Constructing and working a reasonably large storage system known as S3

But at the time, what I was convinced I really wanted to do was to be a university professor. I applied for a bunch of faculty jobs and wound up finding one at UBC (which worked out really well, because my wife already had a job in Vancouver and we love the city). I threw myself into the faculty role and foolishly grew my lab to 18 students, which is something that I’d encourage anyone that’s starting out as an assistant professor to never, ever do. It was thrilling to have such a large lab full of amazing people and it was absolutely exhausting to try to supervise that many graduate students all at once, but, I’m pretty sure I did a horrible job of it. That said, our research lab was an incredible community of people and we built things that I’m still really proud of today, and we wrote all sorts of really fun papers on security, storage, virtualization, and networking.

A little over two years into my professor job at UBC, a few of my students and I decided to do another startup. We started a company called Coho Data that took advantage of two really early technologies at the time: NVMe SSDs and programmable ethernet switches, to build a high-performance scale-out storage appliance. We grew Coho to about 150 people with offices in four countries, and once again it was an opportunity to learn things about stuff like the load bearing strength of second-floor server room floors, and analytics workflows in Wall Street hedge funds – both of which were well outside my training as a CS researcher and teacher. Coho was a wonderful and deeply educational experience, but in the end, the company didn’t work out and we had to wind it down.

And so, I found myself sitting back in my mostly empty office at UBC. I realized that I’d graduated my last PhD student, and I wasn’t sure that I had the strength to start building a research lab from scratch all over again. I also felt like if I was going to be in a professor job where I was expected to teach students about the cloud, that I might do well to get some first-hand experience with how it actually works.

I interviewed at some cloud providers, and had an especially fun time talking to the folks at Amazon and decided to join. And that’s where I work now. I’m based in Vancouver, and I’m an engineer that gets to work across all of Amazon’s storage products. So far, a whole lot of my time has been spent on S3.

When I joined Amazon in 2017, I arranged to spend most of my first day at work with Seth Markle. Seth is one of S3’s early engineers, and he took me into a little room with a whiteboard and then spent six hours explaining how S3 worked.

It was awesome. We drew pictures, and I asked question after question non-stop and I couldn’t stump Seth. It was exhausting, but in the best kind of way. Even then S3 was a very large system, but in broad strokes — which was what we started with on the whiteboard — it probably looks like most other storage systems that you’ve seen.

S3 is an object storage service with an HTTP REST API. There is a frontend fleet with a REST API, a namespace service, a storage fleet that’s full of hard disks, and a fleet that does background operations. In an enterprise context we might call these background tasks “data services,” like replication and tiering. What’s interesting here, when you look at the highest-level block diagram of S3’s technical design, is the fact that AWS tends to ship its org chart. This is a phrase that’s often used in a pretty disparaging way, but in this case it’s absolutely fascinating. Each of these broad components is a part of the S3 organization. Each has a leader, and a bunch of teams that work on it. And if we went into the next level of detail in the diagram, expanding one of these boxes out into the individual components that are inside it, what we’d find is that all the nested components are their own teams, have their own fleets, and, in many ways, operate like independent businesses.

All in, S3 today is composed of hundreds of microservices that are structured this way. Interactions between these teams are literally API-level contracts, and, just like the code that we all write, sometimes we get modularity wrong and those team-level interactions are kind of inefficient and clunky, and it’s a bunch of work to go and fix it, but that’s part of building software, and it turns out, part of building software teams too.

Before Amazon, I’d worked on research software, I’d worked on pretty widely adopted open-source software, and I’d worked on enterprise software and hardware appliances that were used in production inside some really large businesses. But by and large, that software was a thing we designed, built, tested, and shipped. It was the software that we packaged and the software that we delivered. Sure, we had escalations and support cases and we fixed bugs and shipped patches and updates, but we ultimately delivered software. Working on a global storage service like S3 was completely different: S3 is effectively a living, breathing organism. Everything, from developers writing code running next to the hard disks at the bottom of the software stack, to technicians installing new racks of storage capacity in our data centers, to customers tuning applications for performance, everything is one single, continuously evolving system. S3’s customers aren’t buying software, they are buying a service and they expect the experience of using that service to be continuously, predictably fantastic.

