To do a static analysis, inspect the log files, schema files, and the actual index for issues. If possible, connect to the live Solr instance while simultaneously examining log files so you can compare the schema with the index. The schema and the index can be out of sync in situations where the schema is changed, but the index was never rebuilt. Some hints are:

• A high number or proportion of 0-match queries. This indicates that the user-facing part of the application is making it easy for users to enter queries for which there are no matches. In Cloudera Search, given the size of the data, this should be an extremely rare event.
• Queries that match an excessive number of documents. All documents that match a query have to be scored, and the cost of scoring a query goes up as the number of hits increases. Examine any frequent queries that match millions of documents. An exception to this case is “constant score queries”. Queries, such as those of the form ":" bypass the scoring process entirely.
• Overly complex queries. Defining what constitutes overly complex queries is difficult to do, but a very general rule is that queries over 1024 characters in length are likely to be overly complex.
• High autowarm times. Autowarming is the process of filling caches. Some queries are executed before a new searcher serves the first live user request. This keeps the first few users from having to wait. Autowarming can take many seconds or can be instantaneous. Excessive autowarm times often indicate excessively generous autowarm parameters. Excessive autowarming usually has limited benefit, with longer runs effectively being wasted work.
  • Cache autowarm. Each Solr cache has an autowarm parameter. You can usually set this value to an upper limit of 128 and tune from there.
  • FirstSearcher/NewSearcher. The solrconfig.xml file contains queries that can be fired when a new searcher is opened (the index is updated) and when the server is first started. Particularly for firstSearcher, it can be valuable to have a query that sorts relevant fields.
    Note: The aforementioned flags are available from solrconfig.xml
• Exceptions. The Solr log file contains a record of all exceptions thrown. Some exceptions, such as exceptions resulting from invalid query syntax are benign, but others, such as Out Of Memory, require attention.
• Excessively large caches. The size of caches such as the filter cache are bounded by maxDoc/8. Having, for instance, a filterCache with 10,000 entries is likely to result in Out Of Memory errors. Large caches occurring in cases where there are many documents to index is normal and expected.
• Caches with low hit ratios, particularly filterCache. Each cache takes up some space, consuming resources. There are several caches, each with its own hit rate.
  • filterCache. This cache should have a relatively high hit ratio, typically around 80%.
  • queryResultCache. This is primarily used for paging so it can have a very low hit ratio. Each entry is quite small as it is basically composed of the raw query as a string for a key and perhaps 20-40 ints. While useful, unless users are experiencing paging, this requires relatively little attention.
  • documentCache. This cache is a bit tricky. It’s used to cache the document data (stored fields) so various components in a request handler don’t have to re-read the data from the disk. It’s an open question how useful it is when using MMapDirectory to access the index.
• Very deep paging. It is uncommon for user to go beyond the first page and very rare to go through 100 pages of results. A "&start=<pick your number>" query indicates unusual usage that should be identified. Deep paging may indicate some agent is completing scraping.
    Note: Solr is not built to return full result sets no matter how deep. If returning the full result set is required, explore alternatives to paging through the entire result set.
• Range queries should work on trie fields. Trie fields (numeric types) store extra information in the index to aid in range queries. If range queries are used, it’s almost always a good idea to be using trie fields.
• "fq" clauses that use bare NOW. “fq” clauses are kept in a cache. The cache is a map from the "fq" clause to the documents in your collection that satisfy that clause. Using bare NOW clauses virtually guarantees that the entry in the filter cache is not be re-used.
• Multiple simultaneous searchers warming. This is an indication that there are excessively frequent commits or that autowarming is taking too long. This usually indicates a misunderstanding of when you should issue commits, often to simulate Near Real Time (NRT) processing or an indexing client is improperly completing commits. With NRT, commits should be quite rare, and having more than one simultaneous autowarm should not happen.
• Stored fields that are never returned ("fl=" clauses). Examining the queries for “fl=” and correlating that with the schema can tell if stored fields that are not used are specified. This mostly wastes disk space. And "fl=*" can make this ambiguous. Nevertheless, it’s worth examining.
• Indexed fields that are never searched. This is the opposite of the case where stored fields are never returned. This is more important in that this has real RAM consequences. Examine the request handlers for “edismax” style parsers to be certain that indexed fields are not used.
• Queried but not analyzed fields. It’s rare for a field to be queried but not analyzed in any way. Usually this is only valuable for “string” type fields which are suitable for machine-entered data, such as part numbers chosen from a pick-list. Data that is not analyzed should not be used for anything that humans enter.
• String fields. String fields are completely unanalyzed. Unfortunately, some people confuse “string” with Java’s “String” type and use them for text that should be tokenized. The general expectation is that string fields should be used sparingly. More than just a few string fields indicates a design flaw.
• Whenever the schema is changed, re-index the entire data set. Solr uses the schema to set expectations about the index. When schemas are changed, there’s no attempt to retrofit the changes to documents that are currently indexed, but any new documents are indexed with the new schema definition. So old and new documents can have the same field stored in vastly different formats (for example, String and TrieDate) making your index inconsistent. This can be detected by examining the raw index.
• Query stats can be extracted from the logs. Statistics can be monitored on live systems, but it is more common to have log files. Here are some of the statistics you can gather.
  • Longest running queries
  • 0-length queries
  • average/mean/min/max query times
  • You can get a sense of the effects of commits on the subsequent queries over some interval (time or number of queries) to see if commits are the cause of intermittent slowdowns.
• Too-frequent commits have historically been the cause of unsatisfactory performance. This is not so important with NRT processing, but it is valuable to consider.
• Optimizing an index, which could improve search performance before, is much less necessary now. Anecdotal evidence indicates optimizing may help in some cases, but the general recommendation is to use “expungeDeletes”, instead of committing.
  • Modern Lucene code does what “optimize” used to do to remove deleted data from the index when segments are merged. Think of this process as a background optimize. Note that merge policies based on segment size can make this characterization inaccurate.
  • It still may make sense to optimize a read-only index.
  • “Optimize” is now renamed “forceMerge”.