Solr + Hadoop = Big Data Love
Many people use the Hadoop open source project to process large data sets because it’s a great solution for scalable, reliable data processing workflows. Hadoop is by far the most popular system for handling big data, with companies using massive clusters to store and process petabytes of data on thousands of servers.
Since it emerged from the Nutch open source web crawler project in 2006, Hadoop has grown in every way imaginable – users, developers, associated projects (aka the “Hadoop ecosystem”).
Starting at roughly the same time, the Solr open source project has become the most widely used search solution on planet Earth. Solr wraps the API-level indexing and search functionality of Lucene with a RESTful API, GUI, and lots of useful administrative and data integration functionality.
The interesting thing about combining these two open source projects is that you can use Hadoop to crunch the data, and then serve it up in Solr. And we’re not talking about just free-text search; Solr can be used as a key-value store (i.e. a NoSQL database) via its support for range queries.
Even on a single server, Solr can easily handle many millions of records (“documents” in Lucene lingo). Even better, Solr now supports sharding and replication via the new, cutting-edge SolrCloud functionality.
I started using Hadoop & Solr about five years ago, as key pieces of the Krugle code search startup I co-founded in 2005.
Back then, Hadoop was still part of the Nutch web crawler we used to extract information about open source projects. And Solr was fresh out of the oven, having just been released as open source by CNET.
At Bixo Labs we use Hadoop, Solr, Cascading, Mahout, and many other open source technologies to create custom data processing workflows. The web is a common source of our input data, which we crawl using the Bixo open source project.
During a web crawl, the state of the crawl is contained in something commonly called a “crawl DB”. For broad crawls, this has to be something that works with billions of records, since you need one entry for each known URL. Each “record” has the URL as the key, and contains important state information such as the time and result of the last request.
For Hadoop-based crawlers such as Nutch and Bixo, the crawl DB is commonly kept in a set of flat files, where each file is a Hadoop “SequenceFile”. These are just a packed array of serialized key/value objects.
Sometimes we need to poke at this data, and here’s where the simple flat-file structure creates a problem. There’s no easy way run queries against the data, but we can’t store it in a traditional database since billions of records + RDBMS == pain and suffering.
Here is where scalable NoSQL solutions shine. For example, the Nutch project is currently re-factoring this crawl DB layer to allow plugging in HBase. Other options include Cassandra, MongoDB, CouchDB, etc.
But for simple analytics and exploration on smaller datasets, a Solr-based solution works and is easier to configure. Plus you get useful and surprising fun functionality like facets, geospatial queries, range queries, free-form text search, and lots of other goodies for free.
So what exactly would such a Hadoop + Solr system look like?
As mentioned previously, in this example our input data comes from a Bixo web crawler’s CrawlDB, with one entry for each known URL. But the input data could just as easily be log files, or records from a traditional RDBMS, or the output of another data processing workflow.
The key point is that we’re going to take a bunch of input data, (optionally) munge it into a more useful format, and then generate a Lucene index that we access via Solr.
For the uninitiated, Hadoop implements both a distributed file system (aka “HDFS”) and an execution layer that supports the map-reduce programming model.
Typically data is loaded and transformed during the map phase, and then combined/saved during the reduce phase. In our example, the map phase reads in Hadoop compressed SequenceFiles that contain the state of our web crawl, and our reduce phase write out Lucene indexes.
The focus of this article isn’t on how to write Hadoop map-reduce jobs, but I did want to show you the code that implements the guts of the job. Note that it’s not typical Hadoop key/value manipulation code, which is painful to write, debug, and maintain. Instead we use Cascading, which is an open source workflow planning and data processing API that creates Hadoop jobs from shorter, more representative code.
The snippet below reads SequenceFiles from HDFS, and pipes those records into a sink (output) that stores them using a LuceneScheme, which in turn saves records as Lucene documents in an index.
Tap source = new Hfs(new SequenceFile(CRAWLDB_FIELDS), inputDir);
Pipe urlPipe = new Pipe("crawldb urls");
urlPipe = new Each(urlPipe, new ExtractDomain());
Tap sink = new Hfs(new LuceneScheme(SOLR_FIELDS,
FlowConnector fc = new FlowConnector();
fc.connect(source, sink, urlPipe).complete();
We defined CRAWLDB_FIELDS and SOLR_FIELDS to be the set of input and output data elements, using names like “url” and “status”. We take advantage of the Lucene Scheme that we’ve created for Cascading, which lets us easily map from Cascading’s view of the world (records with fields) to Lucene’s index (documents with fields). We don’t have a Cascading Scheme that directly supports Solr (wouldn’t that be handy?), but we can make-do for now since we can do simple analysis for this example.
