Lucene, shadow query classes, and the visitor pattern

A good Lucene result is achieved from an index and a query working together.

Defining the schema for the index is mostly an upfront design task: What are the fields? What fields are stored and what are indexed? How are terms in fields parsed? Are found terms supplemented? Are multiple fields combined? Etc.

Once you have made these decisions revisiting them can be prohibitive without planning. Ie, reindexing is found to be too expensive in time or resources, or, worse, the source documents are no longer available. Most Lucene uses error on the side of having a cautious schema where a lot more is stored and indexed than is needed now with the hope of a successful schema refactoring in the future. My experience has been that you don't know enough now about your data, how to index it, how to query it, and how users want to access it to have a viable cautious schema. It is better to store all your sources so you can reindex when you know better how the indexes and queries can cooperate. This is just a cost to using a new technology (and, as always, it can be mitigated).

The upshot in the near term is that the index schema is static. You have to apply flexibility with the query.

As mentioned in a previous posting, there is a tendency to think of Lucene queries like SQL queries. That is, there is a single, correct rendition. Discard that thinking. There are no correct results; there are only better results. To achieve better results you need to watch what your users searches and workout how the queries need to adjust. For example, perhaps you discover that there is a shift in vocabulary happening. What was once "OS X" is now "macOS". When a user queries for OS X you need to also include macOS in the query.

The Lucene API contains a number of query subtypes. These are combined to construct an expression that characterizes the user's search intent. This nested data structure should be considered a starting point. This data structure will be augmented and reformed to reflect your current understanding of how to best use the indexes. In the example above, you want to include macOS when OS X is used.
The Lucene API query subclasses are too rigid for direct augmentation and reforming. In the past the API was downright unbending and so I developed a set of shadow query classes that were amenable for use with the Visitor pattern. For example, this shows two visitors, the first adds the macOS variant and the second converts the query to a Solr expression:

Query q = ... new TermQuery(10.0f, "f", "osx") ...
...
Map<String, List<String>> variants = new HashMap<>();
variants.put("a", Arrays.asList("osx, "macos"));
VariantsQueryVistor vistor = new VariantsQueryVistor(0.0001f, variants);
...
q = vistor.visitQuery(q);
...
s = new SolrLuceneQueryVistor().visitQuery(q).toString();

and s is

( "osx" OR "macos" ^ 0.0001 ) ^ 10.0

Remain in control of your search

Lucene, Solr, and Elasticsearch have a powerful query language and a convenient textual representation. I have seen some API providers allow users to directly use this representation within their search API. Unfortunately, when you do so you loose significant control over your development and operations. You have exposed how you process your source data for full-text searching and so can no longer make behind-the-scenes schema changes. All your queryable data must be indexed in Lucene even if doing so is questionable and/or searchable elsewhere. You are locked into using Lucene to execute the query and at scale this can be exorbitantly expensive. To fix any of these issues you end up having to break backwards compatibility. How often will your API users accept this?

Instead, design your own query language. You can base it on Lucene's syntax if you like or an S-expression as it is trivial to parse in any programming language. No matter how you express the syntax, however, you have control over the semantics and the execution. For example, perhaps one of the searchable fields is not in your full-text index, but in a relational database and its presence in the query signals a programmatic join of the indexed and relational results. You can do that. Even if your queries are easily handled by Lucene at scale you are still better off translating yours syntax, like this old-school Google query

 -a +b c d

to the Lucene equivalent

 not a and ( b and (c d) )

because you remain in control.

Update: I fixed the term transposition in the Lucene equivalent of the Google expression. Doh!

Lucene boosting

Lucene has a powerful query language. The same language is used by Solr and Elasticsearch. I have found that users often don't utilize it well as they mistakenly apply their SQL experiences to it. With SQL your queries return exact results. Nothing in the result set is irrelevant. SQL query performance rewards tight queries consisting of few terms, few indexes, and few joins. With Lucene yours queries return ranked results. Much in the result set is actually peripheral. In fact, unless you limit the result set Lucene will return all the documents. A Lucene query is about getting a good ranking of results rather than exact results. To this end your query and indexes need to work together to achieve this. In general this means you need to make good use of broadening and narrowing terms, and boosting matches.

For example, if your query simply looked for the term "mouse" and you indexed your documents verbatim you should not expect to find any "mice". (Recall that Lucene sees your words as numbers and so "mouse" might be #23 and "mice" might be #6078.) It is therefore better to search for

mouse mice

When searching for "mouse mice" your results will be ordered so that either term gives equal weight to the document's rank. This is unlikely the correct course. The query was for "mouse" and you broadened it to include "mice". Documents matching "mouse" should be ranked higher than documents matching "mice". In Lucene's query language you do this by boosting the weight of terms. Documents matching "mouse" should be boosted orders of magnitude higher than "mice", ie

mouse^1000 mice

You will often see small boost values in other peoples examples. My experience has been that small boosts do not adequately differentiate documents. Big boosts do.

You likely noticed that my query was for "mouse mice" and not "mouse or mice". With Lucene, as with SQL, a boolean "or" kills performance. By not using "or" in your Lucene query you are allowing it to rank higher documents that contain both terms over documents that contain only one of them. Since the higher documents do contain the wanted term, "mouse", I don't see a need to repress ranking them equally to documents containing only "mouse" (ie, no use of "mice"). The performance cost is usually not worth it, especially when your queriers will be more complex than this simple example.

What does my pile of tech at Crossref look like?

What does my pile of tech at Crossref look like?

Service infrastructure is 29 deployments -- mostly 4 CPUs and 8G RAM -- handling 100M external (unique) queries and 1B internal requests per month.

Server infrastructure is Tomcat, ActiveMQ, MySql, and Oracle.

Service development is primarily in Java w/ Spring. Infrastructure operations aided with Bash and Perl scripts.

Primary datastores are RDBS using MySql and Oracle.

Secondary datastores are NoSQL using Oracle Berkeley DB, Solr, and bespoke solutions.

Full text search uses Solr, and bespoke Lucene solutions.

Data originates primarily in XML, JSON, tabular, and semi-structured text.

Lots of Linux operations experience. Some AWS operations experience.

No server, disk, or network hardware configuration and operations experience.

Source code managed in Subversion, developed in NetBeans, bespoke CI, and bespoke automated deployments.

Oh, and my trusty MacBook Plus.