Is it sensible to try to `` teach'' cloud computing?

Many years ago, I was teaching a rather polished course on distributed computing, and students seemed to like it very much: we looked at distributed programming models and algorithms, and the students learned to create new protocols of their own, solving problems like atomic multicast.  Mostly I taught from my own textbook, initially created back in 1996 on a prior sabbatical.

When distributed systems suddenly expanded enormously and started to be known as cloud infrastructure systems (I would date this roughly to 2005 when AWS was introduced), it seemed like an obvious moment to modernize that class and shift it to become a cloud computing course. 

What followed became a cycle: I would revise my distributed systems textbook, update all my slides,  and then the course would feel reasonably current: I did this in 2010, then again in 2012.

But today I'm facing an interesting puzzle: as part of my sabbatical I thought I might again attempt to revamp the class, but the more I learn about the current state of the art, the less clear it is that this is even possible!  Cloud computing may have become too large a topic, and too arcane, to permit a proper treatment.  Here are a few of the more important topics:

• Edge platforms.  Devices like iPhones and iPads were completely new back when the cloud emerged, and fairly simple.  Today, they are incredibly complex and very powerful platforms, and it would probably take a full year-long course to just cover the full range of technologies embedded into them.
• The network itself is increasingly sophisticated, and I use the word "increasingly" with some hesitation, because after decades of continuous evolution and change, it is hard for a person who thinks of the Internet in terms of the old TCP/IP structure to fathom quite how changed the modern network actually has become.  Today's network is a computational structure: it dynamically adapts routing, is able to model each separate enterprise as a distinct domain with its own security and quality of service, it can cache huge amounts of data, and it understands mobility and adapts proactively, so that by the time your car emerges from the tunnel, the network is ready to reconnect and deliver the next bytes of your children's videos.  With the push towards P4, the network is increasingly programmable: an active entity that computes on data flows running at rates of 100Gbs or more.
• Any serious cloud computing company operates a lot of data centers, at locations spread globally (obviously, smaller players lease capacity from data centers operated by specialists, then customize their slice with their own software layers).  Some systems are just limited-functionality cache and web service structures: simple points of presence; others are full-featured data warehouses that do extensive computation.   Thus the cloud is a heavily distributed structure, with a hierarchy.  Routing of client requests is heavily managed.
• Within any single data center we have layers of functionality: edge systems that run from cache and are mostly stateless (but this is changing), then back-end systems that track dynamic data, and compute engines that apply iterative computational steps to the flow: data arrives, is persisted, is compressed or analyzed, this creates new meta-data artifacts that in turn are processed, and the entire infrastructure may run on tens or hundreds of thousands of machines.
• Big data is hosted entirely in the cloud, simply because there is so much of it.  So we also have these staggeringly-large data sets of every imaginable kind, together with indices of various kinds intended to transform all that raw stuff into useful "content".
• We have elaborate scalable tools that are more and more common: key-value stores for caching (the basic MemCacheD model), transactional ones that can support SQL queries, even more elaborate key-value based database systems. 
• The cloud is a world of extensive virtualization, and virtualized security enclaves.  All the issues raised by multitenancy arise, and those associated with data leakage, ORAM models, and then technologies like ISGX that offer hardware remedies.
• Within the cloud, the network itself is a complex and dynamic creation, increasingly supporting RDMA communication, with programmable network interface cards, switches and routers that can perform aspects of machine learning tasks, such as in-network reductions and aggregation.
• There are bump-in-the-wire processors: NetFPGA and other ASIC devices, plus GPU clusters, and these are all interconnected via new and rather exotic high speed bus technologies that need to be carefully managed and controlled, but permit amazingly fast data transformations.
• File systems and event notification buses have evolved and proliferated, so in any given category one has an endless list of major players.  For example, beyond the simple file systems like HDFS we have ones that offer strong synchronization, like Zookeeper, ones that are object oriented, like Ceph, real-time versions like Cornell's Freeze Frame (FFFS), big-data oriented ones, and the list goes on and on.  Message bus options might include Kafka, Rabbit, OpenSlice, and these are just three of a list that could extend to include 25.  There are dozens of key-value stores.  Each solution has its special feature set, advantages and disadvantages.
• There are theories and counter-theories: CAP, BASE, FLP, you name it.  Most are actually false theories, in the sense that they do apply to some specific situation but don't generalize.  Yet developers often elevate them to the status of folk-legend: CAP is so true in the mind of the developers that it almost doesn't matter if CAP is false in any strong technical sense.
• We argue endlessly about consistency and responsiveness and the best ways to program asynchronously.  The technical tools support some models better than others, but because there are so many tools, there is no simple answer.
• Then in the back-end we have all the technologies of modern machine learning: neural networks and MapReduce/Hadoop and curated database systems with layers of data cleaning and automated index creation and all the functionality associated with those tasks.

