Alvaro VidelaRead my Thoughts. Follow my Leads.

What We Talk About When We Talk About Distributed Systems

For quite some time now I’ve been trying to learn about distributed systems, and it’s appropriate to say that once you start digging, there seems to be no end to it, the rabbit hole goes on and on. The literature in distributed systems is quite extensive, with lots of papers coming from different universities, plus quite a few books to choose from. For a total beginner like me, it proved hard to decide what paper to read, or what book to buy.

At the same time, I’ve found that several bloggers recommend this or that paper that one must know in order to become a distributed systems engineer (whatever that means). So the list of things to read grows: FLP, Zab, Time, Clocks and the Ordering of Events in a Distributed Systems, Viewstamped Replication, Paxos, Chubby and on and on. My problem is that many times I don’t see a justification of why I should read this or that paper. I love the idea of learning just for knowledge’s sake, to satisfy curiosity, but at the same time one needs to prioritise what to read, since the day only has 24hs.

Apart from the abundance of papers and research material, as mentioned above, there are also lots books. Having bought quite a few of them and read chapters here and there, I started to see that a book that had a promising title was quite unrelated to what I was looking for, or that the content didn’t directly target the problems I would have liked to solve. Therefore, I would like to go over what I think are the main concepts of distributed systems, citing papers, books or resources where one can learn about them.

As I’m continuously learning while writing these words, please have some patience, expect some mistakes, and be aware that I will try to expand whatever I end up writing here.

Before we start, I must tell you that I have presented this blog post at various conferences, so here are the slides if you are interested:

And here’s a video of when I presented this talk at the Erlang User Conference in Stockholm:

Let’s start with the article:

Main Concepts

Distributed systems algorithms can be classified according to different kinds of attributes. Some classifications are: the timing model; the kind of interprocess communication used; the failure model assumed for the algorithm; and many others as we will see.

Here are the main concepts we will see:

  • Timing Model
  • Interprocess Communication
  • Failure Modes
  • Failure Detectors
  • Leader Election
  • Consensus
  • Quorums
  • Time In Distributed Systems
  • A Quick Look At FLP
  • Conclusion
  • References
  • Timing Model

    Here we have the synchronous model, the asynchronous model and the partially synchronous model.

    The synchronous model is the simplest one to use; here components take steps simultaneously, in what are called synchronous rounds. The time for a message to be delivered is usually known, and we also can assume the speed of each process, ie.: how long it takes for a process to execute one step of the algorithm. The problem with this model is that it doesn’t reflect reality so well, even less in a distributed system context, where one can send a message to another process, and wish that the stars are aligned for us, so the message arrives to said process. The good thing is that using this model it is possible to achieve theoretical results that later can be translated to other models. For example, due to the guarantees of this model about time, we could see that if a problem cannot be solved under these timing guarantees, then it would probably be impossible to solve once we relax them (think of a Perfect Failure Detector for instance).

    The asynchronous model gets a bit more complex. Here components can take steps in whatever order they choose and they don’t offer any guarantee about the speed in which they will take such steps. One problem with this model is that while it can be simple to describe and closer to reality, it still doesn’t reflect it properly. For example, a process might take infinitely long to respond to a request, but in a real project, we would probably impose a timeout on said request, and once the timeout expires we will abort the request. A difficulty that comes with this model is how to assure the liveness condition of a process. One of the most famous impossibility results, the “Impossibility of Consensus with one Faulty Process” belongs to this timing model, where it is not possible to detect if a process has crashed, or if the process is just taking an infinitely long time to reply to a message.

    In the partially synchronous model, components have some information about timing, having access to almost synchronised clocks, or they might have approximations of how long messages take to be delivered, or how long it takes a processes to execute a step.

    The book Distributed Algorithms by Nancy Lynch is actually organised in sections based on these timing models.

    Interprocess Communication

    Here we need to think about how processes in the system exchange information. They can do it by sending messages to each other, in the message passing model, or by using the shared memory model, where they share data by accessing shared variables.

    One thing to keep in mind is that we can use a message passing algorithm to build a distributed shared memory object. A common example in books is the implementation of a read/write register. We also have queues and stacks, which are used by some authors to describe consistency properties, like linearizabilty. We should not confuse shared memory as a way to share data between processes by accessing a shared variable with shared memory abstractions built on top of message passing, like the ones just mentioned.

