nested documents performance anomaly

classic Classic list List threaded Threaded
3 messages Options
Reply | Threaded
Open this post in threaded view
|

nested documents performance anomaly

roiwexler

Hi, 
we're at the process of testing Solr for its indexing speed which is very impotent to our application. 
we've witnessed strange behavior that we wish to understand before using it. 
when we indexed 1M docs it took about 63 seconds but when we indexed the same documents only now we've nested them as 1000 parented with 1000 child documents each, it took only 27 seconds. 

we know that Lucene don't support nested documents for it has a flat object model, and we do see that in fact it does index each of the child documents as a separate document. 

we have tests shows that we get the same results in case we index all documents flat (without childs) or when we index them as 1000 parents with 1000 nested documents each. 

do we miss something here? 
why does it behave like that? 
what kind of constraints does child documents have, or what is the price we pay to get this better index speed? 
we're trying to establish if this is a valid way to get a better performance in index speed.. 

any help will be appreciated. 

 

Reply | Threaded
Open this post in threaded view
|

Re: nested documents performance anomaly

Dale Richardson
Hi Roi,
My understanding of how the nested relationship is implemented in Lucene is that the child document references are physically stored in the same index segment as the parent document reference.  For normal queries which index segment a document reference is stored in is completely transparent to the query result, but the block join operator used for parent-child joins takes advantage of this low-level detail to provide for super-fast joins between parent and child documents.  A trade off for this technique is that the relevant index segment needs to be re-written when any part of the parent-child relationship changes.

I suspect that if you are writing all the children documents for a parent, you are helpfully batching up all updates to a single index segment into a single update, with the subsequent increase in speed.

The constraints that apply in return for this speed boost is that you must have all the children document ready to write in one go, and the index updates are likely done in a single transaction for each parent (i.e. all or none).  I suspect (but have not tested the fact) that indexing/storing 1000 child documents to a 1000 parent documents one document at a time would actually be slower than just indexing 1 million documents 1 document at a time.  

I hope this increases your understanding of the situation.

Regards,
Dale.

From: Roi Wexler <[hidden email]>
Sent: Sunday, 14 April 2019 6:59 AM
To: [hidden email]
Subject: nested documents performance anomaly
 

Hi, 
we're at the process of testing Solr for its indexing speed which is very impotent to our application. 
we've witnessed strange behavior that we wish to understand before using it. 
when we indexed 1M docs it took about 63 seconds but when we indexed the same documents only now we've nested them as 1000 parented with 1000 child documents each, it took only 27 seconds. 

we know that Lucene don't support nested documents for it has a flat object model, and we do see that in fact it does index each of the child documents as a separate document. 

we have tests shows that we get the same results in case we index all documents flat (without childs) or when we index them as 1000 parents with 1000 nested documents each. 

do we miss something here? 
why does it behave like that? 
what kind of constraints does child documents have, or what is the price we pay to get this better index speed? 
we're trying to establish if this is a valid way to get a better performance in index speed.. 

any help will be appreciated. 

 

Reply | Threaded
Open this post in threaded view
|

Re: nested documents performance anomaly

Jeff Wartes

 

This is more a solr-user conversation, but one other possibility is that for the “1M docs test” you sent 1M insert requests, and for the “1000 parent doc test” you sent 1000 insert requests.

Batching multiple documents into a single insert request will yield *much* better throughput, and the nested-doc approach essentially forces you to do that as a side effect of how the insert request is structured.

 

So basically Dale’s theory, but applied to the HTTP level instead of the segment level.

 

Some other random tips for indexing speed:

  • For hard commits, set openSearcher=false
  • For soft commits, set the commit interval as large as you can stand.

 

-Jeff Wartes

 

From: Dale Richardson <[hidden email]>
Reply-To: "[hidden email]" <[hidden email]>
Date: Sunday, April 14, 2019 at 3:58 AM
To: "[hidden email]" <[hidden email]>
Subject: Re: nested documents performance anomaly

 

Hi Roi,

My understanding of how the nested relationship is implemented in Lucene is that the child document references are physically stored in the same index segment as the parent document reference.  For normal queries which index segment a document reference is stored in is completely transparent to the query result, but the block join operator used for parent-child joins takes advantage of this low-level detail to provide for super-fast joins between parent and child documents.  A trade off for this technique is that the relevant index segment needs to be re-written when any part of the parent-child relationship changes.

 

I suspect that if you are writing all the children documents for a parent, you are helpfully batching up all updates to a single index segment into a single update, with the subsequent increase in speed.

 

The constraints that apply in return for this speed boost is that you must have all the children document ready to write in one go, and the index updates are likely done in a single transaction for each parent (i.e. all or none).  I suspect (but have not tested the fact) that indexing/storing 1000 child documents to a 1000 parent documents one document at a time would actually be slower than just indexing 1 million documents 1 document at a time.  

 

I hope this increases your understanding of the situation.

 

Regards,

Dale.


From: Roi Wexler <[hidden email]>
Sent: Sunday, 14 April 2019 6:59 AM
To: [hidden email]
Subject: nested documents performance anomaly

 

Hi, 
we're at the process of testing Solr for its indexing speed which is very impotent to our application. 
we've witnessed strange behavior that we wish to understand before using it. 
when we indexed 1M docs it took about 63 seconds but when we indexed the same documents only now we've nested them as 1000 parented with 1000 child documents each, it took only 27 seconds. 

we know that Lucene don't support nested documents for it has a flat object model, and we do see that in fact it does index each of the child documents as a separate document. 

we have tests shows that we get the same results in case we index all documents flat (without childs) or when we index them as 1000 parents with 1000 nested documents each. 

do we miss something here? 
why does it behave like that? 
what kind of constraints does child documents have, or what is the price we pay to get this better index speed? 
we're trying to establish if this is a valid way to get a better performance in index speed.. 

any help will be appreciated.