Posted on

Thrift on Steroids: A Tale of Scale and Abstraction

Apache Thrift is an RPC framework developed at Facebook for building “scalable cross-language services.” It consists of an interface definition language (IDL), communication protocol, API libraries, and a code generator that allows you to build and evolve services independently and in a polyglot fashion across a wide range of languages. This is nothing new and has been around for over a decade now.

There are a number of notable users of Thrift aside from Facebook, including Twitter (mainly by way of Finagle), Foursquare, Pinterest, Uber (via TChannel), and Evernote, among others—and for good reason, Thrift is mature and battle-tested.

The white paper explains the motivation behind Thrift in greater detail, though I think the following paragraph taken from the introduction does a pretty good job of summarizing it:

As Facebook’s traffic and network structure have scaled, the resource demands of many operations on the site (i.e. search, ad selection and delivery, event logging) have presented technical requirements drastically outside the scope of the LAMP framework. In our implementation of these services, various programming languages have been selected to optimize for the right combination of performance, ease and speed of development, availability of existing libraries, etc. By and large, Facebook’s engineering culture has tended towards choosing the best tools and implementations available over standardizing on any one programming language and begrudgingly accepting its inherent limitations.

Basically, as Facebook scaled, they moved more and more away from PHP and the LAMP stack and became increasingly polyglot. I think this same evolution is seen at most startups as they grow into themselves. We saw a similar transition in my time at Workiva, moving from our monolothic Python application on Google App Engine to a polyglot service-oriented architecture in AWS. It was an exciting but awkward time as we went through our adolescence as an engineering culture and teams started to find their identities. Teams learned what it meant to build backward-compatible APIs and loosely coupled services, how to deprecate APIs, how to build resilient and highly available systems, how to properly instrument services and diagnose issues, how to run and manage the underlying infrastructure, and—most importantly—how to collaborate with each other. There was lots of stumbling and mistakes along the way, lots of postmortems, lots of stress, but with that comes the learning and growing. The payoff is big but the process is painful. I don’t think it ever isn’t.

With one or two services written in the same language and relatively few developers, it was easy to just stick with “REST” (in quotes because it’s always a bastardized version of what REST ought to be), sling some JSON around, and call it a day. As the number of tech stacks and integration points increase, it becomes apparent that some standards are important. And once things are highly polyglot with lots of developers and lots of services running with lots of versions, strict service contracts become essential.

Uber has a blog post on building microservices that explains this and why they settled on Thrift to solve this problem.

Since the number of service calls grows rapidly, it is necessary to maintain a well-defined interface for every call. We knew we wanted to use an IDL for managing this interface, and we ultimately decided on Thrift. Thrift forces service owners to publish strict interface definitions, which streamlines the process of integrating with services. Calls that do not abide by the interface are rejected at the Thrift level instead of leaking into a service and failing deeper within the code. This strategy of publicly declaring your interface emphasizes the importance of backwards compatibility, since multiple versions of a service’s Thrift interface could be in use at any given time. The service author must not make breaking changes, and instead must only make non-breaking additions to the interface definition until all consumers are ready for deprecation.

Early on, I was tasked with building a unified messaging solution that would help with our integration challenges. The advantages of a unified solution should be obvious: reusability (before this, everyone was solving the problem in their own way), focus (allow developers to focus on their problem space, not the glue), acceleration (if the tools are already available, there’s less work to do), and shared pain points (it’s a lot easier to prioritize your work when everyone is complaining about the same thing). Also, a longer term benefit is developing the knowledge of this shared solution into an organizational competency which has a sort of “economy of scale” to it. Our job was not just to ship a messaging platform but evangelize it and help other teams to be successful with it. We did this through countless blog posts, training sessions, workshops, talks, and even a podcast.