I could easily go on at far greater length.  In fact I was able to attend a presentation on Amazon's latest tools for AWS and then, soon after, one for Microsoft's latest Azure offerings, and then a third one focused on the Google cloud.  Every one of these presentations reveals myriad personalities you can pick and chose from: Azure, for example, is really an Azure PaaS offering for building scalable web applications, an Azure fabric and storage infrastructure, an Azure IaaS product that provides virtual machines running various forms of Linux (yes, Microsoft is a Linux vendor these days!), a container offering based on Mesos/Docker, and then there is Azure HPC, offering medium-size clusters that run MPI supercomputing codes over Infiniband.  All of this comes with storage and compute managers and developer tools to help you build and debug your code, and endless sets of powerful tools you can use at runtime.

Yet all of this is also terribly easy to misuse.  A friend who runs IT for a big company was just telling me about their move to the cloud: a simple mistake ran up a $100k AWS bill one weekend by generating files that simply got bigger and bigger and bigger. On a single computer, you run out of space and your buggy code crashes.  In the cloud, the system actually has the capacity to store exobytes... so if you do this, it just works.  But the, on Monday the bill arrives.  Microsoft, Amazon and Google are vying for the best ways to protect your IT department against surprises, but of course you need to realize that the dashboard has those options, and enable them, much like credit card alerts from Visa or MasterCard.

So even cloud dashboards have become an elaborate topic.

Where does this story end?  The short answer is that it definitely isn't going to end, not for decades.  If anything, it will accelerate: with the trend towards Internet of Things, we'll be moving all sorts of mission-critical real-time systems into the cloud, and even larger data flows: self-driving cars, self-managed smart highways and cities and buildings, you name it. 

And you know what?  I don't see a way to teach this anymore!  I was just talking to Idit Keidar yesterday at Technion, quite possibly the world's very best distributed systems researcher, and she told be a little about her graduate class in distributed systems.  Five years ago, I might have thought to myself that wow, she should really be teaching the cloud, that students will insist on it.  Yesterday my reaction was exactly the opposite: when I resume teaching this fall at Cornell, I might just do what she's doing and just return to my real roots.  Cloud computing was a fun course to teach for a while, but the growth of the area has simply taken it completely out of the scope of what we as faculty members can possibly master and cover in any one class.

The CAP conjecture is dead. Now what?

CAP has been around now for something like 15 years.  There are some circles in which the acronym is even more famous than FLP or SAT!  But CAP was never a theorem (let's call it a "folk-theorem"), and by now we have more and more examples of systems that violate CAP. 

Yet even if CAP is false, it was also a fantastic rule of thumb that was incredibly valuable to developers tasked with building high performance distributed systems on the cloud.  If we start to teach students that CAP is false, what can replace it in this role?

