Troubleshooting Cloudera Search
After installing and deploying Cloudera Search, use the information in this section to troubleshoot problems.
Troubleshooting
The following table contains some common troubleshooting techniques.
Note: In the URLs in the following table, you must replace entries such as <server:port> with values from your environment. The port defaults value is 8983, but see /etc/default/solr or /opt/cloudera/parcels/CDH-*/etc/default/solr for the port if you are in doubt.
|
Symptom |
Explanation |
Recommendation |
|---|---|---|
|
All |
Varied |
Examine Solr log. By default, the log can be found at /var/log/solr/solr.out. |
|
No documents found |
Server may not be running |
Browse to http://server:port/solr to see if the server responds. Check that cores are present. Check the contents of cores to ensure that numDocs is more than 0. |
|
No documents found |
Core may not have documents |
Browsing http://server:port/solr/[collection name]/select?q=*:*&wt=json&indent=true should show numFound, which is near the top, to be more than 0. |
| The secure Solr Server fails to respond to Solrj requests, but other clients such as curl can communicate successfully | This may be a version compatibility issue. Httpclient 4.2.3, which ships with solrj in Search 1.x, has a dependency on commons-codec 1.7. If an earlier version of commons-codec is on the classpath, httpclient may be unable to communicate using Kerberos. | Ensure your application is using commons-codec 1.7 or later. Alternatively, use httpclient 4.2.5 instead of version 4.2.3 in your application. Version 4.2.3 behaves correctly with earlier versions of commons-codec. |
Dynamic Solr Analysis
Any JMX-aware application can query Solr for information and display results dynamically. For example, Zabbix, Nagios, and many others have been used successfully. When completing Static Solr Log Analysis, many of the items related to extracting data from the log files can be seen from querying Solr, at least the last value (as opposed to the history which is available from the log file). These are often useful for status boards. In general, anything available from the Solr admin page can be requested on a live basis from Solr. Some possibilities include:
- numDocs/maxDoc per core. This can be important since the difference between these numbers indicates the number of deleted documents in the index. Deleted documents take up disk space and memory. If these numbers vary greatly, this may be a rare case where optimizing is advisable.
- Cache statistics, including:
- Hit ratios
- Autowarm times
- Evictions
- Almost anything available on the admin page. Note that drilling down into the “schema browser” can be expensive.
Other Troubleshooting Information
Since the use cases for Solr and search vary, there is no single solution for all cases. That said, here are some common challenges that many Search users have encountered:
- Testing with unrealistic data sets. For example, a users may test a prototype that uses faceting, grouping, sorting, and complex schemas against a small data set. When this same system is used to load of real data, performance issues occur. Using realistic data and use-cases is essential to getting accurate results.
- If the scenario seems to be that the system is slow to ingest data, consider:
- Upstream speed. If you have a SolrJ program pumping data to your cluster and ingesting documents at a rate of 100 docs/second, the gating factor may be upstream speed. To test for limitations due to upstream speed, comment out only the code that sends the data to the server (for example, SolrHttpServer.add(doclist)) and time the program. If you see a throughput bump of less than around 10%, this may indicate that your system is spending most or all of the time getting the data from the system-of-record.
- This may require pre-processing.
- Indexing with a single thread from the client. ConcurrentUpdateSolrServer can use multiple threads to avoid I/O waits.
- Too-frequent commits. This was historically an attempt to get NRT processing, but with SolrCloud hard commits this should be quite rare.
- The complexity of the analysis chain. Note that this is rarely the core issue. A simple test is to change the schema definitions to use trivial analysis chains and then measure performance.
- When the simple approaches fail to identify an issue, consider using profilers.