Before we set out on building a common messaging solution, there were a few key principles we used to guide ourselves. We wanted to provide a core set of tools, libraries, and infrastructure for service integration. We wanted a solution that was rigid yet flexible. We provide only a minimal set of messaging patterns to act as generic building blocks with strict, strongly typed APIs, and promote design best practices and a service-oriented mindset. This meant supporting service evolution and API iteration through versioning and backward compatibility, allowing for resiliency patterns like timeouts, retries, circuit breakers, etc., and generally advocating asynchronous, loosely coupled communication. Lastly, we had to keep in mind that, at the end of the day, developers are just trying to ship stuff, so we had to balance these concerns out with ergonomics and developer experience so they could build, integrate, and ship quickly.

As much as I think RPC is a bad abstraction, it’s what developers want. If you don’t provide them with an RPC solution, they will build their own, so we had to provide first-class support for it. We evaluated solutions in the RPC space. We looked at GRPC extensively, which is the new RPC hotness from Google, but it had a few key drawbacks, namely its “newness” (it was still in early beta at the time and has since been almost entirely rewritten), it’s coupled to HTTP/2 as a transport (which at the time had fairly limited support), and it lacks support for JavaScript (let alone Dart, which is what most of our client applications were being written in). Avro was another we looked at.

Ultimately, we settled on Thrift due to its maturity and wide use in production, its performance, its architecture (it separates out the transports, protocols, and RPC layer with the first two being pluggable), its rich feature set, and its wide range of language support (checking off all the languages we standardized on as a company including Go, Java, Python, JavaScript, and Dart). Thrift is not without its problems though—more on this in a bit.

In addition to RPC, we wanted to promote a more asynchronous, message-passing style of communication with pub/sub. This would allow for greater flexibility in messaging patterns like fan-out and fan-in, interest-based messaging, and reduced coupling and fragility of services. This enables things like the worker pattern where we can distribute work to a pool of workers and scale that pool independently, whereas RPC tends to promote more stateful types of services. In my experience, developers tend to bias towards stateful services since this is how we’ve built things for a long time, but as we’ve entered the cloud-native era, things are running in containers which are autoscaled, more ephemeral, and more distributed. We have to grapple with the complexity imposed by distributed systems. This is why asynchronous messaging is important and why we wanted to support it from the onset.

We selected NATS as a messaging backplane because of its simplicity, performance, scalability, and adoption of the cloud-native mentality. When it comes to service integration, you need an always-on dial tone and NATS provides just that. Because of Thrift’s pluggable transport layer, we could build a NATS RPC transport while also providing HTTP and TCP transports.

Unfortunately, Thrift doesn’t provide any kind of support for pub/sub, and we wanted the same guarantees for it that we had with RPC, like type safety and versioning with code-generated APIs and service contracts. Aside from this, Thrift has a number of other, more glaring problems:

  • Head-of-line blocking: a single, slow request will block any subsequent requests for a client.
  • Out-of-order responses: an out-of-order response puts a Thrift transport in a bad state, requiring it to be torn down and reestablished, e.g. if a slow request times out at the client, the client issues a subsequent request, and a response comes back for the first request, the client blows up.
  • Concurrency: a Thrift client cannot be shared between multiple threads of execution, requiring each thread to have its own client issuing requests sequentially. This, combined with head-of-line blocking, is a major performance killer. This problem is compounded when each transport has its own resources, such as a socket.
  • RPC timeouts: Thrift does not provide good facilities for per-request timeouts, instead opting for a global transport read timeout.
  • Request headers: Thrift does not provide support for request metadata, making it difficult to implement things like authentication/authorization and distributed tracing. Instead, you are required to bake these things into your IDL or in a wrapped transport. The problem with this is it puts the onus on service providers rather than allowing an API gateway or middleware to perform these functions in a centralized way.
  • Middleware: Thrift does not have any support for client or server middleware. This means clients must be wrapped to implement interceptor logic and middleware code must be duplicated within handler functions. This makes it impossible to implement AOP-style logic in a clean, DRY way.