The first observation was that I was going to have to change, and really broaden how I thought about software systems and how they behave. This didn’t just mean broadening thinking about software to include those hundreds of microservices that make up S3, it meant broadening to also include all the people who design, build, deploy, and operate all that code. It’s all one thing, and you can’t really think about it just as software. It’s software, hardware, and people, and it’s always growing and constantly evolving.

The second observation was that despite the fact that this whiteboard diagram sketched the broad strokes of the organization and the software, it was also wildly misleading, because it completely obscured the scale of the system. Each one of the boxes represents its own collection of scaled out software services, often themselves built from collections of services. It would literally take me years to come to terms with the scale of the system that I was working with, and even today I often find myself surprised at the consequences of that scale.

It probably isn’t very surprising for me to mention that S3 is a really big system, and it is built using a LOT of hard disks. Millions of them. And if we’re talking about S3, it’s worth spending a little bit of time talking about hard drives themselves. Hard drives are amazing, and they’ve kind of always been amazing.

So, with all this in mind, one of the biggest and most interesting technical scale problems that I’ve encountered is in managing and balancing I/O demand across a really large set of hard drives. In S3, we refer to that problem as heat management.

By heat, I mean the number of requests that hit a given disk at any point in time. If we do a bad job of managing heat, then we end up focusing a disproportionate number of requests on a single drive, and we create hotspots because of the limited I/O that’s available from that single disk. For us, this becomes an optimization challenge of figuring out how we can place data across our disks in a way that minimizes the number of hotspots.

Hotspots are small numbers of overloaded drives in a system that ends up getting bogged down, and results in poor overall performance for requests dependent on those drives. When you get a hot spot, things don’t fall over, but you queue up requests and the customer experience is poor. Unbalanced load stalls requests that are waiting on busy drives, those stalls amplify up through layers of the software storage stack, they get amplified by dependent I/Os for metadata lookups or erasure coding, and they result in a very small proportion of higher latency requests — or “stragglers”. In other words, hotspots at individual hard disks create tail latency, and ultimately, if you don’t stay on top of them, they grow to eventually impact all request latency.

As S3 scales, we want to be able to spread heat as evenly as possible, and let individual users benefit from as much of the HDD fleet as possible. This is tricky, because we don’t know when or how data is going to be accessed at the time that it’s written, and that’s when we need to decide where to place it. Before joining Amazon, I spent time doing research and building systems that tried to predict and manage this I/O heat at much smaller scales – like local hard drives or enterprise storage arrays and it was basically impossible to do a good job of. But this is a case where the sheer scale, and the multitenancy of S3 result in a system that is fundamentally different.

The more workloads we run on S3, the more that individual requests to objects become decorrelated with one another. Individual storage workloads tend to be really bursty, in fact, most storage workloads are completely idle most of the time and then experience sudden load peaks when data is accessed. That peak demand is much higher than the mean. But as we aggregate millions of workloads a really, really cool thing happens: the aggregate demand smooths and it becomes way more predictable. In fact, and I found this to be a really intuitive observation once I saw it at scale, once you aggregate to a certain scale you hit a point where it is difficult or impossible for any given workload to really influence the aggregate peak at all! So, with aggregation flattening the overall demand distribution, we need to take this relatively smooth demand rate and translate it into a similarly smooth level of demand across all of our disks, balancing the heat of each workload.

In storage systems, redundancy schemes are commonly used to protect data from hardware failures, but redundancy also helps manage heat. They spread load out and give you an opportunity to steer request traffic away from hotspots. As an example, consider replication as a simple approach to encoding and protecting data. Replication protects data if disks fail by just having multiple copies on different disks. But it also gives you the freedom to read from any of the disks. When we think about replication from a capacity perspective it’s expensive. However, from an I/O perspective – at least for reading data – replication is very efficient.

So, redundancy schemes let us divide our data into more pieces than we need to read in order to access it, and that in turn provides us with the flexibility to avoid sending requests to overloaded disks, but there’s more we can do to avoid heat. The next step is to spread the placement of new objects broadly across our disk fleet. While individual objects may be encoded across tens of drives, we intentionally put different objects onto different sets of drives, so that each customer’s accesses are spread over a very large number of disks.

