利用 MongoDB 和 Solr 实现全文搜索

In this article I am explaining how you can tie MongoDB and Solr together. With the help of a little helper script in PHP that uses MongoDB's replication features we automatically generate updates in Solr. We will first look at the MongoDB side and then the Solr side.

Replication

MongoDB's replication works by recording all operations done on a database in a log file, called the oplog. The local database contains a collection called oplog.rs that stores all those operations. The oplog is a capped collection which means they have a fixed maximum size. The operations that are stored in the oplog can be quite easily read, as it is just a normal collection in a normal database:

<?php  $m = new Mongo( 'mongodb://localhost:13000', array( 'replSet' => 'seta' ) );  $c = $m->local->selectCollection( 'oplog.rs' );  $r = $c->findOne( array( 'ns' => 'demo.article', 'op' => 'i' ) );  var_dump( $r );  ?>

This script connects to MongoDB and enables a replica set connection (array( 'replSet' => 'seta' )). It then selects the local database and the oplog.rs collection. We then find one document with findOne().

The ns field can be used to restrict for which databases and collections you would want to see the operations. In this case, the findOne() call limits the operations to the demo database and the article collection.

The op field tells you what sort of operation was recorded. In some cases, the records in the oplog are not exactly what they are when you send them through a driver, as they are broken up in parts. A remove() call for example will generate a delete operation for each document in the oplog. In any case, the op types are: i for inserts, u for updates and d for deletes. n is used for notices such as the reconfiguration of the replicaset.

When running the script above, you get an output like:

array(5) {    'ts' =>    class MongoTimestamp#6 (2) {      public $sec =>      int(1338716530)      public $inc =>      int(1)    }    'h' =>    int(5050582189860592111)    'op' =>    string(1) "i"    'ns' =>    string(12) "demo.article"    'o' =>    array(3) {      '_id' =>      string(8) "w4442243"      'name' =>      string(16) "Brondesbury Road"      'tags' =>      array(3) {        'highway' =>        string(9) "secondary"        'ref' =>        string(4) "B451"        'source_ref' =>        string(22) "OS OpenData StreetView"      }    }  }

And an update operation (type u) looks like:

array(6) {    'ts' =>    class MongoTimestamp#6 (2) {      public $sec =>      int(1338716534)      public $inc =>      int(1)    }    'h' =>    int(644574934620412462)    'op' =>    string(1) "u"    'ns' =>    string(12) "demo.article"    'o2' =>    array(1) {      '_id' =>      string(8) "w4442243"    }    'o' =>    array(1) {      '$set' =>      array(1) {        'source_ref' =>        string(6) "survey"      }    }  }

Tailable Cursors

One of the cool things about capped collections, is that they support something called a tailable cursor. A tailable cursor does not use an index and can only return documents in natural order. Which means the order into which documents are inserted into a collection. Hence, we can write the following PHP script to connect to the oplog and wait until new operations are made:

<?php  $m = new Mongo( 'mongodb://localhost:13000', array( 'replSet' => 'seta' ) );  $c = $m->local->selectCollection( 'oplog.rs' );  $cursor = $c->find( array( 'ns' => 'demo.article' ) );  $cursor->tailable( true );    while (true) {      if (!$cursor->hasNext()) {          // we've read all the results, exit          if ($cursor->dead()) {              break;          }          sleep(1);      } else {          var_dump( $cursor->getNext() );      }  }  ?>

In the script that we run on the command line, we use the tailable() method to configure the cursor as tailable. When we run this script however, it would just loop and use CPU time because hasNext() doesn't actually wait for new data coming in. The sleep(1) prevents the script from using 100% CPU time. From the upcoming driver releases (1.2.11/1.3.0) there will be a new awaitData() method to make hasNext() wait until there is actually new data available for the cursor. The script would then look like:

<?php  $m = new Mongo( 'mongodb://localhost:13000', array( 'replSet' => 'seta' ) );  $c = $m->local->selectCollection( 'oplog.rs' );  $cursor = $c->find( array( 'ns' => 'demo.article', 'op' => 'i' ) );  $cursor->tailable( true );  $cursor->awaitData( true );    while (true) {      if (!$cursor->hasNext()) {          // we've read all the results, exit          if ($cursor->dead()) {              break;          }      } else {          var_dump( $cursor->getNext() );      }  }  ?>

When we run this script it just sits there and loops for new information to be available for the cursor. In our case, that would be if the oplog has a new item for our article collection in the demo database.

Solr

Now we have a way to wait for updates on the MongoDB side, we need to tie this into Solr. To talk to Solr I use the Solr PECL extension from http://pecl.php.net/solr that I've installed by running pecl install solr and then added extension=solr.so into my php.ini file.