Twitter’s Finagle addresses many of these issues but is solely for the JVM, so we decided to address Thrift’s shortcomings in a cross-platform way without completely reinventing the wheel. That is, we took Thrift and extended it. What we ended up with was Frugal, a superset of Thrift recently open sourced that aims to solve the problems described above while also providing support for asynchronous pub/sub APIs—a sort of Thrift on steroids as I’ve come to call it. Its key features include:

  • Request multiplexing: client requests are fully multiplexed, allowing them to be issued concurrently while simultaneously avoiding the head-of-line blocking and out-of-order response problems. This also lays some groundwork for asynchronous messaging patterns.
  • Thread-safety: clients can be safely shared between multiple threads in which requests can be made in parallel.
  • Pub/sub: IDL and code-generation extensions for defining pub/sub APIs in a type-safe way.
  • Request context: a first-class request context object is added to every operation which allows defining request/response headers and per-request timeouts. By making the context part of the Frugal protocol, headers can be introspected or even injected by external middleware. This context could be used to send OAuth2 tokens and user-context information, avoiding the need to include it everywhere in your IDL and handler logic. Correlation IDs for distributed tracing purposes are also built into the request context.
  • Middleware: client- and server- side middleware is supported for RPC and pub/sub APIs. This allows you to implement interceptor logic around handler functions, e.g. for authentication, logging, or retry policies. One can easily integrate OpenTracing as a middleware, for example.
  • Cross-language: support for Go, Java, Dart, and Python (2.7 and 3.5).

Frugal adds a second kind of transport alongside Thrift’s RPC transport for pub/sub. With this, we provide a NATS transport for both pub/sub and RPC (internally, Workiva also has an at-least-once delivery pub/sub transport built around Amazon SQS for mission-critical data). In addition to this, we built a SDK which developers use to connect to the messaging infrastructure (such as NATS) with minimal ceremony. The messaging SDK played a vital role not just in making it easy for developers to adopt and integrate, but providing us a shim where we could introduce sweeping changes across the organization in one place, such as adding instrumentation, tracing, and authentication. This enabled us to roll critical integration components out to every service by making a change in one place.

To support pub/sub, we extended the Thrift IDL with an additional top-level construct called a scope, which is effectively a pub/sub namespace (basically what a service is to RPC). We wrote the IDL using a parsing expression grammar which allows us to generate a parser. We then implemented a code generator for the various language targets. The Frugal compiler is written in Go and is, at least in my opinion, much more maintainable than Thrift’s C++ codebase. However, the language libraries make use of the existing Thrift APIs, such as protocols, transports, etc. This means we didn’t need to implement any of the low-level mechanics like serialization.

I’ve since left Workiva (and am now actually working on NATS), but as far as I know, Frugal helps power nearly every production service at the company. It was an interesting experience from which I learned a lot. I was happy to see some of that work open sourced so others could use it and learn from it.

Of course, if I were starting over today, things would probably look different. GRPC is much more mature and the notion of a “service mesh” has taken the container world by storm with things like Istio, Linkerd, and Envoy. What we built was Workiva’s service mesh, we just didn’t have a name for it, so we called it a “Messaging SDK.” The corollary to this is you don’t need to adopt bleeding-edge tech to be successful. The concepts are what’s important, and if enough people are working on the same types of problems in parallel, they will likely converge on solutions that look very similar to each other given enough time and enough people working on them.

I think there’s a delicate balance between providing solutions that are “easy” from a developer point of view but may provide longer term drawbacks when it comes to building complex systems the “right” way. I see RPC as an example of this. It’s an “easy” abstraction but it hides a lot of complexity. Service meshes might even be in this category, but they have obvious upsides when it comes to building software in a way that is scalable. Peter Alvaro’s Strange Loop talk “I See What You Mean” does a great job of articulating this dilemma, which I’ve also written about myself. In the end, we decided to optimize for shipping, but we took a principled approach: provide the tools developers need (or want) but help educate them to utilize those tools in a way that allows them to ship products that are reliable and maintainable. Throwing tools or code over the wall is not enough.

