High availability is necessary in mission-critical systems, but blind faith that the high availability solution is working and the architecture is always available (no matter how fault tolerant) is just crazy! Monitoring the architecture from top to bottom is necessary in a mission-critical system to ensure that failed pieces are caught early and dealt with swiftly before their effects propagate and cause the facing services to malfunction.
The typical approach to monitoring services is from the bottom up. Most monitoring services are managed by the operations group and as such tend to address their immediate needs and grow perpetually as needed. This methodology is in no way wrong; however, it is incomplete.
In my early years as a novice systems administrator, I was working on a client's systems after an unplanned outage and was asked why the problem wasn't caught earlier. The architecture in question was large and relatively complex (500,000 lines of custom perl web application code, 3 Oracle databases, and about 250 unique and unrelated daily jobs). The problem was that cron (the daemon responsible for dispatching daily jobs) got stuck writing to /var/log/ocron (its log file) and simply stopped running jobs. I won't explain in more detail, not because it is outside the scope of this book, but rather because I don't really understand why it malfunctioned. We had no idea that this could happen, and no bottom-up monitoring we had in place alerted us to this problem.
Where is this going? Bear with me a bit longer. I explained to the client that we monitor disk space, disk performance metrics, Oracle availability and performance, and a billion other things, but we weren't monitoring cron. He told me something extremely valuable that I have never forgotten: "I don't care if cron is running. I don't care if the disks are full or if Oracle has crashed. I don't care if the @#$% machine is on fire. All that matters is that my business is running."
I submit that top-down monitoring is as valuable as bottom-up monitoring. If your business is to sell widgets through your website, it doesn't really matter whether your machine is always on fire if you are smoothly selling widgets. Obviously, this statement is the other extreme, but you get the picture.
A solid and comprehensive monitoring infrastructure monitors systems from the bottom up (the systems and services), as well as from the top down (the business level). From this point on, we will call them systems monitors and business monitors, respectively.
Saying that you are going to monitor a web server is one thing; actually doing it is another. Even systems monitors are multifaceted. It is important to ensure that the machine is healthy by monitoring system load, memory-use metrics, network interface errors, disk space, and so on. On top of this, a web server is running, so you must monitor the number of hits it is receiving and the type and number of each type of response code (200 "OK," 302 "Redirect," 404 "Not Found," 500 "Internal Error," and so on). And, regardless of the metrics you get from the server itself, you should monitor the ability to contact the service over HTTP and HTTPS to ensure that you can load pages. These are the most basic systems monitors.
Aside from checking the actual service using the protocols it speaks (in this case web-specific protocols), how do you get all the metrics you want to monitor? There isn't one single answer to this, but a standard protocol is deployed throughout the industry called SNMP (simple network management protocol). Almost every commercially sold product comes instrumented with an SNMP agent. This means that by using the same protocol, you can ask every device on your network metric questions such as "How much disk space is used?" and "How many packets have you received?"
SNMP pretty much rules the monitoring landscape. Not only do most vendors implement SNMP in the architectural components they sell, all monitoring implementations (both commercial and free) support querying SNMP-capable devices. You'll notice that I said "most commercial components" and didn't mention free/open components. Sadly, many of the good open-source tools and free tools available are not instrumented with SNMP agents.
Probably the largest offender of this exclusion practice is the Mail Transport Agent (MTA). Most MTAs completely ignore the fact that there are standardized SNMP mechanisms called Management Information Bases (MIBs). The point here is not to complain about MTAs not implementing SNMP (although they should), but rather to illustrate that you will inevitably run into something you need to monitor on the systems level that doesn't speak SNMP. What happens then?
There are two ways to handle components that do not expose the needed information over SNMP. The first method is to fetch this metric via some unrelated network mechanism, such as SSH or HTTP, which has the tremendous advantage of simplicity. The second is to take this metric and export it over SNMP, making it both efficient and trivial to integrate into any monitoring system. Both methods require you first to develop an external means of determining the information you need (via a hand-written script or using a proprietary vendor-supplied tool).
Although the second choice may sound like the "right way," it certainly has a high overhead. SNMP is based on a complicated hierarchy of MIBs defined by a large and scattered set of documents. Without a very good MIB tool, it is difficult to decipher the information and an unbelievable pain to author a new MIB. To export the custom metrics you've determined over SNMP, you must extend or author a new MIB specifically for this and then either integrate an agent into the product or find an agent flexible enough to publish your MIB for you.
When is this worth it? That is a question that can only be answered by looking at the size of your architecture, the rate of product replacement, and the rate of change of your monitoring requirements. Remember, the monitoring system needs to monitor all these system metrics, as well as handle the business metrics. Business metrics, in my experience, are rarely exported over SNMP.
Some services are distributed in nature. Spread is one such example that cannot be effectively monitored with SNMP. The purpose of Spread is to allow separate nodes to publish data to other nodes in a specific fashion. Ultimately, the health of a single Spread daemon is not all that important. Instead, what you really want to monitor is the service that Spread provides, and you can only do so by using the service to accomplish a simple task and reporting whether the task completed as expected.
Criteria for a Capable Monitoring System
What makes a monitoring system good? First and foremost, it is the time invested in it and winning, or at least fighting, the constant battles brought on by architecture, application, and business changes. Components are added and removed on an ongoing basis. Application changes happen. Business metrics are augmented or deprecated. If the monitoring infrastructure does not adapt hand-in-hand with these changes, it is useless and even dangerous. Monitoring things no longer of importance while failing to monitor newly introduced metrics can result in a false sense of security that acts like blinders.
The second crucial criterion is a reliable and extensible monitoring infrastructure. A plethora of commercial, free, proprietary, and open monitoring frameworks are available. Rather than do a product review, we'll look at the most basic capabilities required of a monitoring solution:
- SNMP support—This is not difficult to find and will support most of the systems monitoring requirements.
- Extensibility—The ability to plug in ad hoc monitors. This is needed for many custom systems monitors and virtually all business monitors. Suppose that a business sells widgets and decides that (based on regressing last week's data) it should sell at least 3 widgets every 15 minutes and at least 20 widgets every hour between 6 a.m. and 10 p.m. and at least 10 widgets every hour between 10 p.m. and 6 a.m. If, at any point in time, those goals aren't met, it is likely that something technical is wrong. A good monitoring system allows an operator to place arbitrary rules such as this one alongside the other systems monitors.
- Flexible notifications—The system must be able to react to failures by alerting the operator responsible for the service. Not only should an operator be notified, but the infrastructure should also support automatic escalation of failures if they are not resolved in a given time period.
- Custom reaction—Some events that occur can be rectified or further investigated by an intelligent agent. If expected website traffic falls below a reasonable lower bound, the system should notify the operator. The operator is likely to perform several basic diagnostic techniques as the first phase of troubleshooting. Instead of wasting valuable time, the monitoring system could have performed the diagnostics and placed the output in the payload of the failure notification.
- Complex scheduling—Each individual monitor should have a period that can be modified. Metrics that are expensive to evaluate and/or are slow to change can be queried less frequently than those that are cheap and/or volatile.
- Maintenance scheduling—Monitors should never be taken offline. The system should support input from an administrator that some service or set of services is expected to be down during a certain time window in the future. During these windows, the services will be monitored, but noticed failures will not be escalated.
- Event acknowledgment—When things break that do not affect the overall availability and quality of service, they often do not need to be addressed until the next business day. Disabling notifications manually is dangerous. Instead, the system should acknowledge the failure and suspend notifications for a certain period of time or until a certain fixed point in time.
- Service dependencies—Each monitor that is put in place is not an automaton. On any given web server there will be 10 or more individual system checks (connectivity, remote access, HTTP, HTTPS, time skew, disk space, system load, network metrics, and various HTTP response code rates, just to name a few). That web server is plugged into a switch, which is plugged into a load balancer, which is plugged into a firewall, and so on. If your monitoring infrastructure exists outside your architecture, there is a clear service dependency graph. There is no sense in alerting an operator that there is a time skew on a web server if the web server has crashed. Likewise, there is no sense in alerting that there is a time skew or a machine crash if there is a failure of the switch to which the machine in question is attached. A hierarchy allows one event to be a superset of another. System alerts that go to a person's pager should be clear, be concise, and articulate the problem. Defining clear service dependencies and incorporating those relationships into the notification logic is a must for a complete monitoring system.