Should Derecho do optimistic early delivery?

In the Isis and Vsync systems, we supported a form of early delivery that could be described as "optimistic":  messages were delivered as soon as they arrived, before the system was certain safety had been achieved, enabling the application to start processing them, asynchronously with respect to the background exchanges of acknowledgements required to learn that stability had been reached.  

For example, suppose that a client process, P, asked a process Q to take some action.  Perhaps as a side effect of this, Q updates data replicated in group G with members {Q, R, S}.  In an optimistic case, the update multicast itself might return as soon as the multicast was in process, and the delivery events could occur as soon as ordering was determined.  Latency is thus minimal.

The trouble with optimistic early delivery is that a crash can erase the multicast at the first stages of such a protocol.  Thus there might be a window of a milisecond or so during which a power failure could kill some process after it sent (or received and delivered ) such a multicast, but before it was fully delivered at other members.  Perhaps we would end up with R having delivered (and acted upon the message), and use of the crash, S never gets a copy.

This is a bad behavior: normally, one wants failure atomicity.  Accordingly, the Isis or Vsync application developer who used these optimistic protocols was told to invoke a flush primitive before taking actions that might remain visible after a crash, like giving out cash from the ATM, or sending a reply to an external client application.  The flush primitive simply waited until stability was achieved, returning after the message had become safe.  Thus if R called flush before handing out cash, either the multicast stabilized and we distribute the money, or the crash happens while flush is waiting, and R is killed,  but no money was handed out, hence nothing bad has occurred.

The idea was that optimistic delivery plus flush would be equivalent to pessimistic delivery, but that during the period before flush was called, computing overlaps with background message passing and extra work can be done.  In Vsync and Isis, the totally ordered multicast was  optimistic, whereas the "safe" primitive was pessimistic.  You may have heard of vertical Paxos protocol: this is also a pessimistic protocol.  Pessimistic protocols in effect call flush before the application ever sees the message, so that the application layer is always safe.

As a teaching device, we would ask new Isis or Vsync developers to think about the way that buffered disk writes are done in a file system.  The application writes some text, but it ends up buffered first by the file I/O library, and then in the kernel.  Seconds can elapse before the data is on the disk.  The value of buffering is speed: performance is dramatically improved, far better than if fflush is called character by character or even line by line.

Today I'm revisiting this same question in Derecho.  Our basic OrderedSend is pessimistic, and pretty fast.  We also have a raw version of send that lacks ordering guarantees or failure atomicity and runs at the highest speeds possible.  But do we need an optimistic delivery too?

The case for optimistic deliver centers on this question: are there applications that need strong properties, yet also need the small degree of overlapped computing afforded by optimism?  A pessimistic delivery rarely delays more than a few microseconds relative to an optimistic one.  So you can deliver a small multicast after, say, 12us in an optimistic mode, but might have to wait 25us to deliver it with full safety (pessimistically).

In our experiments, the performance gap is most extreme for small messages.  Our impression, so far, is that these suffer from two kinds of delay.  First, with very small multicasts, the actual time needed to report stability through the SST subsystem is proportionally larger relative to data transfer delays, so when you compare a small message delivery to a large one, the wait for stabilization simply looks larger.  And then the other issue is that because these messages need to be buffered, Derecho runs out of space and at high rates, new incoming messages are delayed waiting for an admission slot.

In contrast, optimistic delivery mode lets Derecho hand off incoming messages to the application instantly, hence the system buffers only data that is in-flight, and won't incur this admission backlog buildup.  Stabilization takes the same amount of time, but by the time flush is invoked by a handler processing message m, there is more hope that m will have fully stabilized and that flush won't delay at all.

But keep in mind that with raw mode, there is no delay and no need to use flush at all.     The question becomes this: will people who would find optimistic delivery useful be content to use raw mode instead?  If so, the demand for optimistic mode would be low, and we should then omit the extra mechanisms and extra complexity.  Or do they need stronger guarantees when failures occur?  And if the latter, just what would those applications look like, that need atomicity but also need to shave 25us or so off the delivery delay?

Let us know if you have a great example of an application that would really benefit from optimistic early delivery.  For now, we know how to support optimism within Derecho, but are learning towards not doing so!

