Coevolving with new technology

A little bit of history

I remember the sinking feeling I had when the Communications Decency Act (CDA) passed, back in the 1990s. I was stunned. A veil was ripped off for me. I had grown up with the U.S. government doing lots of things I didn't understand but that sounded like things that were going to protect me. Our president was building up our defenses while also negotiating with the leader on the other side for disarmament and peace. (Though, our side didn't disarm?) I heard tales about how recreational drugs were going to destroy us, and the government was running these big campaigns to protect us all, because one taste is all it takes before your life will be destroyed. I was enough of a STEM student to be very confused why "drugs" will be the end of your life if you have one single taste, but everywhere around me people were drinking and smoking and not having such a bad result; how do you draw the line for such a thing? I went along with it, though, and sort of thought that smarter people than me must know the reasons for all this. Plus, as Paul Graham says, I was mindful of What you can't say. These topics I mention would get you instantly socially ostracized if you weren't on the right side of them.

The CDA was different, in that it covered a topic I do know about. I knew enough about computer software to see that one individual could set up and run a bulletin board system that could then host a billion people using it. I could picture how this would work, and whenever people talked about social forums, my mind's eye would fill in the gaps of what they are saying with a number of the details--servers, networks, databases, and so on. In the world of the CDA, these rooms are supposed to be monitored, so the picture in my mind can no longer be implemented by one person. The CDA vision was that a chat room would have maybe 1 chaperone for every 1000 users, so to support a chat room with a billion users, you would need one developer and one million chaperones. The CDA vision would make a basic chat room literally one million times more expensive to run than it was before the CDA.

The cost to humanity can be very high. I can't say what we are missing from social networking, because we live in a post-CDA world, and we don't know what we are missing out on by raising the bar on who can even run an experiment. I can given an example from another kind of high, however. Research and limited trials have begun, today, for mind-altering substances including a parade of the villains from my 1980s classrooms: ketamine, psilocybin, and LSD. This is a complex subject, but imagine for a moment that the new trials are at least sometimes heading in the right direction. If that is true, then the American people lost 60 years where we could not access something that, it is now looking like, can turn people's lives completely around. 60 years is a long time. My brother was born and then died during that 60-year window of dark-out.

When I think about how to regulate AI, we are in a similar situation to chatrooms of the 90s or psychedelics in the 60s. We don't yet know what is possible. We don't know what people will try, or what the results will be. Some of those experiments will lead to good results, and some of them will lead to bad results. How do we pursue these experiments while keeping ourselves safe?

My general sense, based on the history of technology and of governance to date, is that it's preliminary to have any blanket AI regulation right now that can do more good than harm. Instead, it is better to wait for those slow-burn problems that take a while to happen, but that don't get phased out all by themselves. Let me paint a picture of what this can look like, from my point of view as a technologist for four decades.

The concept of coevolution

I believe we can think of AI-based technology as co-evolving with humanity. We can think of technology as moving forward in a way akin to evolution, and we can approach our responses to that evolution based on our experience with other instances of coevolution that are by now very familiar.

The mycelial network. I enjoyed the new form of space travel in Star Trek: Discovery, but I have learned that there is a real-world mycelial network that is even more interesting. It turns out that fungi are everywhere. Tree root systems of the real world are permeated by mycelia that stabilize the soil and provide vital nutrients. All living plants include endophytic fungi that they can't live without. The larger ecosystem would not function as we know it without fungi recycling nutrients from dead plants and animals.

The living world would simply not exist without this weird third kingdom of life that permeates it in many different ways. Imagine, now, a national government trying to regulate the evolution of fungi, and the challenges they would face. We need fungi for human life as we know it to exist, but also fungi can harm us. There's no one simple way that fungi interact with us and with our environment, so there's also no one simple rule that fungi need to follow in order to be safe for us. In a territory like this, the perspective that seems best to me is one of growing together. We are here, and fungi are here, and we need to move forward by inches, with diversity in each kingdom and some way to respond if a change looks harmful.

The Butlerian Jihad. Movies like Dune are always grasping for a reason to make heroes with fantastically advanced technology still resort to low-grade sword fighting. In the case of the Dune universe, the in-story answer is called the Butlerian Jihad, and it's based on a real-world novelist named Samuel Butler. Butler explored the idea that the machines around us are developing from a selection mechanism that is similar to the natural selection of Charles Darwin.

