot of things and people are using it to do web search, etc. at very large scales (I have personally talked w/ people doing it). So, what I worry about is that either we are: a) missing something in our defaults setup b) missing something in our docs and our education efforts, or c) we are missing some capability in our indexing such that it is crashing Now, what is to be done? It may well be nothing, but I just want to make sure we are comfortable with that decision or whether it is worth asking for a volunteer who has access to the WT10g docs to go have a look at it and see what happens. I personally don't have access to these docs, otherwise I would try it out. What we don't want to happen is for potential supporters/contributors to read that paper and say Lucene isn't for me because of this. Sometimes, when something like this comes up, it gives you the opportunity to take a step back and ask what are the things we really want Lucene to be going forward (the New Year is good for this kind of assessment as well) What are it's strengths and weaknesses? What can we improve in the short term and what needs to improve in the longer term? Maybe it's just that time of year to send out your Lucene Wish List... :-) Cheers, Grant PS: Samir, any chance of contributing back your ranking algorithms? :-) On Dec 7, 2007, at 5:41 PM, Samir Abdou wrote: There is an expression in French that says comparer des pommes et des poires which literally means to compare apples and pears. That's what this paper is about. For my point of view, such a comparison would be interesting only if a cross analysis of different criterions (for example, retrieval effectiveness (aka search quality), search time, indexing time, index size, query language, index structure, and so on...) is done. Comparing different systems based only on one criterion is not well-grounded. There is always a kind of trade-off: for example, beside other parameters (ranking algorithm, frequencies statistics, document structure, etc.), indexing with zettair is much faster than indexing with lucene but if we consider searching time lucene is better than zettair. Why? Because of many reasons but probably zettair hasn't the complex document structure of lucene besides the ranking algorithm (Okapi BM25 vs. tf- idf). Some systems computes and stores the scores at indexing time which make them faster at searching time but less flexible if you want to change/ implement a new ranking algorithm. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, If we consider the search quality, that's simply not true if we know how to implement in Lucene popular ranking algorithm such OkapiBM25 (at least). I've been working with Lucene for four years now, all experiments of my thesis have been done using Lucene (with many adaptations to implement the most recent ranking algorithm including different language model, divergence from randomness, etc.). I also participated to major IR campaigns (NTCIR, CLEF and TREC) and the results are not bad at all (see http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings5/data/CLIR/NTCIR5 -OV-CLIR-KishidaK.pdf for NTCIR-5 or http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings6/NTCIR/NTCIR6-OVE RVIEW.pdf for NTCIR-6, for CLEF have a look at http://www.clef-campaign.org/2006/working_notes/workingnotes2006/dinunzioOCL EF2006.pdf, ...) for other information search the web ;-) Samir -Message d'origine- De : Mark Miller [mailto:[EMAIL PROTECTED] Envoyé : vendredi 7 décembre 2007 21:01 À : java-dev@lucene.apache.org Objet : Re: O/S Search Comparisons Yes, and even if they did not use the stock defaults, I would bet there would be complaints about what was done wrong at every turn. This seems like a very difficult thing to do. How long does it take to fully learn how to correctly utilize each search engine for the task at hand? I am sure longer than these busy men could possibly take. It seems that such a comparison could only be done legitimately if experts for each search engine set up the indexing/searching
RE: O/S Search Comparisons
There is an expression in French that says comparer des pommes et des poires which literally means to compare apples and pears. That's what this paper is about. For my point of view, such a comparison would be interesting only if a cross analysis of different criterions (for example, retrieval effectiveness (aka search quality), search time, indexing time, index size, query language, index structure, and so on...) is done. Comparing different systems based only on one criterion is not well-grounded. There is always a kind of trade-off: for example, beside other parameters (ranking algorithm, frequencies statistics, document structure, etc.), indexing with zettair is much faster than indexing with lucene but if we consider searching time lucene is better than zettair. Why? Because of many reasons but probably zettair hasn't the complex document structure of lucene besides the ranking algorithm (Okapi BM25 vs. tf-idf). Some systems computes and stores the scores at indexing time which make them faster at searching time but less flexible if you want to change/implement a new ranking algorithm. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, If we consider the search quality, that's simply not true if we know how to implement in Lucene popular ranking algorithm such OkapiBM25 (at least). I've been working with Lucene for four years now, all experiments of my thesis have been done using Lucene (with many adaptations to implement the most recent ranking algorithm including different language model, divergence from randomness, etc.). I also participated to major IR campaigns (NTCIR, CLEF and TREC) and the results are not bad at all (see http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings5/data/CLIR/NTCIR5 -OV-CLIR-KishidaK.pdf for NTCIR-5 or http://research.nii.ac.jp/ntcir/workshop/OnlineProceedings6/NTCIR/NTCIR6-OVE RVIEW.pdf for NTCIR-6, for CLEF have a look at http://www.clef-campaign.org/2006/working_notes/workingnotes2006/dinunzioOCL EF2006.pdf, ...) for other information search the web ;-) Samir -Message d'origine- De : Mark Miller [mailto:[EMAIL PROTECTED] Envoyé : vendredi 7 décembre 2007 21:01 À : java-dev@lucene.apache.org Objet : Re: O/S Search Comparisons Yes, and even if they did not use the stock defaults, I would bet there would be complaints about what was done wrong at every turn. This seems like a very difficult thing to do. How long does it take to fully learn how to correctly utilize each search engine for the task at hand? I am sure longer than these busy men could possibly take. It seems that such a comparison could only be done legitimately if experts for each search engine set up the indexing/searching processes. Even then the results seem like they could be difficult to measure...eg was each search engine configured so that they would only break on spaces for indexing and do nothing else special at all? So many small settings and knowledge need to ensure each engine is on level ground... I doubt it will ever happen, but some sort of open source search off would be pretty cool g. Then each camp could properly configure their search engine for each task. - Mark Mike Klaas wrote: There is a good chance that they were using stock indexing defaults, based on: Lucene: In the present work, the simple applications bundled with the library were used to index the collection. On 7-Dec-07, at 10:27 AM, Grant Ingersoll wrote: Yeah, I wasn't too excited over it and I certainly didn't lose any sleep over it, but there are some interesting things of note in there concerning Lucene, including the claim that it fell over on indexing WT10g docs (page 40) and I am always looking for ways to improve things. Overall, I think Lucene held up pretty well in the evaluation, and I know how suspect _any_ evaluation is given the myriad ways of doing search. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, I don't think we can dismiss them out of hand. So regardless of the tests being right or wrong, they are worth either addressing the failures in Lucene or the failures in the test such that we make sure we are properly educating our users on how best to use Lucene. I emailed the authors asking for information on how the test was run etc., so we'll see if anything comes of it. On Dec 7, 2007, at 12:04 PM, robert engels wrote: I wouldn't get too excited over this. Once again, it does not seem the evaluator understands the nature of GC based systems, and the memory statistics are quite out of whack. But it is hard to tell because there is no data on how memory consumption was actually measured. A far better way of measuring memory consumption is to cap the process at different levels (max ram sizes), and compare the performance at each level. There is also fact that a process
Re: O/S Search Comparisons
Yes, and even if they did not use the stock defaults, I would bet there would be complaints about what was done wrong at every turn. This seems like a very difficult thing to do. How long does it take to fully learn how to correctly utilize each search engine for the task at hand? I am sure longer than these busy men could possibly take. It seems that such a comparison could only be done legitimately if experts for each search engine set up the indexing/searching processes. Even then the results seem like they could be difficult to measure...eg was each search engine configured so that they would only break on spaces for indexing and do nothing else special at all? So many small settings and knowledge need to ensure each engine is on level ground... I doubt it will ever happen, but some sort of open source search off would be pretty cool g. Then each camp could properly configure their search engine for each task. - Mark Mike Klaas wrote: There is a good chance that they were using stock indexing defaults, based on: Lucene: In the present work, the simple applications bundled with the library were used to index the collection. On 7-Dec-07, at 10:27 AM, Grant Ingersoll wrote: Yeah, I wasn't too excited over it and I certainly didn't lose any sleep over it, but there are some interesting things of note in there concerning Lucene, including the claim that it fell over on indexing WT10g docs (page 40) and I am always looking for ways to improve things. Overall, I think Lucene held up pretty well in the evaluation, and I know how suspect _any_ evaluation is given the myriad ways of doing search. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, I don't think we can dismiss them out of hand. So regardless of the tests being right or wrong, they are worth either addressing the failures in Lucene or the failures in the test such that we make sure we are properly educating our users on how best to use Lucene. I emailed the authors asking for information on how the test was run etc., so we'll see if anything comes of it. On Dec 7, 2007, at 12:04 PM, robert engels wrote: I wouldn't get too excited over this. Once again, it does not seem the evaluator understands the nature of GC based systems, and the memory statistics are quite out of whack. But it is hard to tell because there is no data on how memory consumption was actually measured. A far better way of measuring memory consumption is to cap the process at different levels (max ram sizes), and compare the performance at each level. There is also fact that a process takes memory from disk cache, and visa versa, that heavily affects search performance, etc. Since there is no detailed data (that I could find) about system configuration, etc. the results are highly suspect. There is also no mention of performance on multi-processor systems. Some systems (like Lucene) pay a penalty to support multi-processing (both in Java and Lucene), and only realize this benefit when operating in a multi-processor environment. Based on the shear speed of XMLSearch and Zettair those seem likely candidates to inspect their design. On Dec 7, 2007, at 7:03 AM, Grant Ingersoll wrote: Was wondering if people have seen http://wrg.upf.edu/WRG/dctos/Middleton-Baeza.pdf Has some interesting comparisons. Obviously, the comparison of Lucene indexing is done w/ 1.9 so it probably needs to be done again. Just wondering if people see any opportunities to improve Lucene from it.I am going to try and contact the authors to see if I can get what there setup values were (mergeFactor, Analyzer, etc.) as I think it would be interesting to run the tests again on 2.3. -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
Re: O/S Search Comparisons
There is a good chance that they were using stock indexing defaults, based on: Lucene: In the present work, the simple applications bundled with the library were used to index the collection. On 7-Dec-07, at 10:27 AM, Grant Ingersoll wrote: Yeah, I wasn't too excited over it and I certainly didn't lose any sleep over it, but there are some interesting things of note in there concerning Lucene, including the claim that it fell over on indexing WT10g docs (page 40) and I am always looking for ways to improve things. Overall, I think Lucene held up pretty well in the evaluation, and I know how suspect _any_ evaluation is given the myriad ways of doing search. Still, when a well-respected researcher in the field says Lucene didn't do so hot in certain areas, I don't think we can dismiss them out of hand. So regardless of the tests being right or wrong, they are worth either addressing the failures in Lucene or the failures in the test such that we make sure we are properly educating our users on how best to use Lucene. I emailed the authors asking for information on how the test was run etc., so we'll see if anything comes of it. On Dec 7, 2007, at 12:04 PM, robert engels wrote: I wouldn't get too excited over this. Once again, it does not seem the evaluator understands the nature of GC based systems, and the memory statistics are quite out of whack. But it is hard to tell because there is no data on how memory consumption was actually measured. A far better way of measuring memory consumption is to cap the process at different levels (max ram sizes), and compare the performance at each level. There is also fact that a process takes memory from disk cache, and visa versa, that heavily affects search performance, etc. Since there is no detailed data (that I could find) about system configuration, etc. the results are highly suspect. There is also no mention of performance on multi-processor systems. Some systems (like Lucene) pay a penalty to support multi- processing (both in Java and Lucene), and only realize this benefit when operating in a multi-processor environment. Based on the shear speed of XMLSearch and Zettair those seem likely candidates to inspect their design. On Dec 7, 2007, at 7:03 AM, Grant Ingersoll wrote: Was wondering if people have seen http://wrg.upf.edu/WRG/dctos/ Middleton-Baeza.pdf Has some interesting comparisons. Obviously, the comparison of Lucene indexing is done w/ 1.9 so it probably needs to be done again. Just wondering if people see any opportunities to improve Lucene from it.I am going to try and contact the authors to see if I can get what there setup values were (mergeFactor, Analyzer, etc.) as I think it would be interesting to run the tests again on 2.3. -Grant - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] - To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