The ultimate Paxos (and Atomic Multicast) protocol

Many years ago, probably around 1986, Barbara Simons organized a workshop on replication at Asilomar, a California resort that had an iconic role in early systems research (it took a while, but eventually our community was too large to fit there). 

Her idea was to bring people together across areas: theory and practice, and within the practitioners, databases as well as other forms of distributed systems.  It was a good idea, and a great workshop.

In Jim Gray's talk, he pointed out that the interaction pattern we associate with the 2-phase commit protocol was seemingly universal, and that perhaps we were wrong to think of 2-phase commit as primarily a database construct.  Instead, he then asked, if 2-phase commit is really concerned with something deeper than database transactions, what is this deeper core concept?  The essential point of his talk was that the 2-phase pattern seen in 2PC clearly allowed a distributed system to "learn" something.  And from this, he suggested, that if  we could start to appreciate the minimal "knowledge" required for a distributed system to be correct, we could build a minimal implementation of that protocol and solve the problem once and for all.

A long time has passed.  Jim is no longer with us, and we know all about knowledge in distributed systems, and how processes learn.  Even so, Jim's question still sometimes comes up.  I mentioned this to Idit Keidar today over lunch, and her response was definitive: she convinced me that today, we really do (finally) understand the answer to Jim's question.    Her thinking centers on work she published in 2006 with Alex Shraer.

Here's what she explained.

First, she observed, distributed computing is ultimately about just one question: fault-tolerant consensus.  For her, Jim had noticed this, but was understanding it as something about the 2PC communication pattern rather than appreciating that the thing that matters more is consensus: moving from a state in which some decision is uncertain to one in which a decision has been reached.  As Idit views the question, one can transform consensus into other forms, we can route the messages in all sorts of patterns, but the bottom line is that either we achieve agreement via consensus or we simply aren't building systems that can make logically sound claims about their behavior. 

Next, she pointed out, agreement is trivial while failures aren't happening.  While a system remains stable, the author of a new piece of information simply needs to multicast it, and this accomplishes consensus.    If there might be many authors, you can circulate a token around them either before they send (as in the old Totem and Transis protocols) or after the messages arrive, to assign them a delivery ordering.

And so in Idit's view, the only hard issue is to handle failures.  Obviously, in the case of 2PC, particularly back in 1986, the field started out with a somewhat muddled understanding of this aspect: 2PC starts by stating that there is a set of processes that need to achieve agreement, but where does that set come from?  In a modern system like Derecho, the set itself is the output of a consensus layer.  2PC uses a crash failure model, but this is one of many models.  Back in 1986 people were very focused on models: crash failures, failstop failures, omission to send, omission to receive, timing errors, Byzantine models.

But these days all that really matters is to handle crash failures correctly.  Some organizations toy with Byzantine fault-tolerance, but honestly, I've never met anyone who had a service that actually came under a Byzantine attack.  Maybe the blockchain people will finally have that experience.

So let's focus on crash failures for a moment.  In light of the FLP theorem, we know now that without a perfect failure detector, protocols can't be proved correct: you can show safety, but not liveness.  In practice, we solve this by forcing a seemingly faulty process to leave the system and then, if you want, it can rejoin. 

So failure becomes a purely behavioral abstraction: if the majority of the system deems some process to be worthy of excluding, for whatever arbitrary reason it desires, than out goes the offending process, end of story. 

Where did the majority constraint come from?  The role of the majority restriction is to prevent logical partitioning. 

So, we end up with a rather dynamic form of membership: the system defines its own membership, excluding processes that seem faulty, and makes progress as long as it can maintain a dynamic form of majority.  To whit: at any point in time, some epoch defines the composition of the system.  In order to move to a new epoch, the systems needs agreement by the majority of whoever is in the active epoch.

So here's where all of this leads: Derecho is a concrete implementation of this new-age perspective on Jim's question.  Indeed, it is the ultimate story in the sense that the system is optimal in many different ways.

To appreciate this, you'll need to read the Derecho papers, but in a nutshell, the system maps consensus onto RDMA hardware in an exceptionally efficient way.  But the protocol itself runs consensus on the configuration of each epoch, and then uses a cheaper consensus protocol within an epoch (one that can assume membership is stable and that failures won't occur), leaving a very clean realization of distributed computing.

Indeed, Derecho is a constructive lower bound in every interesting dimension.  It is an optimal solution to the problem of distributed computing, using consensus-based methods.

Why do I make this claim?  Well, first, during a given epoch, the system is as quick to deliver messages as is possible: one can prove that any system that delivers messages with fewer "exchanges of information" between its processes is either incorrect, or at last at risk of needing to stop because of some single fault.

Next, one can show that Derecho's agreement on the state of each membership epoch is optimal.   Here, Derecho makes use of an all-to-all pattern of information exchange, and I asked Idit if she thought the protocol could be improved.  Idit pointed again to her 2006 papers with Alex.  Without prior knowledge of which processes failed, and how many failed, she explained, this pattern of information exchange is the quickest way to reach agreement on the membership of the next epoch.

Finally, we can show that Derecho makes progress with a failure model called <>P: eventually perfect failure detection.  Idit explained to me that in the past, people actually thought it might make sense to focus on progress with weaker detection models, like <>W, but that if you do so, you either end up with a fault model equivalent to <>P, or you end up with slower agreement protocols that look much more like Byzantine agreement.  So, for the style of quick progress Derecho is after, she explains, <>P is the right goal.  And Derecho is live with <>P, provided that no more than a minority of processes fail in any epoch.  Which again, is the best one can do.

Now, Idit is a theory person, but it is interesting for me as a practitioner to also think about practicalities.  As the old adage goes: "In theory, theory and practice are the same, but in practice, they differ."

As it happens, Derecho is optimal in a communications-engineering (networking) sense. 

Given that reliable networks always have a one-to-one acknowledgement based layer, reliable multicast over a tree is an optimal data dissemination pattern: if you try and disseminate multicasts over an unreliable 1-N protocol, the cost of detecting lost packets and resending them will be very high compared to a tree of unicast transfers.  (Perhaps optical networks with optically aggregated acknowledgements could offer a slight hardware speedup, but even this isn't clear, since the optical aggregator would itself be a tree). 

Now if you happen to be a cloud computing engineer, and read the above, what will jump to mind is heavy-tailed behaviors and multi-tenancy: Derecho protocols sometimes relay data, so if a relayer is very slow, everyone waits.  Paxos classically would avoid this using quorums: the system doesn't wait for the slowest process.  So how can I claim that Derecho is optimal in a practical sense if it doesn't use quorums?

There are a few answers.  A long-winded one would talk about work Fernando Pedone and his student, Parissa Jallali, did while visiting my group a few years ago.  Basically, they studied quorum behaviors in Paxos and discovered that while Paxos can briefly "outrun" a slow process, eventually work piles up and either flow-control causes the protocol to pause, or a shifting pattern of just who is the slow process switches the slow guy into the quorum and some previous quorum member becomes slow.  Either way, Paxos basically halts until everyone catches up and is back in sync.  So quorum patterns do not evade the disruption caused by heavy tailed behaviors.

Conversely, offloading the protocol into hardware actually can eliminate that issue, because the hardware is dedicated: the network spends 100% of its time communicating, so if you can describe a pattern of communication, then let it rip, the network is the ideal engine for moving data.  As it happens, Derecho generates deterministic data transfer schedules, hence given adequately programmable NICs we can hand the entire sequence of block transfers to the NIC.  So we can even make "optimal" use of the network, and since a NIC never sleeps or pauses, quorum behaviors aren't needed even if end-applications sometimes are a bit slow.

So a road that for me started around when Jim asked his question about 2PC seems to have reached its end: Derecho implements the ultimate Paxos protocols for atomic multicast (often called vertical Paxos) and for persisted memory (classic Paxos).  We could add an optimistic early delivery protocol with a flush, too, as in Isis and Vsync, but we decided to keep the system simple and omitted it: most people who would use that feature probably just want raw RDMC from the Derecho API, and this we do offer.

And so the Paxos problem is solved.  And you know what?  Its about time!  Idit feels that it was solved back in 2006, actually.  As for me, well, until I can write an application using the ultimate protocol, the problem is open.  But today, I finally can.  (I don't mind at all that Derecho employs a protocol that is actually extremely similar to our first Isis protocols from 1985.)

So should we all pack our bags and go home? 

Not quite yet.  First, there are other forms of distributed consistency, like convergent (gossip) protocols, self-stabilization, and of course, Byzantine Agreement.  It would be nice to see how those fit into this picture and whether there can be a single integrated story that combines all the elements.

A second issue is the engineering complexity of modern platforms.  I'll write a whole blog posting on this sometime soon, but suffice it to say that in a data center with physical topology (racks, switches, TOR switches, failure-independence domains...), GPU clusters, NetFPGA accelerators, Intel SGX protection enclaves... it just isn't obvious how to write code for such environments.  Derecho is just part of the answer, and not the whole story.

Then beyond all of this are dimensions we have yet to tackle in Derecho itself.  For example, even if Derecho is the ultimate data center protocol, is it also the ultimate WAN version?  As it happens, I suspect that this may be true too, but more attention to the question will be needed.  Anyhow, until we have it running, I won't believe the story even if I figure it out "in theory".  After all, in theory, theory and practice are the same...  but in practice, they are enormously different.

So I wouldn't despair: very likely we are finally at the end of the road for Paxos, but it certainly isn't the end of the road for distributed systems.  Watch this space: I can promise plenty of interesting new questions, and new answers, in years to come.

On systematic errors in complex computer programs

Society as a whole, including many computing professionals, seems to assume that mistakes made by unsupervised deep learning and other forms of machine-learning will be like Heisenbugs: rare, random, and controllable through a more professional code development and testing process. This belief underlies a growing tendency to trust computing systems that embody unknown (indeed, unknowable) machine-learned behaviors.

Why is this a concern?  

In many parts of computer science, there is a belief in "ultimate correctness": a perception that if we merely formalize our goals and our algorithms and successfully carry out a proof that our code accomplishes the goals, then we will have created a correct artifact.  If the goals covered a set of safety objectives, then the artifact should be safe to use, up to the quality of our specification.  Highly professional development practices and extensive testing strengthen our confidence; model-checking or other forms of machine-checked proofs that run on the real software carry this confidence to the point at which little more can be said, if you trust the specification.

Yet we also know how unrealistic such a perspective can be.  Typical computing systems depend upon tens or hundreds of millions of lines of complex code, spread over multiple modules, perhaps including firmware programs downloaded into the hardware.  All of this might then be deployed onto the cloud, and hence dependent on the Internet, and on the cloud data-center's thousands of computers. The normal behavior of such programs resides somewhere in the cross-product of the individual states of the component elements: an impossibly large space of possible configurations. 

While we can certainly debug systems through thorough testing, software is never even close to flawless: by some estimates, a bug might lurk in every few tens or hundreds of lines of logic.  The compiler and the operating system and various helper programs are probably buggy too.  Testing helps us gain confidence that these latent bugs are rarely exercised under production conditions, not that the code is genuinely correct.

Beyond what is testable we enter the realm of Heisen-behaviors: irreproducible oddities that can be impossible to explain: perhaps caused by unexpected scheduling effects, or by cosmic rays that flip bits, or by tiny hardware glitches.  The hardware itself is often better modelled as being probabilistic than deterministic: 1 + 1 certainly should equal 2, but perhaps sometimes the answer comes out as 0, or as 3.  Obviously, frequent problems of that sort would make a machine sufficiently unreliable that we would probably repair or replace it.  But an event happening perhaps once a week?  Such problems often pass unnoticed.

Thus, ultimate correctness is elusive; we know this, and we've accommodated to the limitations of computing systems by doing our best to fully specify systems, holding semi-adversarial code reviews aimed at finding design bugs, employing clean-room development practices, then red-team testing, then acceptance testing, then integration testing.  The process works surprisingly well, although patches and other upgrades commonly introduce new problems, and adapting a stable existing system to operate on new hardware or in a new setting can reveal surprising unnoticed issues that were concealed or even compensated for by the prior pattern of use, or by mechanisms that evolved specifically for that purpose.

The puzzle with deep learning and other forms of unsupervised or semi-supervised training is that we create systems that lack a true specification.  Instead, they have a self-defined behavioral goal: reinforcement learning trains a system to respond to situations resembling ones it has seen before by repeating whatever response dominated in the training data.  In effect: "when the traffic light turns yellow, drive very fast."

Thus we have a kind of autonomously-learned specification, and because the specification is extracted automatically by training against a data set, the learned model is inherently shaped by the content of the data set.

Train such a system on a language sample in which plurals always end in "s", and it won't realize that "cattle" and "calamari" are plural.  Train it on images in which all the terrorists have dark hair and complexions, and the system will learn that anyone with dark hair or skin is a potential threat.  Teach it to drive in California, where every intersection either has stop signs on one or both streets, or has a traffic signal, and it won't understand how to drive in Europe, where many regions use a "priority to the right" model, whereby incoming traffic (even from a small street) has priority over any traffic from the left (even if from a major road). 

Machine learning systems trained in this way conflate correlation with causation.  In contrast, human learning teases out causal explanations from examples.  The resulting knowledge is different from a knowledge model learned by training today's machine learning technologies, no matter how impressive the machine learning system's ability to do pattern matching.

Human knowledge also understands time, and understands that behavior must evolve over time.  Stephen Gould often wrote about being diagnosed as a young adult with a fatal circulatory cancer.  Medical statistics of the period gave him a life expectancy of no more than a few months, perhaps a year at best.  But as it happened, a new medication proved to be a true magic bullet: he was cured.   The large-population statistics were based on prior treatments and hence not predictive of the outcomes for those who received this new treatment.  The story resonated in Gould's case because in his academic life, he studied "punctuated equilibria", which are situations in which a population that has been relatively static suddenly evolves in dramatic ways: often, because of some significant change in the environment.  Which is precisely he point.

Those who fail to learn from the past are doomed to repeat it.  But those who fail to appreciate that the past may not predict the future are also doomed.  Genuine wisdom comes not from raw knowledge, but also from the ability to reason about novel situations in robust ways.

Machine learning  systems tend to learn a single set of models at a time.  They squeeze everything into a limited collection of models, which blurs information if the system lacks a needed category: "drives on the left", or "uses social networking apps".  Humans create models, revise models, and are constantly on the lookout for exceptions.  "Is  that really a  pile of  leaves, or has the cheetah realized it can hide in a pile of leaves?  It never did that before.  Clever cheetah!"   Such insights once were of life-or-death importance.

Today, a new element enters the mix: systematic error in which a system is programmed to learn a pattern, but overgeneralizes and consequently behaves incorrectly every time a situation arises that exercises the erroneous generalization.  Systematic error is counterintuitive, and perhaps this explains our seeming inability to recognize the risk: viewing artificially intelligent systems as mirrors of ourselves, we are blind to the idea that actually, they can exhibit bizarre and very non-random misbehavior.  Indeed, it is in the nature of this form of machine learning to misbehave in unintuitive  ways!

My concern is this: while we've learned to create robust solutions from somewhat unreliable components, little of what we know about reliability extends to this new world of machine-learning components that can embody systematic error, model inadequacies, or an inability to adapt and learn as conditions evolve.  This exposes us to a wide range of new failure modalities never before seen, and that could challenge the industry and the computer science community to overcome.  We lack systematic ways to recognize and respond to these new kinds of systematic flaws.

Systematic error also creates new and worrying attack surfaces that hackers and others might exploit.  Knowing how a machine learning system is trained, a terrorist might circulate some photoshopped images of him or herself with very pale makeup and light brown or blond hair, to bind other biometrics that are harder to change (like fingerprints, corneal patterns) with interpretations suggesting "not a threat".  Knowing how a self-driving car makes decisions, a hacker might trick it into driving into a pylon.

Welcome to the new world of complex systems with inadequate artificial intelligences.  The public places far too much confidence in these systems, and our research community has been far too complacent.  We need to open our eyes to the risks, and to teach the public about them, too.