    Back to the message passing model, we have another abstraction to consider when trying to understand algorithms: the kind of link used between processes (think of channels used to send messages back and forth between processes). These links will offer certain guarantees to the algorithm using them. For example there’s the Perfect Links abstraction that has reliable delivery and sends no duplicates; this abstraction also assures exactly once delivery. We can easily see that this abstraction doesn’t reflect the real world either, therefore there are other kinds of link abstractions used by algorithm designers when they try to design models that are closer to real systems. Keep in mind that even if the Perfect Links abstraction is not so real, it can still be useful, for example if we can prove a problem is not possible to solve even assuming perfect links, then we know of a whole bunch of related problems which might not be solvable as well. On the topic of links authors usually consider or assume FIFO message ordering, like in Zab

    Failure Modes

    I already wrote an article about failure modes in distributed systems but it is worth reiterating here. A property of a distributed system model is what kind of process failures are assumed. On the crash-stop failure mode, a process is assumed to be correct until it crashes. Once it crashes, it never recovers. There’s also the crash-recovery model, where processes can recover after a fault, in this case, some algorithms also include a way for the process to recover the state it had before before crashing. This can be done either by reading from persistent storage or by communicating with other processes in a group. It’s worth noting that for some group membership algorithms, a process that crashes and then recovers could not be considered as the same process that was alive before. This usually depends if we have dynamic groups or fixed groups.

    There are also failure modes where processes fail to receive or send messages, these are called omission failure modes. There are different kind of omissions as well, a process can fail to receive a message, or to send a message. Why does this matter? Imagine the scenario where a group of processes implement a distributed cache. If a process is failing to reply to requests from other processes on the group, even though it is able to receive requests from them, that process will still have its state up to date, which means it can reply to read requests from clients.

    A more complex failure mode is the one called Byzantine or arbitrary failures mode, where processes can send wrong information to their peers; they can impersonate processes; reply to other process with the correct data, but garble its local database contents, and more.

    When thinking about the design of a system, we should consider which kind of process failures we want to cope with. Birman (see Guide to Reliable Distributed Systems) argues that usually we don’t need to cope with Byzantine failures. He cites work done at Yahoo! where they concluded that crash failures are way more common than Byzantine failures.

    Failure Detectors

    Depending on the process failure mode and timing assumptions we can construct abstractions that take care of reporting to the system if a process has crashed, or if it is suspected to have crashed. There are Perfect Failure detectors that never give a false positive. Having a crash-stop failure mode plus a synchronous system, we can implement this algorithm by just using timeouts. If we ask processes to periodically ping back to the Failure Detector Process, we know exactly when a ping should arrive to the failure detector (due to the synchronous model guarantees). If the ping doesn’t arrive after a certain configurable timeout, then we can assume the other node has crashed.

    On a more realistic system, it might not be possible to always assume the time needed for a message to reach its destination, or how long it will take for a process to execute a step. In this case we can have a failure detector p that would report a process q as suspected if q doesn’t reply after a timeout of N milliseconds. If q later replies, then p will remove q from the list of suspected processes, and it will increase N, since it doesn’t know what the actual network delay between itself and q is, but it wants to stop suspecting q of having crashed, since q was alive, as it took longer than N to ping back. If at some point q crashes, then p will first suspect it has crashed, and it will never revise its judgement (since q will never ping back). A better description of this algorithm can be found in Introduction to Reliable and Secure Distributed Programming under the name “Eventually Perfect Failure Detector”.

    Failure Detectors usually offer two properties: completeness and accuracy. For the Eventually Perfect Failure Detector type, we have the following:

    Failure detectors have been crucial in solving consensus in the asynchronous model. There’s a quite famous impossibility result presented in the FLP paper mentioned above. This paper talks about the impossibility of consensus in asynchronous distributed systems where one process might fail. One way to go around this impossibility result is to introduce a failure detector that can circumvent the problem.

    Leader Election

    Related to the problem of failure detectors is that of actually doing the opposite, to determine which process hasn’t crashed and is therefore working properly. This process will then be trusted by other peers in the network and it will be considered as the leader that can coordinate some distributed actions. This is the case of protocols like Raft or Zab that depend on a leader to coordinate actions.