There are two big benefits to spreading the objects within each bucket across lots and lots of disks:

1. A customer’s data only occupies a very small amount of any given disk, which helps achieve workload isolation, because individual workloads can’t generate a hotspot on any one disk.
2. Individual workloads can burst up to a scale of disks that would be really difficult and really expensive to build as a stand-alone system.

For instance, look at the graph above. Think about that burst, which might be a genomics customer doing parallel analysis from thousands of Lambda functions at once. That burst of requests can be served by over a million individual disks. That’s not an exaggeration. Today, we have tens of thousands of customers with S3 buckets that are spread across millions of drives. When I first started working on S3, I was really excited (and humbled!) by the systems work to build storage at this scale, but as I really started to understand the system I realized that it was the scale of customers and workloads using the system in aggregate that really allow it to be built differently, and building at this scale means that any one of those individual workloads is able to burst to a level of performance that just wouldn’t be practical to build if they were building without this scale.

Beyond the technology itself, there are human factors that make S3 – or any complex system – what it is. One of the core tenets at Amazon is that we want engineers and teams to fail fast, and safely. We want them to always have the confidence to move quickly as builders, while still remaining completely obsessed with delivering highly durable storage. One strategy we use to help with this in S3 is a process called “durability reviews.” It’s a human mechanism that’s not in the statistical 11 9s model, but it’s every bit as important.

When an engineer makes changes that can result in a change to our durability posture, we do a durability review. The process borrows an idea from security research: the threat model. The goal is to provide a summary of the change, a comprehensive list of threats, then describe how the change is resilient to those threats. In security, writing down a threat model encourages you to think like an adversary and imagine all the nasty things that they might try to do to your system. In a durability review, we encourage the same “what are all the things that might go wrong” thinking, and really encourage engineers to be creatively critical of their own code. The process does two things very well:

1. It encourages authors and reviewers to really think critically about the risks we should be protecting against.
2. It separates risk from countermeasures, and lets us have separate discussions about the two sides.

When working through durability reviews we take the durability threat model, and then we evaluate whether we have the right countermeasures and protections in place. When we are identifying those protections, we really focus on identifying coarse-grained “guardrails”. These are simple mechanisms that protect you from a large class of risks. Rather than nitpicking through each risk and identifying individual mitigations, we like simple and broad strategies that protect against a lot of stuff.

But additionally, technically, it was enjoyable to debate Dynamo, as a result of Dynamo was finally constant, so it was attainable in your procuring cart to be unsuitable.

I beloved this, as a result of it was the place we’d talk about what you do, virtually, in manufacturing, when Dynamo was unsuitable. When a buyer was capable of place an order solely to later understand that the final merchandise had already been offered. You detected the battle however what may you do? The shopper was anticipating a supply.

This instance could have stretched the Dynamo paper’s story a bit bit, but it surely drove to an incredible punchline. As a result of the scholars would usually spend a bunch of dialogue making an attempt to provide you with technical software program options. Then somebody would level out that this wasn’t it in any respect. That finally, these conflicts had been uncommon, and you possibly can resolve them by getting assist employees concerned and making a human choice. It was a second the place, if it labored effectively, you possibly can take the category from being important and engaged in serious about tradeoffs and design of software program methods, and you possibly can get them to appreciate that the system is likely to be larger than that. It is likely to be an entire group, or a enterprise, and possibly a few of the similar pondering nonetheless utilized.

Now that I’ve labored at Amazon for some time, I’ve come to appreciate that my interpretation wasn’t all that removed from the reality — when it comes to how the providers that we run are hardly “simply” the software program. I’ve additionally realized that there’s a bit extra to it than what I’d gotten out of the paper when educating it. Amazon spends lots of time actually centered on the thought of “possession.” The time period comes up in lots of conversations — like “does this motion merchandise have an proprietor?” — that means who’s the only individual that’s on the hook to essentially drive this factor to completion and make it profitable.