A little historical context: CAP is short for "you can only have two from Consistency, Availabilty and Partition Tolerance",  This assertion was initially put forward by Eric Brewer in a PODC keynote talk he gave in 2000. The justification he offered ran along the following lines.  First, he observed that in today's increasingly globalized web systems, we invariably deploy services with high latency WAN links between them.  These are balky, so services are forced to make a choice: either respond right now based on data locally available, or await restoration of the WAN link.  A tradeoff that obviously favors availability over consistency, and already tells us that many web services will have to find a way to prevent responses that reflect stale data from being noticed by users, or if they are noticed, from causing problems.  He used the term "partition tolerance" for this kind of fault-tolerance (namely, giving a response even when some link is down).  Hence the P in CAP.

He suggested that we are favoring "A and P over C".

Then he observed that even in a single data center, if you want the highest possible levels of performance, you'll need to handle web requests on edge nodes that take their data from cache, without first talking a backend server first: again weaker consistency, but higher availability.  So again, we see the A, as a kind of synonym for rapid response.

So he looked at all the different pairing: C and A over P, C and P over A, A and P over C.  He concluded that in practice we always seem to find that by taking A and P we get the best scalability.  But CAP itself asserted that although other mixes work, you always have to pick the two you like best, and you'll erode the third.

Although the definitions of A and P are a bit strange (P seems to have a bit of A mixed in, and also seems to sometimes mean fault-tolerance, since partitions never really arise within a data center), CAP is sort of catchy.  But like many such acronyms, much depends on the details: if you try and give rigorous definitions, to turn CAP into a theorem (people have done so), you find that it only holds in some narrow situations.

The result is that as professors teaching cloud computing, we generally treat CAP as a clever acronym, but one that works mostly as a rule of thumb.  In some sense, CAP is a useful kind of folklore.

CAP took hold back in 2000 because at that time, companies like eBay and Amazon were struggling with the high costs of ACID transactions in systems that the database SQL programming model.  Scalable database performance poses issues that are much more nuanced than the ones Eric had in mind: there were puzzles of lock conflict, complexity, data pipelines with non-trivial asynchronous ordering requirements, etc.  But the bottom line is that performance of the big systems at eBay and Amazon was erratic and often strayed outside the magic 100ms target for web-service and web-page responses.  This is the point at which a human user feels that the system is "snappy" and everyone wants to be in the sub-100ms range.

So, the technology leaders at eBay began to tell their developers that it was absolutely fine to write SQL code as a starting point in web application design (a pragmatic decision: they people they were hiring mostly had extensive SQL experience from their database courses), but that once the SQL code was working, to weaken the transactions by turning them into a series of individual atomic actions.   eBay began to talk about the resulting development methodology using a new acronym: BASE, by which they meant "Basically Available, Softstate systems with Eventual Consistency."

Notice that BASE doesn't abandon consistency.  Instead, it points out to the developer that many web systems just don't need ACID guarantees to work correctly.   ACID and SQL are great for creating a quick prototype that will be robust and easy to debug, but then you "optimize" it by taking away the ACID properties, without breaking the behavior in ways that violate the specification.

Amazon embraced BASE too.  Around 2006, they decided to rebuild many of their core applications around a key-value technology called Dynamo, but SQL users found it hard to use, and by 2008 the adoption of Dynamo began to falter.  To make the transition easier, Amazon layered in a NoSQL API for Dynamo, called Dynamo-DB: now SQL code could run on Dynamo, but with weaker guarantees than for a full SQL system (for example, NoSQL lacks join operations), and Dynamo-DB happened to be especially well-matched to BASE.

So you can see from these examples why CAP would be such convenient way to motivate developers to make the switch: it more or less tells them that if they optimize their code using BASE, it won't scale properly.  Moreover, the eBay memos about BASE include step by step instructions to explain precisely how to go about doing it.

Today, fifteen years later,  we've discovered more and more ways to implement scalable, high performance cloud services, edge caches that maintain coherence even at global scale, fully scalable transactional key-value storage systems, and the list goes on.  Derecho is one example: it helps you build systems that are highly Available, Replicated and Consistent.  Call this ARC.