Posted on

If State Is Hell, SOA Is Satan

More and more companies are describing their success stories regarding the switch to a service-oriented architecture. As with any technological upswing, there’s a clear and palpable hype factor involved (Big Data™ or The Cloud™ anyone?), but obviously it’s not just puff.

While microservices and SOA have seen a staggering rate of adoption in recent years, the mindset of developers often seems to be stuck in the past. I think this is, at least in part, because we seek a mental model we can reason about. It’s why we build abstractions in the first place. In a sense, I would argue there’s a comparison to be made between the explosion of OOP in the early 90’s and today’s SOA trend. After all, SOA is as much about people scale as it is about workload scale, so it makes sense from an organizational perspective.

The Perils of Good Abstractions

While systems are becoming more and more distributed, abstractions are attempting to make them less and less complex. Mesosphere is a perfect example of this, attempting to provide the “datacenter operating system.” Apache Mesos allows you to “program against your datacenter like it’s a single pool of resources.” It’s an appealing proposition to say the least. PaaS like Google App Engine and Heroku offer similar abstractions—write your code without thinking about scale. The problem is you absolutely have to think about scale or you’re bound to run into problems down the road. And while these abstractions are nice, they can be dangerous just the same. Welcome to the perils of good abstractions.

I like to talk about App Engine because I have firsthand experience with it. It’s an easy sell for startups. It handles spinning up instances when you need them, turning them down when you don’t. It’s your app server, database, caching, job scheduler, task queue all in one, and it does it at scale. There’s vendor lock-in, sure, yet it means no ops, no sysadmins, no overhead. Push to deploy. But it’s a leaky abstraction. It has to be. App Engine scales because it’s distributed, but it allows—no, encourages—you to write your system as a monolith. The datastore, memcache, and task queue accesses are masked as RPCs. This is great for our developer mental model, but it will bite you if you’re not careful. App Engine imposes certain limitations to encourage good design; for instance, front-end requests and datastore calls are limited to 60 seconds (it used to be much less), but the leakiness goes beyond that.

RPC is consistently at odds with distributed systems. I would go so far as to say it’s an anti-pattern in many cases. RPC encourages writing synchronous code, but distributed systems are inherently asynchronous. The network is not reliable. The network is not fast. The network is not your friend. Developers who either don’t understand this or don’t realize what’s happening when they make an RPC will write code as if they were calling a function. It will sure as hell look like just calling a function. When we think synchronously, we end up with systems that are slow, fault intolerant, and generally not scalable. To be quite honest, however, this is perfectly acceptable for 90% of startups as they are getting off the ground because they don’t have workloads at meaningful scale.

There’s certainly some irony here. One of the selling points of App Engine is its ability to scale to large amounts of traffic, yet the vast majority of startups would be perfectly suited to scaling up rather than out, perhaps with some failover in place for good measure. Stack Overflow is the poster child of scale-up architecture. In truth, your architecture should be a function of your access patterns, not the other way around (and App Engine is very much tailored to a specific set of access patterns). Nonetheless, it shows that vertical scaling can work. I would bet a lot of startups could sufficiently run on a large, adequately specced machine or maybe a small handful of them.

The cruel irony is that once you hit a certain scale with App Engine, both in terms of your development organization and user base, you’ve reached a point where you have to migrate off it. And if your data model isn’t properly thought out, you will without a doubt hit scale problems. It’s to the point where you need someone with deep knowledge of how App Engine works in order to build quality systems on it. Good luck hiring a team of engineers who understand it. GAE is great at accelerating you to 100 mph, but you better have some nice airbags for the brick wall it launches you into. In fairness, this is a problem every org hits—Conway’s law is very much a reality and every startup has growing pains. To be clear, this isn’t a jab at GAE, which is actually very effective at accelerating a product using little capital and can sustain long-term success given the right use case. Instead, I use it to illustrate a point.

