Tracing Black Boxes II: Monitoring Solr
Your site is indexed on Google, but that doesn’t mean you’re done with search. Content-rich websites provide native search functionality to keep users engaged, maintain visual consistency, and provide content-aware filtering. But it’s very hard to implement an effective, scalable search system, which is why Apache Solr is just about the most popular "black box" in web application infrastructure. This Lucene-backed search appliance has seen wide adoption due to its performance, reliability, and ease of deployment. In fact, it’s become so widely used that many Solr deployments are managed by people who have no other exposure to running Java applications. Documents go in, indexed RESTful search comes out – that is, until something breaks.
TraceView can provide insight into Apache Solr instances by correlating individual traces with JMX metrics, such as the rate of requests over the past five minutes. Even at a very low overall volume, an increased traffic rate is already increasing request latency.
Unlike most web application front-ends, Solr is a complex, stateful application that contains persistent objects, runs background indexing processes, and maintains multiple tiers of caches. There are a lot of ways to deploy and configure Solr, and that means there are a lot of ways to make mistakes. But even when you have everything up and running, there’s always the lingering question of whether you could be getting more out of your Solr instances (or reducing the number of them!).
One of the best ways to get insight into Solr’s internal abstractions – such as cores, handlers, and components – is to monitor them directly via JMX. I’ve previously written about using JMX metrics to keep tabs on JVM memory internals, but JMX is a common API for collecting data from your Java applications and any application can make use of it. Because of this it’s been widely adopted in the Java ecosystem to centralize the provision of application-specific performance data.
Solr provides JMX metrics on a variety of internals, such as queryResultCache.
Solr exposes hundreds of JMX metrics across dozens of categories, and efficient use of them can help you delve into Solr performance in a variety of ways. Some metrics are better for providing a high-level view of Solr’s overall workflow. The
queryResultCachecategory, pictured above, provides a snapshot of how often your data was successfully cached, as well as how often cache entries had to be evicted due to insufficient space. Other metric categories are more granular and provide detail at the level of classes, or even objects. An update request will be routed to a different handler depending on whether the data was provided in XML, CSV, or JSON; each of these update handlers exposes metrics independently, like how long it has been running and the number of errors.
JMX metrics can even provide insight into advanced Solr use cases, like modifying result scoring to permit n-dimensional spatial searches or customizing results based on user data stored in Redis. Even without adding custom JMX metrics, Solr will report enough data to allow you to separately track the effectiveness of these custom searches relative to more traditional queries.
Let’s look at a practical example. You just got paged because half of your distributed Solr cluster lost connectivity in a widespread EC2 outage. It looks like it might last a while, so you decide to add additional capacity in one of the functioning availability zones. Rather than spending time re-indexing your content, you decide to replicate your existing Solr data to the new servers. Using the high-level metrics provided by
ReplicationHandler, you determine that replication is proceeding smoothly. Halfway through your second replication, though, you realize that the first replicated server is taking five times as long as your original servers to respond to the same user queries, even though it’s running on the same hardware. Checking out the cache metrics for a specific search handler, it looks like the hit ratios on its caches are abysmal – but wait, what’s actually in those caches? After checking the metrics for that node’s active Searcher instance, you realize you didn’t set up Solr to warm the cache – it was starting off empty! Now you know to make a quick configuration change next time you spin up an instance so that the first users routed to it will have acceptable performance.
So, that sounds awesome – but how do you do it? The easiest approach is to view Solr’s JMX statistics through its web interface (in Solr 3.x,
/solr/admin/stats.jsp, while in Solr 4.x it’s available at a collection-based URL like
/solr/#/collection1/plugins/). However, web access won’t be an option for most deployments. Alternately, you could use remote jconsole, but that requires either a complex remote configuration that’s a tremendous hassle to set up or the glacially slow option of SSH X11 forwarding (e.g.,
ssh -X solr jconsole).
In practice, those approaches all suck. Solr is stunningly verbose: it exposes hundreds of JMX metrics out of the box, and that number expands quickly as you add additional handlers and components. Purpose-built JMX monitoring tools like
jconsole are great for browsing the available metrics to see what’s available, but they’re horrible for pulling out the ones you want in a hurry. They also allow write operations like initiating garbage collection or clearing caches – definitely not something you want to give out to every developer!
TraceView automatically monitors the JMX metrics of every node involved in this distributed Solr Cloud trace.
On a day to day basis, it’s more common to read JMX metrics via automated, read-only monitoring tools like Nagios, Ganglia, or AppNeta TraceView. These tools not only present a number of metrics at once, but they also generally let you filter down to a meaningful subset of the hundreds of lines exposed by Solr. On the other hand, “health check”-style metrics aren’t necessarily the only way to look the problem. Each request has a number of metrics it can generate, and bringing together these data sources in one application has some real advantages. Looking at an individual request can tell you exactly what went wrong and it’s often the context of JMX data that says why. Examining the concurrent host activity can disambiguate between whether a pause was due to a garbage collection event in the JVM or an overloaded document cache in Solr forcing additional disk access.
Next time, we’ll talk about how TraceView captures these request-based metrics directly from the Solr internals. In the meantime, if you’ve got a Solr installation, sign up for your free account, put it on that server, and take a look inside that black box!