Chapter 11. mnoGoSearch cluster

Introduction

Starting from the version 3.3.0, mnoGoSearch provides a clustered solution, which allows to scale search on several computers, extending database size up to several dozens or even hundreds million documents.

How it works

A typical cluster consists of several database machines (cluster nodes) and a single front-end machine. The front-end machine receives HTTP requests from a user's browser, forwards search queries to the database machines using HTTP protocol, receives back a limited number of the best top search results (using a simple XML format, based on OpenSearch specifications) from every database machine, then parses and merges the results ordering them by score, and displays the results applying HTML template. This approach distributes operations with high CPU and hard disk consumption between the database machines in parallel, leaving simple merge and HTML template processing functions to the the front-end machine.

Operations done on the database machines

In a clustered mnoGoSearch installation, all hard operations are done on the database machines.

This is an approximate distribution chart of time spent on different search steps:

• 5% - fetching word information

• 30% - grouping words by doc ID

• 30% - calculating score for each document

• 20% - sorting documents according to score

• 10% - fetching cached copies and generating excerpts for the best 10 documents

• 5% - generating XML or HTML results using template

How a typical XML response looks like

<rss>
  <channel>

    <!-- General search results information block -->
    <openSearch:totalResults>20</openSearch:totalResults>
    <openSearch:startIndex>1</openSearch:startIndex>
    <openSearch:itemsPerPage>2</openSearch:itemsPerPage>

    <!-- Word statistics block -->
    <mnoGoSearch:WordStatList>
      <mnoGoSearch:WordStatItem order="0" count="300" word="apache"/>
      <mnoGoSearch:WordStatItem order="1" count="103" word="web"/>
      <mnoGoSearch:WordStatItem order="2" count="250" word="server"/>
    </mnoGoSearch:WordStatList>

    <!-- document information block -->
    <item>
      <id>1</id>
      <score>70.25%</score>
      <title>Test page of the Apache HTTP Server</title>
      <link>http://hostname/index.html</link>
      <description>...to use the images below on Apache
      and Fedora Core powered HTTP servers. Thanks for using Apache ...
      </description>
      <updated>2006-12-13T18:30:02Z</updated>
      <content-length>3956</content-length>
      <content-type>text/html</content-type>
    </item>

    <!-- more items, typically 10 items total -->

  </channel>
</rss>

Operations done on the front-end machine

The front-end machine receives XML responses from every database machine. On the first query, the front-end machine requests top 10 results from every database machine. An XML response for the top 10 results page is about 5Kb. Parsing of each XML response takes less than 1% of time. Thus, a cluster consisting of 50 machines is about 50% slower than a cluster consisting of a single machine, but allows to search through a 50-times bigger collection of documents.

If the user is not satisfied with search results returned on the first page and navigates to higher pages, then the front-end machine requests ps*np results from each database machine, where ps is page size (10 by default), and np is page number. Thus, to display the 5th result page, the front-end machine requests 50 results from every database machine and has to do five-times more parsing job, which makes search on higher pages a little bit slower. But typically, users look through not more than 2-3 pages.

Cluster types

mnoGoSearch supports two cluster types. A "merge" cluster is to join results from multiple independent machines, each one created by its own indexer.conf. This type of cluster is recommended when it is possible to distribute your web space into separate databases evenly using URL patterns by means of the Server or Realm commands. For example, if you need to index three sites siteA, siteB and siteC with an approximately equal number of documents on each site, and you have three cluster machines nodeA, nodeB and nodeC, you can put each site to a separate machine using a corresponding Server command in the indexer.conf file on each cluster machine:

# indexer.conf on machine nodeA:
DBAddr mysql://root@localhost/test/
Server http://siteA/

# indexer.conf on machine nodeB:
DBAddr mysql://root@localhost/test/
Server http://siteB/

# indexer.conf on machine nodeC:
DBAddr mysql://root@localhost/test/
Server http://siteC/

A "distributed" cluster is created by a single indexer.conf, with "indexer" automatically distributing search data between database machines. This type of cluster is recommended when it is hard to distribute web space between cluster machines using URL patterns, for example when you don't know your site sizes or the site sizes are very different.

Installing and configuring a "merge" cluster

Installing and configuring a "distributed" cluster

DBAddr mysql://hostname1/dbname/?dbmode=blob
DBAddr mysql://hostname2/dbname/?dbmode=blob
DBAddr mysql://hostname3/dbname/?dbmode=blob

#
# The entire top level domain .ru,
# using http://www.ru/ as a start point
#
Server http://www.ru/
Realm http://*.ru/*

Using dump/restore tools to add cluster nodes

Starting from the version 3.3.9, mnoGoSearch allows to add new nodes into a cluster (or remove nodes) without having to re-crawl the documents once again.

Suppose you have 5 cluster nodes and what extend the cluster to 10 nodes. Please go through the following steps:

1. Stop all indexer processes.

2. Create all new 10 SQL databases and create a new .conf file with 10 DBAddr commands. Note, the old and the new SQL databases can NOT overlap. The new databases must be freshly created empty databases.

3. Run

   indexer -d /path/to/new.conf -Ecreate

   to create table structure in all 10 new SQL databases.

4. Make sure you have enough disk space - you'll need about 2 times extra disk space of the all original SQL databases size.

5. Create a directory, say, /usr/local/mnogosearch/dump-restore/, where you will put the dump file, then go to this directory.

6. Run

   indexer -Edumpdata | gzip > dumpfile.sql.gz 

   It will create the dump file.

7. Run

   zcat dumpfile.sql.gz | indexer -d /path/to/new.conf -Esql -v2

   It will load information from the dump file and put it into the new SQL databases. Note, all document IDs will be automatically re-assigned.

8. Check that restoring worked fine. These two commands should report the same statistics:

   indexer -Estat
   indexer -Estat -d /path/to/new.conf
   

9. Run

   indexer -d /path/to/new.conf -Eblob

   to create inverted search index in the new SQL databases;

10. Configure a new search front-end to use the new SQL databases and check that search bring the same results from the old and the new databases.

Cluster limitations

• As of version 3.3.0, mnoGoSearch allows to join up to 256 database machines into a single cluster.

• Only "body" and "title" sections are currently supported. Support for other standard sections, like meta.keywords and meta.description, as well as for user-defined sections will be added in the future 3.3.x releases.

• Cluster of clusters is not fully functional yet. It will likely be implemented in one of the near future 3.3.x release.

• Popularity rank does not work well with cluster. Links between documents residing on different cluster nodes are not taken into account. This will be fixed in future releases.