Mastering Apache Solr 7.x: An expert guide to advancing, optimizing, and scaling your enterprise search

Doug Cutting created Lucene in 2000, which is the core technology behind Solr.

Solr was made in 2004 by Yonik Seeley at CNET Networks for a homegrown project to provide search capability for the CNET Networks website.

Later in 2006, CNET Networks published the Solr source code to ASF. By early 2007, Solr had found its place in some of the top projects. It was then that Solr kept on adding new features to attract customers and contributors.

Solr 1.3 was released in September 2008. It included major performance enhancements and features such as distributed search.

In January 2009, Yonik Seeley, Grant Ingersoll, and Erik Hatcher joined Lucidworks; they are the prime faces of Solr and enterprise search. Lucidworks started providing commercial support and training for Solr.

Solr 1.4 was released in November 2009. Solr had never stopped providing enhancements; 1.4 was no exception, with indexing, searching, faceting, rich document processing, database integration, plugins, and more.

In 2011, Solr versioning was revised to match up with the versions of Lucene. Sometime in 2010, the Lucence and Solr projects were merged; Solr had then became an integral subproject of Lucene. Solr downloads were still available separately; however, it was developed together by the same set of contributors. Solr was then marked as 3.1.

Solr 4.0 was released in October 2012, which introduced the SolrCloud feature. There were a number of follow-ups released over a couple of years in the 4.x line. Solr kept on adding new features, becoming more scalable and further focusing on reliability.

Solr 5.0 was released in February 2015. It was with this release that official support for the WAR bundle package ended. It was packaged as a standalone application. And later, in version 5.3, it also included an authentication and authorization framework.

Solr 6.0 was released in April 2016. It included support for executing parallel SQL queries across SolrCloud. It also included stream expression support and JDBC driver for the SQL interface.

Finally, Solr 7.0 was released in September 2017, followed by 7.1.0 in October 2017, as shown in the following diagram. We will discuss the new features as we move ahead in this chapter, in the What is new in Solr 7 section.

We have depicted the history of Solr in the preceding image for a much better view and understanding.

So by now, we have a brief understanding of Solr, along with its history. We must also have a good understanding of why we should be using Solr. Let's get the answer to this question too.

Lucene is an open source project that provides text search engine libraries. It is widely adopted for many search engine technologies. It has strong community contributions, which makes it much stronger as a technology backend. Lucene is a simple code library that you can use to write your own code by using the API available for searching, indexing, and much more.

For Lucene, a document consists of a collection of fields; they are name-value pairs consisting of either text or numbers. Lucene can be configured as a text analyzer that tokenizes a field’s text to a series of words. It can also do further processing, such as substituting with synonyms or other similar processes. Lucene stores its index on the disk of the server, which consists of indexing for each of the documents. The index is an inverted index that stores the mapping of a field to its relevant document, along with the position of the word from the text of the document. Once you have the index in place, you can search for documents with the input of a query string that is parsed accordingly to Lucence. Lucene manages to score a value for each of the relevant documents and the ones that are high-scoring documents are displayed.

One of the basic needs a search engine should support is a keyword search, as that's the primary goal behind the search engine platform. In fact it is the first thing a user will start with. Keyword search is the most common technique used for a search engine and also for end users on our websites. It is a pretty common expectation nowadays to punch in a few keywords and quickly retrieve the relevant results. But what happens in the backend is something we need to take care of to ensure that the user experience doesn't deteriorate. Let's look at a few areas that we must consider in order to provide better outcomes for search engine platforms using Solr:

These features can be easily managed by Solr; so our next challenge is to provide relevant results with improved user experience.

Solr is not limited to finding relevant results for a user's search. Providing the end user with selection of the most relevant results, that are sorted, is important as well. We will be doing this using SQL to find relevant matching pattern results and sorting them into columns in either ascending or descending order. Similarly, Solr also does sorting of the result set retrieved based on the search pattern, with a score that would match the relevancy strength in the dataset.

Ranked results is very important, primarily because the volume of data that search engine platforms have to dig through is huge. If there is no control on ranked results, then the result set would be filled with no relevancy and would have so much data that it wouldn't be feasible to display it either. The other important aspect is user experience. All of us are now used to expecting a search engine to provide relevant results using limited keywords. We are getting restless, aren't we? But we expect a search engine platform to not get annoyed and provide us relevant ranked results with few keywords. Hold on, we are not talking of Google search here! So for users like us, Solr can help address such situations by providing higher rankings based on various criteria: fields, terms, document name, and a few more. The ranking of the dataset can vary based on many factors, but a higher ranking would generally be based on the relevancy of the search pattern. With this, we can also have criteria such as gender; with the rankings of certain documents being at the top.

LinkedIn, a well known professional social media site, uses Lucene/Solr search. Lucene has a powerful faceting system that allows us to pivot and navigate by user or company attributes abstracted from user profile data. LinkedIn has an excellent feature that is backed up by Solr: its ranking of results by people's relationship with you. This data is not fixed, and being derived by Lucene in real time, it's all based on the arithmetic calculations of the relationships in your connections list.