    Having a leader in a protocol introduces asymmetry between nodes, since non-leader nodes will then be followers. This will have the consequence that the leader node will end up being a bottleneck for many operations, so depending on the problem we are trying to solve, using a protocol that requires leader election might not be what we want. Note that most protocols that achieve consistency via some sort of consensus use a leader process and a set of followers. See Paxos, Zab or Raft for some examples.

    Consensus

    The consensus or agreement problem was first introduced in the paper Reaching Agreement in the Presence of Faults by Pease, Shostak and Lamport. There they introduced the problem like this:

    Fault-tolerant systems often require a means by which independent processors or processes can arrive at an exact mutual agreement of some kind. It may be necessary, for example, for the processors of a redundant system to synchronise their internal clocks periodically. Or they may have to settle upon a value of a time-varying input sensor that gives each of them a slightly different reading.

    So consensus is a problem of reaching agreement among independent processes. These processes will propose some values for a certain problem, like what’s the current reading of their sensor, and then agree on a common action based on the proposed values. For example, a car might have various sensors providing it with information about the breaks temperature levels. These readings might have some variations depending on the precision of each sensor and so on, but the ABS computer needs to agree on how much pressure it should apply on the breaks. That’s a consensus problem being solved in our everyday lives. The book Fault-Tolerant Real-Time Systems explains consensus and other problems in distributed systems in the context of the automotive industry.

    A process that implements some form of consensus works via exposing an API with propose and decide functions. A process will propose a certain value when consensus starts and then it will have to decide on a value, based on the values that were proposed in the system. These algorithms must satisfy certain properties, which are: Termination, Validity, Integrity and Agreement. For example for Regular Consensus we have:

    For more details on consensus please consult the original paper mentioned above. Also the following books are a great reference:

    Quorums

    Quorums are a tool used for designing fault-tolerant distributed systems. Quorums refer to intersecting sets of processes that can be used to understand the characteristic of a sytem when some processes might fail.

    For example if we have an algorithm where N processes have crash-failure modes, we have a quorum of processes whenever we have a majority of processes applying certain operation to the system, for example a write to the database. If a minority of process might crash, that is N/2 - 1 process crashes, we still have a majority of processes that know about the last operation applied to the system. For example Raft uses majorities when committing logs to the system. The leader will apply an entry to its state machine as soon as half the servers in the cluster have replied to its request of log replication. The leader plus half the servers constitute a majority. This has the advantage that Raft doesn’t need to wait for the whole cluster to reply to a log-replication RPC request.

    Another example would be: let’s say we want to limit the access to a shared resource by one process at a time. This resource is guarded by a set S of processes. Whenever process p wants to access the resource it needs to first ask permission to a majority of the processes S guarding the resource. A majority of processes in S grant access to the resource to p. Now process q arrives into the system and tries to access the shared resource. No matter which processes it contacts in S, q will never arrive to a majority of processes that will grant it access to the shared resource until the resource is freed by p. See The Load, Capacity, and Availability of Quorum Systems for more details.

    Quorums don’t always refer to a majority of processes. Sometimes they even need more processes to form a quorum for an operation to succeed, like in the case of a group N processes that can suffer Byzantine failures. In this case if f is the number of tolerated process failures, a quorum will be a set of more than (N + f) / 2 processes See Introduction to Reliable and Secure Distributed Programming.

    If you are interested in this topic, there’s a whole book dedicated to quorums in distributed systems:

    Quorum Systems: With Applications to Storage and Consensus

    Time in Distributed Systems

    Understanding time and its consequences is one of the biggest problems in distributed systems. We are used to the concept of events in our life happening one after the other, with a perfectly defined happened before order, but when we have a series of distributed processes, exchanging messages, accessing resources concurrently, and so on, how can we tell which process event happened before? To be able to answer these kinds of questions, processes would need to share a synchronised clock, and know exactly how long it takes for electrons to move around the network; for CPUs to schedule tasks, and so on. Clearly this is not quite possible on a real-world system.

    The seminal paper that discusses these issues is called Time, Clocks, and the Ordering of Events in a Distributed System which introduced the concept of logical clocks. Logical Clocks are a way of assigning a number to an event in the system; said numbers are not related to the actual passage of time, but to the processing of events by a node in a distributed system.

    There are many kinds of logical clocks algorithms, like Vector Clocks or Interval Tree Clocks.