We indexed all of the fields so that we can perform queries against them. Only the status message contains normal English text, so that’s the only one we have to analyze (i.e., break the text up into terms using spaces and other token delimiters). In addition, the ExtractDomain operation pulls the domain from the URL field and builds a new Solr field containing just the domain. This will allow us to do queries against the domain of the URL as well as the complete URL.
We could also have chosen to apply a custom analyzer to the URL to break it into several pieces (i.e., protocol, domain, port, path, query parameters) that could have been queried individually.
Running the Hadoop Job
For simplicity and pay-as-you-go, it’s hard to beat Amazon’s EC2 and Elastic Mapreduce offerings for running Hadoop jobs. You can easily spin up a cluster of 50 servers, run your job, save the results, and shut it down – all without needing to buy hardware or pay for IT support.
There are many ways to create and configure a Hadoop cluster; for us, we’re very familiar with the (modified) EC2 Hadoop scripts that you can find in the Bixo distribution. Step-by-step instructions are available at http://openbixo.org/documentation/running-bixo-in-ec2/
The code for this article is available via GitHub, at http://github.com/bixolabs/hadoop2solr. The README displayed on that page contains step-by-step instructions for building and running the job.
After the job is done, we’ll copy the resulting index out of the Hadoop distributed file system (HDFS) and onto the Hadoop cluster’s master server, then kill off the one slave we used. The Hadoop master is now ready to be configured as our Solr server.
On the Solr side of things, we need to create a schema that matches the index we’re generating. The key section of our schema.xml file is where we define the fields.
<field name="url" type="string" indexed="true" stored="true" />
<field name="domain" type="string" indexed="true" stored="false" />
<field name="status" type="string" indexed="true" stored="true" />
<field name="statustime" type="string" indexed="true" stored="true" />
<field name="statusmsg" type="simpletext" indexed="true" stored="true" />
These fields have a one-to-one correspondence with the SOLR_FIELDS we defined in our Hadoop workflow. They also need to use the same Lucene settings as what we defined in the static IndexWorkflow.java STORE_SETTINGS and INDEX_SETTINGS.
Once we have this defined, all that’s left is to set up a server that we can use. To keep it simple, we’ll use the single EC2 instance in Amazon’s cloud (m1.large) that we used as our master for the Hadoop job, and run the simple Solr search server that relies on embedded Jetty to provide the webapp container.
Similar to the Hadoop job, step-by-step instructions are in the README for the hadoop2solr project on GitHub. But in a nutshell, we’ll copy and unzip a Solr 1.4.1 setup on the EC2 server, do the same for our custom Solr configuration, create a symlink to the index, and then start it running with:
Giving it a Try
Now comes the interesting part. Since we opened up the default Jetty port used by Solr (8983) on this EC2 instance, we can directly access Solr’s handy admin console by pointing our browser at http://<ec2-public-name>:8983/solr/admin
% cd solr
% java -Dsolr.solr.home=../solr-conf -Dsolr.data.dir=../solr-data -jar start.jar
From here we can run queries against Solr:
We can also use curl to talk to the server via HTTP requests:
The response is XML by default. Below is an example of the response from the above request, where we found 2,546 matches in 94ms.
Now here’s what I find amazing. For an index of 82 million documents, running on a fairly wimpy box (EC2 m1.large = 2 virtual cores), the typical response time for a simple query like “status:FETCHED” is only 400 milliseconds, to find 9M documents. Even a complex query such as (status not FETCHED and not UNFETCHED) only takes six seconds.
Obviously we could use beefier boxes. If we switched to something like m1.xlarge (15GB of memory, 4 virtual cores) then it’s likely we could handle upwards of 200M “records” in our Solr index and still get reasonable response times.
If we wanted to scale beyond a single box, there are a number of solutions. Even out of the box Solr supports sharding, where your HTTP request can specify multiple servers to use in parallel.
Finally, the Katta open source project supports Lucene-level distributed search, with many of the features needed for production quality distributed search that have not yet been added to SolrCloud.
The combination of Hadoop and Solr makes it easy to crunch lots of data and then quickly serve up the results via a fast, flexible search & query API. Because Solr supports query-style requests, it’s suitable as a NoSQL replacement for traditional databases in many situations, especially when the size of the data exceeds what is reasonable with a typical RDBMS.
Solr has some limitations that you should be aware of, specifically:
· Updating the index works best as a batch operation. Individual records can be updated, but each commit (index update) generates a new Lucene segment, which will impact performance.
· Current support for replication, fail-over, and other attributes that you’d want in a production-grade solution aren’t yet there in SolrCloud. If this matters to you, consider Katta instead.
· Many SQL queries can’t be easily mapped to Solr queries.
The code for this article is available via GitHub, at http://github.com/bixolabs/hadoop2solr. The README displayed on that page contains additional technical details.