The cool twist is that with ARC, you get lock-free strong consistency, right in the cloud edge! This is because we end up with very fast replication at the edge, and every replica is consistent.  You do need to think hard about your update sources and patterns if you hope to avoid using locks, but for most applications that aspect is solvable because in the cloud, there is usually some sense in which any particular data item has just once real update source.  So there is an update "pipeline" and once you have consistency in the system, you end up with a wide range of new options for building strongly consistent solutions.

An example I like very much was introduced by Marcos Aguilera in Sinfonia.  In that system, you can make strongly consistent cache snapshots, and run your code on the snapshot.  At massive scale you have as many of these snapshots as needed, and transactions mostly just run at the edge.  But when you want to update the system state, the trick he suggested is this: run your transaction and keep track of what data it read and wrote (version numbers).  Now instead of just responding to the client, generate a "minitransaction" that checks that these version numbers are still valid, and then does all the writes.  Send this to the owner of the true database: it validates your updates and either commits by applying the updates, or rejects the transaction, which you can then retry.

Since so much of the heavy lifting is down when speculatively executing the transactions at the first step, the database owner has way less compute load imposed on it.  Then you can start to think about systems with sharded data that has different owners for each shard, and limits the transactions to run within a single shard at a time (the trick is to design the sharing rule cleverly).  Sinfonia and the follow-on systems had lots of these ideas.

ARC creates a world in which solutions like the Sinfonia one are easy to implement -- and the Sinfonia approach is definitely not the only such option.

I'm convinced that as we move the cloud towards the Internet of Things services, developers will need this model,  because inconsistency in a system that controls smart cars or runs the power grid can wreak havoc and maybe even would be dangerous.

So now I want to claim that ARC is the best BASE story ever!  Why?  Well, remember that BASE is about basically available, soft-state systems with eventual consistency.  I would argue that in examples like the Sinfonia one, we see all elements of the BASE story!

For example, a Sinfonia system is basically available because with enough consistent snapshots you can always do consistent read-only operations at any scale you like.  The state is soft (the snapshots aren't the main copy of the system: they are replicas of it, asynchronously being updated as the main system evolves).  And they are eventually consistent because when you do make an update, the mini-transaction validation step lets you push the update results into the main system, in a consistent way.  It might take a few tries, but eventually, should succeed.

What about the eventual consistency aspect of BASE in the case of ARC?   ARC is about direct management of complex large-scale strongly consistent replicated state.   Well, if you use ARC at the edge, back-end servers tend to be doing updates in a slightly time-lagged way: batched updates and asynchronous pipelines improve performance.  Thus they are eventually consistent, too: the edge queues an update, then responds to the end-user in a consistent way, and the back end catches up soon afterwards.

And ARC isn't the only such story:  my colleague Lorenzo Alvisi has a way to combine BASE with ACID: he calls it SALT, and it works really well.

So, feeling burned by CAP?  Why not check out ARC?  And perhaps you would like a little SALT with that, for your databases?

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.  

C++ just grew up, and wow, what a change!

When I was first learning to program (longer ago than I should admit, but this would have been around 1975), I remember what a revelation it was to start working in C, after first teaching myself languages like Basic, PL/I, SNOBOL and Fortran.  Back then, those were the choices.

When you code in C, you can literally visualize the machine instructions that the compiler will generate.  In fact most C programmers learn to favor loops like these:

     while(--n) { ... do something ... }

or even

    do { ... something ... } while(--n);

because when you code in this style, you know that the compiler will take advantage of the zero/non-zero condition code left in the processor register after doing the --n decrement, and hence will generate one fewer instruction than with a for-loop, where you generally end up with an increment/decrement followed by a separate test.

I fell in love with C.  What an amazing shift of perspective: a language that actually lets you control everything!

Ok, admittedly, only a computer science nut could care.  But as it happens, this little example gets to the core of something important about computer systems: we tend to want our systems to be as functional as possible, but also to want them to be as minimal as possible and as fast as possible.  That tension is at the core of the best systems research: you want a maximally expressive component in which every non-essential element is removed, so that the remaining logic can be executed blindingly rapidly (or with amazing levels of concurrency, or perhaps there are other goals).  This sense of minimalism, though, is almost akin to an artistic aesthetic.  Code in C was beautiful in a way that code in PL/I never, ever, was able to match.