One more use case is Myspace. Myspace is considered one of the world's largest search sites, with almost 200 million active users and adding up to almost 2.5k new users daily. It is expected to have around 50 million videos and adding around 75,000 daily. Myspace consists of almost 900 billion rows of data and 15 billion friend relationship searches by Lucene, with about 1 terabyte of data added every week.

NASA (https://www.nasa.gov/open/nebula.html) uses Solr for its Nebula Cloud Computing Platform. Similarly, Public Library of Science (PLOS) that is a non-profit publisher of research articles on various subjects. VUFind (https://vufind.org/vufind/) is another powerful open source discovery portal of libraries. It is known to have around 25 million records for a few of its implementations.

Having Google using Solr is a milestone for Solr. Google Search Appliance (GSA), is backed up with Solr. GSA uses many features of Solr: metadata sorting, recommendations, spellcheck, auto-suggest, and more.

Similarly, Open Test Search (http://www.opentestsearch.com/) uses Solr to provide a comparison of a few common search engines.

Flipkart is a leading example of Solr. It has more than 900k users and sees more than 20k searches per second. Flipkart product search has a backbone of 175 million listings, ~250 million documents, and ~5,500 categories. The major challenge was real-time results, ranking, autocompletion, high-update rates, and inverted index. It has become a huge success by using Solr for product searches for its e-commerce business.

Netflix uses Solr for the site search feature. Netflix has more than 2 million queries per day for searches and more than 15 million subscribers. It is available in more than 190 countries and supports around 23 languages. The search works based on video title name, genre name, or person name. Features such as autocompletion and ranked results are used by Netflix.

The Guardian, one of the leading newspapers, also uses Solr for its API search platform. There are other users too: MTV, Digg, cnet.com, and many more.

The White House uses Solr for https://www.whitehouse.gov/. It uses features such as search, highlighting, and faceting. Similarly, Federal Communications Commission (FCC) uses Solr for its website search.

Hathitrust is another wonderful use case of Solr. It has almost a couple of terabytes of index, with more than 10 million books provided online. Solr plays a prominent role in searches through its huge library of books. There are many such examples having similar use cases:

Before we understand the new replication methods introduced in Solr 7, let's go through what was available for replication before Solr 7.

Until Solr 7, Solr had two options for replication purposes:

In master-slave replication, also known as index replication, the master shares a copy of indexed data with one or more slave servers. The master server's job is to index the data that is being added into Solr and share it with the slave servers while all read operations are performed in the slaves.

SolrCloud is a clustered environment of Solr that provides high availability and failover capability so that the content indexed using Solr can be distributed equally among multiple servers for scaling. In SolrCloud, one of the servers act as the leader and the rest of the servers in the cluster work as replica shards. Until Solr 7, in case of any issue on the leader server, any of the replica servers could act as a leader and form the leader-replica cluster. So in that case, data had to be shared with each of the nodes in the cluster, as leader shards and replica shards must remain in sync at any time. Each replica node performed the same operations as the leader. This replication, method available in SolrCloud before Solr 7, was known as NRT replicas.

In Solr 7, two new replication methods have been introduced:

Solr has improved its schemaless mode functionality, the way it now detects data for indexing of an incoming field would be text based. By default, it will now be indexed as text_general for incoming fields, which can be modified if required. The name of the field will be the one defined in the document. A copy field rule will now be added in the schema when a collection is created if config set is not defined. It is now schemaless, which would insert the first 256 characters from the text field in a new strings field. It would be named as <name>_str.

The relevant schemaless behavior can be customized to remove a copy field rule or to update the number of characters added into the strings field or type of field used.

Copy field rules can impact the index size as well as slow down the indexing process. It is recommended to use the copy field rule when it is required. If there is no need to do a sort or facet on a field, you should ideally disable the copy field rule that is generated automatically.

The field creation rule can be disabled via the update.autoCreateFields property. You can also use the configuration API with the following command to disable it:

curl http://hostname:8983/solr/collection/config -d '{"set-user-property": {"update.autoCreateFields":"false"}}'

As termed in the documentation of http://lucene.apache.org, the goal of autoscaling is to make SolrCloud cluster management easier by providing a way for changes to the cluster to be more automatic and more intelligent.

So in Solr 7, there are some APIs that monitor some predefined preferences and policies. If any of the rules provided in the policies are violated, Solr changes its configuration automatically as defined in the preferences. With the updated autoscaling feature, we can now have Solr spin up new replicas depending on the monitoring metrics, such as disk space.

Trie*-based numeric files are now replaced by *PointField from Solr 7 onwards. Going forward, from Solr 8, all *PointField types will be removed. You need to work towards moving from *PointFields to the new Trie* fields for your schema. After changing to the new *Pointfields type, data will need to be reindexed in Solr.

Here is the list of spatial fields that have been deprecated:

The following is the list of spatial fields that can be used moving forward:

Here are the changes made in SolrJ:

Here are the changes made in Java Management Extensions (JMX) and MBeans:

Apart from these changes, there are many other features and improvements that have been made in Solr 7: