Developing A Monitoring System
I currently spend most of my spare time developing Bergamot Monitoring . Developing a monitoring throws up some interesting challenges. I want to discuss some of the things, as a web developer that I've realised during the course of this project.
Caching is a technique often used by web applications to improve performance. Often is applied at multiple levels within a web application: data layer caching, view caching, etc. For most web application, the caching of a rendered page provides a massive performance gain. However for a monitoring system, caching is next to useless. The key issue with monitoring systems, is that that everything changes and changes often.
In the worst case (with defaults) a check could be executed every minute by Bergamot. Oh and users need to know the second that something changes, after all that is the point of a monitoring system. This means it is guaranteed that a view will change within one minute, little point in caching that.
This problem is compounded by group views, where the result of multiple checks are displayed. Even on a modest sized system, these views could change every 10 seconds. On larger deployments, the state of a group can change multiple times a second.
The core issue with monitoring, is that stuff is changing all the time.
Failing out of the caching problem, and the constant change problem. Is that users need to see consistent result. To scale the resources of multiple servers are needed. But unlike simpler web applications, coherency needs to be managed across these machines.
When the result of a check is processed all data caches need to be updated and invalidated coherently across the cluster. This is further complicated by the result transition logic being transactional.
Groups form a heirarchial tree, where child groups exist for a parent, fairly normal stuff really. However the state of child groups needs to be encompassed by the parent group. In other words to compute the state of a group at any moment in time, we need to compute the state of the whole tree.
If you naively attempt to solve that problem in your application, be prepared for the performance and scalability hit. The latency of round tripping to the database (even when on Localhost) completely kills performance, due to the sheer volume of queries that need to be executed (even for a trivaially small system).
Enter the joys of recursive SQL, where we can use a single SQL query to compute the state of the entire graph with one query (albeit about 30 lines of SQL). SQL is an often underused powerhouse of data querying and manipulation, fuck that ORM and spend the time learning propper SQL, you'll thank yourself that you did.
Message queues are awesome, Bergamot makes use of RabbitMQ to pass messages between multiple nodes. This is how Bergamot distributes work across multiple servers.
We push alot of routing logic down into RabbitMQ. Using features such as exchange to exchange bindings, alternate exchanges, per message time to live, dead lettering. This allows Bergamot to build a really flexible routing model without having to deal with any of the mechanics of it.
Word of advice, get a peice of papper and sketch out your routing before you code it up.
Websockets are seriously cool, they allow Bergamot to realise updating checks in real time. Websockets implement true push messaging for the web and the technology should not be overlooked, it is super easy to use via the browser.
The server side however is a little more complex. Websockets rely upon a long running TCP / HTTP connection, as such you need to ensure that the backend server is non-blocking / event based (like wise for all servers in the connection path.
Programming for non-blocking / event based servers is very different from programming for threaded servers. Bergamot makes use of Netty to handle websockets, Netty is an event based networking library for Java and has support for websockets. Bergamot uses Netty to bridge between websockets and message queues. The change in state of a check is published to a message queue, Netty is used to simply listen to these messages and transmit them to browsers.
This allows for less than 200ms of latency between telling Bergamot to execute a check in the UI, to Bergamot executing the check and publishing the result to the browser. I deliberately had to have a slow animation effect in the UI so that users could realise that a check had actually updated!