Our programming languages have somewhat lost that purity.  When I first learned Python and Java and then C#, their raw power impressed me enormously.  Yet there is also something a little wrong, a little fast-and-loose, about these technologies: they are insanely expensive.  First, they encourage enormous numbers of very small method calls, which entail pushing return addresses and arguments to the stack, doing maybe two instructions worth of work, and then popping the environment.  Next, because of automated memory management, object creation is cheap and easy, yet costly, and there are unpredictable background delays when garbage collection runs.  Even if you code in a style that preallocates all the objects you plan to use, these costs arise anyhow, because the runtime libraries make extensive use of them.  Finally, the very powerful constructs they offer tend to make a lot of use of runtime introspection features: polymorphism, dynamic type checking, and dynamic request dispatch.  All are quite expensive.

But if you are a patient person, sometimes you get to relive the past.  Along came C++ 11, which has gradually morphed to C++ 14, with C++ 17 around the corner.  These changes to C++ are not trivial, but they are a revelation: for the first time, I feel like I'm back in 1975 programming in C and thinking hard about how many instructions I'm asking the compiler to generate for me. 

I'll start by explaining what makes the new C++ so great, but then I'll complain a little about how very hard it has become to learn and to use.

• A normal program spends a remarkable amount of time pushing stuff onto the stack and popping it back off.  This includes arguments passed by value or even by reference (in Java and C# a reference is like a pointer), return addresses, registers that might be overwritten, etc.  Then the local frame needs to be allocated and initialized, and then your code can run.  So, remember that cool C program that looked so efficient back in 1975?  In fact it was spending a LOT of time copying!  You can ask: how much of that copying was really needed?
  
  In C++  these costs potentially vanish.  There is a new notation for creating an alias: if a method foo() has an argument declared this way: foo(int& x), then the argument x will be captured as an alias to the integer passed in.  So the compiler doesn't need to push the value of x, and it won't even need to push a pointer: it literally accesses the caller's x, which in turn could also be an alias, etc. 
  
  With a bit of effort, foo itself will expand inline, and if foo is recursive but uses some form of constant expression to decide the recursion depth or pattern, the compiler can often simulate the recursive execution and generate just the data-touching code from the unwound method.
• With polymorphic method calls in Java and C#, a runtime dispatch occurs when the system needs to figure out the actual dynamic types of the variables used in a method invocation and match that to a particular entry point.  In C++, you get the same behavior but the actual resolution that matches caller and callee occurs at compile time.  Thus at runtime, only data-touching code runs, which is far faster.
• Although C++ now has dynamic memory management, it comes from a library implementing what are called smart pointers, which are reference-counter objects that C++ creates and automatically manages: when such an object goes out of scope, the compiler calls the destructor method and it decrements the reference count on the object, and then automatically destroys the object itself once the count goes to 0.  This gives a remarkable degree of control over memory allocation and deallocation, once you become familiar with the required coding style.  In fact you can take full control and reach a point where no dynamic allocation or garbage collection would ever take place: you preallocate objects and keep old copies around for reuse.  The libraries, unlike the ones in Java and C#, don't create objects on their own, hence the whole thing actually works.
• C++ can do a tremendous amount of work at compile time, using what are called constant expression evaluation and variadic template expansions.  Basically, the compile-time behavior of the language is that of a full program that you get to write, and that generates the code that will really execute at runtime.  All the work of type checking occurs statically,  many computations are carried out by the compiler and don't need to be performed at runtime, and you end up with very complex .h header files, but remarkably "thin" executables

So with so much fantastic control, what's not to love?