    For a very interesting discussion on time in distributed systems I recommend reading the article There Is No Now by Justin Sheehy.

    I would claim that time and its problems in distributed systems is one of the crucial concepts to understand. The idea of simultaneity is something we have to let go. This is related with the old belief of “Absolute Knowledge”, where we used to think that such a thing as absolute knowledge was attainable. The laws of physics show us that even light requires some time in order to get from one place to another, so whenever it reaches our eyes, and it’s processed by our brains, whatever the light is communicating, it’s an old view of the world. This idea is discussed by Umberto Eco in the book Inventing the Enemy, chapter “Absolute and Relative”.

    A Quick look at FLP

    To finalize this article, let’s take a quick look at the Impossibility of Distributed Consensus with One Faulty Process paper to try to relate the concepts we have just learned about distributed systems.

    The abstract starts like this:

    The consensus problem involves an asynchronous system of processes, some of which may be unreliable.

    So we have an asynchronous system, where no timing assumptions are made, either on processing speed or the time required for messages to reach other processes. We also know that some of these process may crash.

    The issue here is that in usual technical jargon, asynchronous might refer to a way of processing requests, like RPC for example, where a process p sends an asynchronous request to process q, and while q is processing the request, p keeps doing other things, that is: p doesn’t block waiting for a reply. We can see that this definition is completely different from the one used in the distributed systems literature, so without having this knowledge, it’s quite hard to fully understand the meaning of just the first sentence of the FLP paper.

    Later in the paper they say:

    In this paper, we show the surprising result that no completely asynchronous consensus protocol can tolerate even a single unannounced process death. We do not consider Byzantine failures, and we assume that the message system is reliable—it delivers all messages correctly and exactly once.

    So the paper only considers the crash-stop failure mode discussed above (sometimes called fail-stop). We can also see that there are no omission failures, since the message system is reliable.

    And finally they also add this constraint:

    Finally, we do not postulate the ability to detect the death of a process, so it is impossible for one process to tell whether another has died (stopped entirely) or is just running very slowly.

    So we can’t use failure detectors either.

    To recap, this means that the FLP impossibility applies for asynchronous systems with fail-stop processors, with access to a reliable message system, and where detecting the death of a process is not possible. Without knowing the theory related to the different models of distributed systems, it might be possible that we miss many of these details, or we just interpret them in a totally different way from what the authors meant.

    For a more detailed overview of FLP please take a look at this blog post: A Brief Tour of FLP Impossibility

    Also, it is interesting to read the paper Stumbling over Consensus Research: Misunderstandings and Issues by Marcos Aguilera, which has a discussion about what it means for FLP to be an impossibility result for distributed systems (spoiler alert: is not the same level of impossibility as the halting problem).

    Conclusion

    As you can see, learning about distributed systems takes time. It’s a very vast topic, with tons of research on each of its sub-areas. At the same time implementing and verifying distributed systems is also quite complex. There are many subtle places where to commit mistakes can make our implementations totally broken under unexpected circumstances.

    What if we choose the wrong quorum and then our new fancy replication algorithm loses critical data? Or we choose a very conservative quorum that slows down our application without need, making us break SLAs with customers? What if the problem we are trying to solve doesn’t need consensus at all and we can live with eventual consistency? Perhaps our system has the wrong timing assumptions? Or it uses a failure detector unfit for the underlying system properties? What if we decide to optimise an algorithm like Raft, by avoiding a small step here or there and we end up breaking its safety guarantees? All these things and many more can happen if we don’t understand the underlying theory of distributed systems.

    OK, I get it, I won’t reinvent the distributed systems wheel, but with such vast literature and set of problems, where to start, then? As stated at the top of this article, I think randomly reading papers will get you nowhere, as shown with the FLP paper, where understanding the first sentence requires you to know about the various timing models. Therefore I recommend the following books in order to get started:

    Distributed Algorithms by Nancy Lynch. This book is kinda the bible of distributed systems. It covers the various models cited above, having sections with algorithms for each of them.

    Introduction to Reliable and Secure Distributed Programming by Christian Cachin et al. Besides from being a very good introduction, it covers many kinds of consensus algorithms. The book is full of pseudo-code explaining the algorithms which is a good thing to have.

    Of course there are many more books, but I think these two are a good start. If you feel you need to diver deeper, here’s the list of resources used in this article:

    References

    blog comments powered by Disqus