Here at Logentries we have a simple philosophy when it comes to hiring: hire the best people we can find and let them jump in at the deep end. That is how we like to learn. Smart people like to go deep and then find out what they don’t know as they work through some real world problems. And, our job is to give them the mentoring and support they need to overcome the blockers quickly and continue the learning process. We ensure they come to us with a great computer science or mathematical background and we take it from there.
When we thought about designing a training program we built it around the following set of problems:
1. CAP Theorem
Learn to accept that consistency, availability and partition tolerance cannot be provided simultaneously. Decide what is important for your customers and make the best trade-offs you can. But, what you build needs to be understandable. Raft provides a consensus algorithm whose aim is to be understandable, or at least to be easier to comprehend than Paxos or Paxos-influenced algorithms. It’s a good place to start to understand the distributed systems challenge. So we start with the theory then the next step we get practical…
2. Embrace Hardware Failures
Logentries runs is an Amazon EC2-based service – an environment where failures in parts of the system are routine. Netflix has taken this approach to its natural conclusion. If you expect failure and know that when it happens it’s going to happen at the worst possible time; why not plan for it to happen all time. So they built Chaos Monkey, a tool that randomly generates various kinds of failures in your environment. You can run this to schedule so that failures happen when you are closely monitoring things. By combining Raft + Chaos Monkey – you get to understand what happens when things go wrong and how to build fault tolerant systems that can deal with any failure.
3. Embrace the Asynchronous
Blocking calls will kill performance and may impact availability. The slowest system in a chain will cause a backlog on the other nodes. Asynchronous design allows you to avoid the cascade effects that can occur when one node in the system brings down other nodes. It is the best way to build fault tolerant systems. So we start by building a system that relies on synchronous calls, show what happens under extreme load or when failures occur and then evolve this to work asynchronously.
4. Mechanical Sympathy
The LMAX Disruptor team understands that to get every last drop of performance out of your system you sometimes need to know every last detail of the hardware environment you are running on. They call it Mechanical Sympathy. They took what they knew about modern day processor architectures (caches, memory barriers, compare and swap etc.) and turned it into a software solution for high performance inter-thread messaging, which leads to massive performance improvements when used correctly. So next we show our graduates what happens when you move from queue-based interactions to using Disruptor.
5. Horizontal over Vertical Scaling
Vertical scaling can give you quick wins in performance but when it comes to building a business like Logentries, where subscriber growth outpaces technology, the only way to design systems to scale with your success is to scale them horizontally. Don’t wait for technology to set limits on your growth. Also it’s better to scale on commodity hardware or, in our case, to scale on the EC2 servers that give you the best price-performance rather than rely on the latest premium-priced bells and whistles. When it comes to practical side of scaling we focus on splitting out the load but we plan for multiple axes of scale. Scaling out horizontally gives us the first split, where transactions are split over nodes. But we also ensure we have options for scaling by partitioning by types of transactions (log queries, API calls, live tail requests etc.), and by customer etc.
6. Talent Borrows, Genius Steals
We work only ten minutes away from where Oscar Wilde was born (Merrion Square, Dublin), so we do like to keep his words in mind. There are so many great distributed technologies around today – and they are getting better all the time. We have been inspired by Kafka, Hadoop, etcd, BDAS etc. But we like to control our own destiny and there are some very good reasons why we have built our platform from the ground up to do what it does in the best way it can. Turning logs into business insights isn’t easy – so we do like to stand on the shoulders of some distributed systems giants. So the final steps in our training program is to evaluate other distributed systems in the context of what we have learnt in the previous steps. Learn how Kafka implements mechanical sympathy when maximizing disk I/O. Understand Raft in the context of etcd. Understand the challenges on scaling analytics in the context of Hadoop.
Over the next few months – as we put our training program into practice we hope to have some of graduates blog about their experiences. Stay tuned!