• The syntax of C++ has become infuriatingly difficult to understand: a morass of angle braces and & operators and pragmas about constants and constant expressions that actually turn out to matter very much.
• The variadic template layer is kind of weird: it does a syntax-directed style of recursion in which there are order-based matching operations against what often looks, at a glance, like an ambiguous set of possible matching method invocations.  A misplaced comma can lead the compiler off on an infinite loop.  At best, you feel that you are coding in a bizarre fusion of Haskall or O'CaML with the old C++, which honestly doesn't make for beautiful code.
• As a teacher who enjoys teaching object oriented Java classes, I'm baffled by the basic question of how one would teach this language to a student.  For myself, it has taken two years to start to feel comfortable with the language.  I look forward to seeing a really good textbook!  But I'm not holding my breath.
• The language designers have been frustratingly obtuse in some ways.  For example, and this is just one of many, there are situations in which one might want to do compile-time enumerations over the public members of a class, or to access the actual names the programmer used for fields and methods.  For obscure religious reasons that date back to the dark ages, the standards committee has decided that these kinds of things are simply evil and must never, ever, be permitted. 
  
  Why are they needed?  Well, in Derecho we have persistent template classes and would normally want to name the files that hold the persisted data using the name of the variable the data corresponds to, which would be some sort of scoped name (the compile-time qualified path) and then the variable name.  No hope. 
  
  And one wants to iterate the fields in order to automatically serialize the class.  Nope. 
  
  And one wants to iterate over the methods because this would permit us to do fancy kinds of type checking, like to make sure the client invoking a point-to-point method isn't accidentally invoking a multicast entry point.  Sorry, guy.  Not gonna happen.
• The language designers also decided not to support annotations, like the @ notation in Java, or the C# [ something ] notation that you can attach to a class.  This is a big loss: annotations are incredibly useful in polymorphic code, and as C++ gets more and more powerful, we get closer and closer to doing compile-time polymorphic logic.  So why not give us the whole shebang?
• There isn't any way to stash information collected at compile time.  So for example, one could imagine using variadic templates to form a list of methods that have some nice property, such as being read-only or being multicast entry points.  That list would be a constexpr: generated at compile time.  But they just don't let you do this.  No obvious reason for the limitation.

I could complain at greater length, but I'll stop with the remark that even with its many limitations, and even with the practical problem that compilers have uneven support for the C++ 17 features, the language on a tear and this is clearly the one language that will someday rule them all.

In our work on Derecho, we've become C++ nuts (lead by Matt Milano, an absolutely gifted C++ evangelist and the ultimate C++ nut).  Our API is not trivial and the implementation is quite subtle.  Yet under Matt's guidance, 95% of that implementation involves compile-time logic.  The 5% left that has to be done at runtime is incredibly "thin" code, like the while(--n) { ... } loop: code that genuinely has to execute at that point in time, that actually touches data.  And because this code tends to be very efficient and very highly optimized, just a few instructions often suffice.  All of Derecho may be just instructions away from the RDMA NIC taking control and moving bytes.

Wow...  I feel like it is 1975 again (except that there is no way to fool myself into believing that I'm 20 years old.  Sigh)  But at least we can finally write efficient systems code, after decades of writing systems code in languages that felt like plush goose-down quilts: thick and soft, and nowhere near reality...

Systems engineering viewed as a science

At Cornell University, where I've done my research and taught since 1982, we've always had a reputation for being a top theory department, and because Cornell is relatively small compared to some of our peers, and because many of my colleagues are famous for rigorous systems theory, this reputation includes systems.  I wouldn't call it a bad thing: I actually like to specify protocols as carefully as possible and to tease out a correctness argument, and my work has often benefitted from rigor.

And yet I'm reminded of something Roger Needham loved to stress.  You've probably heard the story, but Roger headed the most widely known systems group in Europe for many years, at Cambridge University, and he went on to found the Microsoft Research Laboratory in Cambridge.

