Solr 4: disable compression on stored fields: how to actually configure custom codec? - solr

Question

The short question is :
I want to disable stored field compression on Solr 4.3.0 index. After reading :
http://blog.jpountz.net/post/35667727458/stored-fields-compression-in-lucene-4-1
http://wiki.apache.org/solr/SimpleTextCodecExample
http://www.opensourceconnections.com/2013/06/05/build-your-own-lucene-codec/
I've decided to follow the path described there, and make my own codec. I'm pretty sure I've followed all the steps, however, when I actually try to use my codec (affectionatelly named "UncompressedStorageCodec"), I get the following error in Solr log:
java.lang.IllegalArgumentException: A SPI class of type org.apache.lucene.codecs.PostingsFormat with name 'UncompressedStorageCodec' does not exist. You need to add the corresponding JAR file supporting this SPI to your classpath.
The current classpath supports the following names: [Pulsing41, SimpleText, Memory, BloomFilter, Direct, Lucene40, Lucene41]
at org.apache.lucene.util.NamedSPILoader.lookup(NamedSPILoader.java:109)
From the output I get that Solr is not picking up the jar with my custom codec, and I don't get why?
Here's all the horriffic details:
I've created a class like this:
public class UncompressedStorageCodec extends FilterCodec {
private final StoredFieldsFormat fieldsFormat = new Lucene40StoredFieldsFormat();
protected UncompressedStorageCodec() {
super("UncompressedStorageCodec", new Lucene42Codec());
}
#Override
public StoredFieldsFormat storedFieldsFormat() {
return fieldsFormat;
}
}
in package: "fr.company.project.solr.transformers.utils"
The FQDN of "FilterCodec" is: "org.apache.lucene.codecs.FilterCodec"
I've created a basic jar file out of this (exported it as jar from Eclipse).
The Solr installation I'm using to test this is the basic Solr 4.3.0 unzipped, and started via it's embedded Jetty server and using the example core.
I've placed my jar with the codec in [solrDir]\dist
In:
[solrDir]\example\solr\myCore\conf\solrconfig.xml
I've added the line:
<lib dir="../../../dist/" regex="myJarWithCodec-1.10.1.jar" />
Then in the schema.xml file, I've declared some fieldTypes that should use this codec like so:
<fieldType name="string" class="solr.StrField" sortMissingLast="true" omitNorms="true" postingsFormat="UncompressedStorageCodec"/>
<fieldType name="string_lowercase" class="solr.TextField" positionIncrementGap="100" omitNorms="true" postingsFormat="UncompressedStorageCodec">
<!--...-->
</fieldType>
Now, if I use the DataImportHandler component to import some data into Solr, at commit time it tells me:
java.lang.IllegalArgumentException: A SPI class of type org.apache.lucene.codecs.PostingsFormat with name 'UncompressedStorageCodec' does not exist. You need to add the corresponding JAR file supporting this SPI to your classpath.
The current classpath supports the following names: [Pulsing41, SimpleText, Memory, BloomFilter, Direct, Lucene40, Lucene41]
at org.apache.lucene.util.NamedSPILoader.lookup(NamedSPILoader.java:109)
What I find strange is that the above mentioned codec jar also contains some Transformers for the DataImportHandler component. And those are picked up fine. Also, other jars placed in the dist folder (and declared in the same way in solrconfig.xml), like the jdbc driver are picked up fine. I'm guessing that for the codec there's this SPI thingy which loads things differentlly, and there's somethign he's missing...
I've also tried placing the codec jar in:
[solrDir]\example\solr-webapp\webapp\WEB-INF\lib\
as well as inside the WEB-INF\lib folder of the solr.war file, which is found in:
[solrDir]\example\webapps\
but I'm still getting the same error.
So basically, my question is, what's missing so that my codec jar is picked up by Solr?
Thanks

Answer 1

I'm going to answer this question myself since it sort of become moot due to some benchmarks I've made: long story short, I had arrived at the (wrong) conclusion that for really large stored fields, Solr 3.x and 4.0 (without field compression) is faster than Solr 4.1 and above (with field compression). However that was mostly due to some errors in my benchmarks. After repeating them I've obtained results where when you go from non-compressed to compressed fields even for very large stored fields, the index time is between 0% and 15% slower, which is really not bad at all, considering that afterwards queries on the compressed fields indexes are 10-20% times faster (the document fetching part).
Also, here's some remarks on how to speed up indexing:
Use the DataImportHandler plugin. It bypasses the Solr Rest (HTTP based) API and writes directly to the Lucene index.
Check out said plugins sources to see how it accomplishes this, and do your own plugin if the DataImportHandler doesn't meet your needs
If for whatever reason you want to stick to the Solr Rest API, use ConcurrentUpdateSolrServer and play around with the queue size and number of threads parameters. It will normally be a lot faster (up to 200% in my case) than the basic HttpSolrServer.
Don't forget to enable the javabin data serialization like this:
ConcurrentUpdateSolrServer solrServer = new ConcurrentUpdateSolrServer("http://some.solr.host:8983/solr", 100, 4);
solrServer.setRequestWriter(new BinaryRequestWriter());
I'm explicitly showing the code because I believe there migth be a small bug here:
If you look at the ConcurrentUpdateSolrServer constructor, you'll see that by default it already sets the request writer to binary:
//the ConcurrentUpdateSolrServer initializes HttpSolrServer objects using this constructor:
public HttpSolrServer(String baseURL, HttpClient client) {
this(baseURL, client, new BinaryResponseParser());
}
However after debugging I've noticed that if you don't explicitly call the setWriter method with the Binary writer argument, it will still use the XmlSerializer.
Going from XML to Binary serialization reduces the size of my documents about 3 times as they are being sent to the server. This makes my index times for this case about 150-200% faster.

Answer 2

I have recently tried and succeeded to get something very similar to work. The only difference is that I want to enable the best compression instead of no compression, and Solr defaults to the fastest compression. I also got the "SPI class [...] does not exist" error at some point, and here is what I have found out from various articles, including the ones you have linked to.
Lucene uses SPI to find the codec classes to load. Lucene requires the list of codec classes be declared in the file "org.apache.lucene.codecs.Codec", and the file must be on the class path. To get Solr to load the file: When you create your JAR file "myJarWithCodec-1.10.1.jar", make sure that it contains a file at "META-INF/services/org.apache.lucene.codecs.Codec". The file should have one full class name per line, like this:
org.apache.lucene.codecs.lucene3x.Lucene3xCodec
org.apache.lucene.codecs.lucene40.Lucene40Codec
org.apache.lucene.codecs.lucene41.Lucene41Codec
org.apache.lucene.codecs.lucene42.Lucene42Codec
fr.company.project.solr.transformers.utils.UncompressedStorageCodec
And in solrconfig.xml, replace:
<codecFactory class="solr.SchemaCodecFactory" />
with:
<codecFactory class="fr.company.project.solr.transformers.utils.UncompressedStorageCodec" />
You might also need to remove postingsFormat="UncompressedStorageCodec" from schema.xml if Solr complains. I think this particular parameter is for specifying the postings format, not the codec. Hope it helps.