The give attention to possession really helps perceive lots of the organizational construction and engineering approaches that exist inside Amazon, and particularly in S3. To maneuver quick, to maintain a very excessive bar for high quality, groups must be homeowners. They should personal the API contracts with different methods their service interacts with, they must be fully on the hook for sturdiness and efficiency and availability, and finally, they should step in and repair stuff at three within the morning when an sudden bug hurts availability. However additionally they must be empowered to replicate on that bug repair and enhance the system in order that it doesn’t occur once more. Possession carries lots of duty, but it surely additionally carries lots of belief – as a result of to let a person or a staff personal a service, it’s a must to give them the leeway to make their very own choices about how they’ll ship it. It’s been an incredible lesson for me to appreciate how a lot permitting people and groups to instantly personal software program, and extra usually personal a portion of the enterprise, permits them to be enthusiastic about what they do and actually push on it. It’s additionally exceptional how a lot getting possession unsuitable can have the alternative end result.

I’ve spent lots of time at Amazon serious about how vital and efficient the give attention to possession is to the enterprise, but additionally about how efficient a person instrument it’s after I work with engineers and groups. I spotted that the thought of recognizing and inspiring possession had really been a very efficient instrument for me in different roles. Right here’s an instance: In my early days as a professor at UBC, I used to be working with my first set of graduate college students and making an attempt to determine how to decide on nice analysis issues for my lab. I vividly keep in mind a dialog I had with a colleague that was additionally a reasonably new professor at one other college. After I requested them how they select analysis issues with their college students, they flipped. They’d a surprisingly annoyed response. “I can’t determine this out in any respect. I’ve like 5 initiatives I would like college students to do. I’ve written them up. They hum and haw and decide one up but it surely by no means works out. I may do the initiatives sooner myself than I can train them to do it.”

And finally, that’s really what this individual did — they had been superb, they did a bunch of actually cool stuff, and wrote some nice papers, after which went and joined an organization and did much more cool stuff. However after I talked to grad college students that labored with them what I heard was, “I simply couldn’t get invested in that factor. It wasn’t my thought.”

As a professor, that was a pivotal second for me. From that time ahead, after I labored with college students, I attempted actually onerous to ask questions, and hear, and be excited and enthusiastic. However finally, my most profitable analysis initiatives had been by no means mine. They had been my college students and I used to be fortunate to be concerned. The factor that I don’t suppose I actually internalized till a lot later, working with groups at Amazon, was that one large contribution to these initiatives being profitable was that the scholars actually did personal them. As soon as college students actually felt like they had been engaged on their very own concepts, and that they might personally evolve it and drive it to a brand new end result or perception, it was by no means troublesome to get them to essentially put money into the work and the pondering to develop and ship it. They only needed to personal it.

And that is in all probability one space of my position at Amazon that I’ve thought of and tried to develop and be extra intentional about than anything I do. As a very senior engineer within the firm, in fact I’ve robust opinions and I completely have a technical agenda. However If I work together with engineers by simply making an attempt to dispense concepts, it’s actually onerous for any of us to achieve success. It’s lots more durable to get invested in an thought that you just don’t personal. So, after I work with groups, I’ve type of taken the technique that my finest concepts are those that different individuals have as a substitute of me. I consciously spend much more time making an attempt to develop issues, and to do a very good job of articulating them, relatively than making an attempt to pitch options. There are sometimes a number of methods to unravel an issue, and selecting the correct one is letting somebody personal the answer. And I spend lots of time being smitten by how these options are growing (which is fairly straightforward) and inspiring of us to determine how you can have urgency and go sooner (which is commonly a bit extra complicated). Nevertheless it has, very sincerely, been one of the rewarding components of my position at Amazon to method scaling myself as an engineer being measured by making different engineers and groups profitable, serving to them personal issues, and celebrating the wins that they obtain.

I got here to Amazon anticipating to work on a very large and sophisticated piece of storage software program. What I discovered was that each facet of my position was unbelievably larger than that expectation. I’ve discovered that the technical scale of the system is so monumental, that its workload, construction, and operations will not be simply larger, however foundationally totally different from the smaller methods that I’d labored on prior to now. I discovered that it wasn’t sufficient to consider the software program, that “the system” was additionally the software program’s operation as a service, the group that ran it, and the client code that labored with it. I discovered that the group itself, as a part of the system, had its personal scaling challenges and offered simply as many issues to unravel and alternatives to innovate. And at last, I discovered that to essentially achieve success in my very own position, I wanted to give attention to articulating the issues and never the options, and to search out methods to assist robust engineering groups in actually proudly owning these options.

I’m hardly accomplished figuring any of these things out, however I certain really feel like I’ve discovered a bunch up to now. Thanks for taking the time to hear.