Roger's health finally failed, and his colleagues gathered to celebrate Roger's accomplishments.  This came near the end of his life, when the cancer he was suffering from had left him very frail.  As a result, Roger was unable to physically attend, but he did send a video, and in this video we see him in a wheelchair holding an engineer's hard hat on his lap.  He starts out by saying a few words about his career (which included a wide range of work, some very theoretical, some very practical).  And then he puts the hat on.  Roger looks directly at the camera and says that he hasn't very long now, and wants people to remember him this way: wearing a hard hat, very much the engineer. 

For Roger, computer systems research is an engineering discipline first: that our primary obligation is to really build the things we invent, and to build production quality software that people will really want to use.  Roger loved ideas, but for him, an unimplemented idea was inferior to a working one, and a working system that people use was the gold standard.  Throughout his career, Roger repeatedly acknowledged the value of rigor and used theoretical tools to achieve rigor.  But the key thing is that for him, theory in the systems arena was a tool, secondary to the actual engineering work itself.  The real value resided in the artifact: the working system.

For me this was one of the most iconic images in my entire career: Roger wearing that hat and underscoring that we are at our best when we build things of high value: useful things, beautifully designed concrete implementations.  And there is a genuine esthetic of systems creation too: a meaningful sense in which great systems are beautiful things. 

Today, I see a huge need to elevate our view of systems engineering: rather than thinking of it as mere implementation work, we need to begin to appreciate the science and beauty of great systems work.  There is an inherent value to this act of creation: building really good software, that solves some really hard problem, and having it work, and be used.  We're at our best when we create artifacts.

When I visit people at the French air traffic control agency, I'm kind of awed that the software we built for them in 1990 is still working today without a single unmanaged failure, in 27 years.  The NYSE floor trading system I developed ran the show for ten years, and there was never a single disruption to trading in all that time.  Every time you install Oracle, watch closely: in the middle of the install script Oracle sets up my old Isis toolkit system as its network management tool.  And this has worked for nearly 30 years.  Nobody phones me to ask how to fix obscure bugs.  (Of course, once Derecho is finally out there, I bet many will call to ask how to migrate to it).

We built these really difficult engineered infrastructures with incredibly demanding specifications, came up with a beautiful and elegant model (virtual synchrony, which turns out to be a variation on the Paxos state machine replication model), and it actually worked!  This is as good as it gets.

Our field continues to struggle with the tension between its mathematically-oriented theory side and its roots in systems engineering.  One sees this again and again.  Papers are brushed to the side by conferences because the contributions are primarily about sound engineering and great performance: "minor and incremental" aspects in the eyes of some PC member (who has probably never coded a large C++ program in his life)!  Funding programs somehow assume that the massive data collection infrastructures that will be needed as hosting environments for their fancy new machine learning technologies are trivial and will just build themselves.  You write a reference letter for a student, and back comes an email query: yes, ok, the student is a gifted distributed systems engineer.  But do they have a talent for theory?  And if not, where's the originality?  I've become accustomed to that sinking feeling of needing to translate my enthusiasm for systems engineering to make sense to a person who just doesn't understand the creativity and originality that the best systems research demands.

So here's my little challenge to the community: we need to start going out of our way to seek out and accept papers on the best systems, and to deliberately validate and celebrate the systems engineering side of the field as we do so.  Over time, we need to recreate a culture of artifacts: by deliberate choice, we need to tilt the field back towards valuing systems to a greater degree.  Roger was right: we're at our best when we wear those hard hats. 

A few concrete suggestions:

• Let's start to write papers that innovate by revealing clever ways to build amazing things.
• Those of us in a position to do so should press program committees to set aside entire sessions to highlight real systems: practical accomplishments that highlight the underlying science of systems engineering. 
• We should be giving awards to the best academic systems research work.  Somehow, over time, prizes like the ACM software systems prize started going purely to massive projects done by big teams in industry (which is fine, if ACM wants to orient the prize that way).  But we also need prizes to recognize the best systems built in academic and research settings.  

Why do these things?  Because we're at our best when we create working code, and when we teach others to appreciate the elegance of the code itself: the science behind the engineering.  Systems building is what we do, and we should embrace that underlying truth.