I then downloaded Solr 3.6 from http://lucene.apache.org/solr/downloads.html into install:

cd install  tar -xvzf apache-solr-3.6.0.tgz  cd apache-solr-3.6.0/example

There is a configuration file in solr/conf/schema.xml. For now we will keep the defaults, but an interesting feature is is that Solr supports dynamic fields. Basically, this allows you to store any field with the type being selected by a suffix. For example the field name name_t will automatically be used as text, and addr_housenumber_l as an integer (long).

After the configuration we can start Solr with:

java -jar start.jar

This starts a Solr instance that you can access through http://localhost:8983/solr/admin/.

Mapping Field Names

In order to store things in Solr for easy searching we need to map the fields in our document to fields that Solr understands. Solr does not support nested arrays or embedded documents so we need some translation. We also will need to add the correct type suffix. As example we use the document from earlier:

<?php  $document =          array(                  '_id' => "w4442243",                  'name' => "Brondesbury Road",                  'tags' => array(                          'highway' => "secondary",                          'ref' =>"B451",                          'source_ref' => "OS OpenData StreetView",                  )          );  ?>

We map the fields in this document to Solr fields. For everything that we don't map, we assume the text type:

<?php  class Mapper  {          static protected $map = array(                  '_id' => 'id_s',                  'name' => 'name_t',                  'tags_highway' => 'tags_highway_t',          );            static function doArray( $prefix, $array )          {                  $fields = array();                  foreach ( $array as $field => $value )                  {                          if ( is_array( $value ) )                          {                                  $fields = array_merge(                                          $fields,                                          self::doArray( $prefix . $field . '_', $value )                                  );                          }                          else if ( isset( self::$map[$prefix . $field] ) )                          {                                  $fields[self::$map[$prefix . $field]] = $value;                          }                          else                          {                                  $fields[$prefix . $field . '_t'] = $value;                          }                  }                  return $fields;          }            static function mapFields( $document )          {                  $solrFields = self::doArray( '', $document );                  return $solrFields;          }  }  ?>

Tying It All Together

Hooking this into our oplog listener is now not very difficult, all we have to do is to include our mapper class file:

<?php  include 'mapper.php';

And establish a connection to Solr:

$options = array( 'hostname' => 'localhost' );  $client = new SolrClient($options);

Keep the original tailable cursor code:

$m = new Mongo( 'mongodb://localhost:13000', array( 'replSet' => 'seta' ) );  $c = $m->local->selectCollection( 'oplog.rs' );  $cursor = $c->find( array( 'ns' => 'demo.article', 'op' => 'i' ) );  $cursor->tailable( true );  // $cursor->awaitData( true ); PHP Driver 1.2.11 and 1.3.0 only

We can then modify our loop, to get the inserted document, pass it through the mapper, create a Solr document and add it to the Solr index:

while (true) {          if (!$cursor->hasNext()) {                  // we've read all the results, exit                  if ($cursor->dead()) {                          break;                  }                  sleep(1); // Only use with PHP Driver < 1.2.11          } else {                  $document = $cursor->getNext();                  $solrFields = Mapper::mapFields( $document['o'] );                    $solrDoc = new SolrInputDocument();                  foreach ( $solrFields as $field => $value )                  {                          $solrDoc->addField( $field, $value );                  }                  $updateResponse = $client->addDocument( $solrDoc );                  print_r($updateResponse->getResponse());                  $client->commit();          }  }

And voilà, we have the Solr full text index update every time a new document is inserted into MongoDB:

db.article.insert( { '_id' : 'w4442244', 'name' : 'Brondesbury Villas', 'tags' : { 'highway' : 'secondary', 'ref' : 'B451', 'oneway' : true } } );

And the data shows up in Solr: http://localhost:8983/solr/select/?q=id%3Aw4442244&version=2.2&start=0&rows=10&indent=on

Obviously, a few things should also be taken into account still, which I will list here, but leave as an exercise for the reader:

• Documents in Solr are required to have an id field, just like MongoDB requires a _id field. Solr however, doesn't auto-generate one for you.

• You need to cast the values before adding them to Solr. For example, if you add string data Brondesbury Road to a name_l (a number) field then it bails out.

• You shouldn't really call $client->commit() after each insert.

• Updates are not handled yet, which you could implement by:

  • Save the document into Solr again.

  • Merge the returned document with the updated values from MongoDB.

  • Querying Solr for the document to be updates.

• Deleted are not handled yet.

• You might want to do something special about arrays in the MongoDB document and map them to multiValued fields in Solr.

Hopefully, this was still a good and simple introduction for tying in MongoDB with Solr. You can find the two PHP scripts at http://derickrethans.nl/files/mongodb-and-solr.tgz .