Peering Through the Abstraction

Eventually SOA makes sense, but our abstractions can cause problems if we don’t understand what’s going on behind the curtain (hence the leakiness). Partial failure is all but guaranteed, and latency, partitioning, and other network pressure happens all the time.

Ken Arnold is famed with once saying “state is hell” in reference to designing distributed systems. In the past, I’ve written how scaling shared data is hard, but with SOA it’s practically a requirement. Ken is right though—state is hell, and SOA is fundamentally competing with consistency. The FLP Impossibility result and the CAP theorem can prove it formally, but really this should be intuitively obvious if we accept the laws of physics.

On the other hand, if you store information that I can’t reconstruct, then a whole host of questions suddenly surface. One question is, “Are you now a single point of failure?” I have to talk to you now. I can’t talk to anyone else. So what happens if you go down?

To deal with that, you could be replicated. But now you have to worry about replication strategies. What if I talk to one replicant and modify some data, then I talk to another? Is that modification guaranteed to have already arrived there? What is the replication strategy? What kind of consistency do you need—tight or loose? What happens if the network gets partitioned and the replicants can’t talk to each other? Can anybody proceed?

Essentially, the more stateful your system is, the harder it’s going to be to scale it because distributing that state introduces a rich tapestry of problems. In practice, we often can’t eliminate state wholesale, but basically everything that can be stateless should be stateless.

Making servers disposable allows you a great deal of flexibility. Former Netflix Cloud Architect Adrian Cockcroft articulates this idea well:

You want to think of servers like cattle, not pets. If you have a machine in production that performs a specialized function, and you know it by name, and everyone gets sad when it goes down, it’s a pet. Instead you should think of your servers like a herd of cows. What you care about is how many gallons of milk you get. If one day you notice you’re getting less milk than usual, you find out which cows aren’t producing well and replace them.

This is effectively how App Engine achieves its scalability. With lightweight, stateless, and disposable instances, it can spin them up and down on the fly without worrying about being in an invalid state.

App Engine also relies on eventual consistency as the default model for datastore interactions. This makes queries fast and highly available, while snapshot isolation can be achieved using entity-group transactions if necessary. The latter, of course, can result in a lot of contention and latency. Yet, people seem to have a hard time grappling with the reality of eventual consistency in distributed systems. State is hell, but calling SOA “satan” is clearly a hyperbole. It is a tough problem nevertheless.

A State of Mind

In the situations where we need state, we have to reconcile with the realities of distributed systems. This means understanding the limitations and accepting the complexities, not papering over them. It doesn’t mean throwing away abstractions. Fortunately, distributed computing is the focus of a lot of great research, so there are primitives with which we can build: immutability, causal ordering, eventual consistency, CRDTs, and other ideas.

As long as we recognize the trade-offs, we can design around them. The crux is knowing they exist in the first place. We can’t have ACID semantics while remaining highly available, but we can use Highly Available Transactions to provide strong-enough guarantees. At the same time, not all operations require coordination or concurrency control. The sooner we view eventual consistency as a solution and not a consequence, the sooner we can let go of this existential crisis. Other interesting research includes BOOM, which seeks to provide a high-level, declarative approach to distributed programming.

State might be hell, but it’s a hell we have to live. I don’t advocate an all-out microservice architecture for a company just getting its start. The complications far outweigh any benefits to be gained, but it becomes a necessity at a certain point. The key is having an exit strategy. PaaS providers make this difficult due to vendor lock-in and architectural constraints. Weigh their advantages carefully.

Once you do transition to a SOA, make as many of those services, or the pieces backing them, as stateless as possible. For those which aren’t stateless, know that the problem typically isn’t novel. These problems have been solved or are continuing to be solved in new and interesting ways. Academic research is naturally at the bleeding edge with industry often lagging behind. OOP concepts date back to as early as the 60’s but didn’t gain widespread adoption until several decades later. Distributed computing is no different. SOA is just a state of mind.