Butler didn't limit this idea to AI. Butler meant the idea very broadly as applying to all of the machines we create and market to each other in order to automate our lives. People are always making machines that don't work very well; for a humorous example, check out Thoren Bradley's review of a 4-way splitting wedge. The machines that catch on are produced more widely and become the basis of the next round of machines. Any software engineer can tell you that new software is created by modifying old software. As a result, the development of machines follows the two ingredients that are needed for natural selection to occur: new entities are replicated from old ones, with a modest rate of mutation; and entities are selected for continued existence by some kind of selection criteria.

Butler's idea seems right to me. Machines are evolving along with us, and it doesn't really change anything that newer machines are doing tasks that used to be things only a human could do. On the contrary, the most helpful machines for humanity are those that remove some of our toil.

Of these two ideas, fungi are the scarier one to me. I know how to turn off a machine, but fungi work at a molecular level that cannot be directly controlled. And yet, fungi have not just gone okay but are an essential part of the world as we know it. Life can exist without fungi, but humans wouldn't be there. Yet, both fungi and machines have so far made human life much, much better than it ever was before.

How to regulate coevolving systems

If we follow the idea of coevolution, then we can think through the conditions where a U.S. intervention is likely to help more than it hurts.

Above all, one of the biggest reasons that we can feel safe about fungi, and less safe about other things, is that fungi exist in a large, diverse ecosystem where any damage will have a limited distance it can travel. It scares me more than any specific technological risk that several categories of technology are near monocultures right now. Chrome, Firefox, and Edge are the only real desktop web browsers; Facebook is ascendant for a certain kind of social media; Gmail is about the only email reader; and Amazon is by far the online market for physically shipped goods. These companies can cause tremendous harm with any mistakes they make, because those mistakes will propagate to all 8 billion of us within seconds of being launched. Monocultures are death traps, and I worry that we have so many of them right now.

For specific technological developments, it seems like there are three categories to bear in mind.

• In some cases, direct contact with a human and the new machine will quickly kill both of them. In a case like this, there's nothing useful for the regulation to do, because the machine will go away on its own all by itself. An example would be the Clippy assistant for Microsoft Word. It died all on its own.
• Many developments are unambiguously simple and positive. For these, as well, any attempt to regulate it is just going to reduce some of the benefits we gain. An example would be the original Google Web Search. Unless you've used some intranet search engine such as Atlassian's, you may have a hard time imagining how bad the web search engines used to be. It's a very good thing for all of us that Page and Brin were able to do their web search experiments without needing a lot of approvals. Imagine what kind of search engine we would have if it took 5-10 years to get an experiment approved, and if the search engine were only able to present results that incorporate today's suite of politically correct speech and representation. It would never have gotten off the ground, and very likely most people would assume that an effective web searcher just can't be made.
• That leaves the category that is neither of the above. Something that is harmful, but not harmful enough that people immediately balk in terror and shut the whole thing down voluntarily. An example in my mind is the slot machine. The slot machine doesn't do anything dramatic like shoot out laser beams or organize armies to go marching around. Yet, it does enough harm, to a significant number of people, that I would say there is a significant practical benefit that a little bit of regulation could establish.

Regulating slot machines is tough after the fact, but take a moment to imagine if someone had tried to regulate them before they took off. Imagine looking at a slot machine before they really caught on, and imagine deciding just based on its design that it might cause some harm. I posit that essentially no one could effectively figure out what the rules should be. The negative effects of slot machines cannot be understood from the technology, which after all is pretty simple. To understand the effects, and therefore to know what to do about them, one has to reason about things like: human psychology; the kinds of establishments that slot machines end up at; the kind of clientele that end up at these establishments; and the larger market developments that would lead to situations of the mass-produced slot machines that have a large enough effect to be worth doing anything about. Based on all of that, a lot of the regulation wouldn't even be about the technology itself; some of the slot machine regulation on the books today is about fairness and about the maximum rate of money that the player will lose. It seems better to me, though, to do things like require payment up front, and to limit payments to once per hour; likewise, it seems likely helpful to put limits on the visual effects and on the advertising claims that surround these machines. It's tricky and is subtle, and it can only be done after the fact, and even after the fact, a lot of the attempts aren't going to go well. I would say that even today, after decades of tinkering, regulation around slot machines is not yet really figured out.

The good news is that slot machines haven't destroyed the world just yet. They have been a slow burn, long in development, because anything that was hyper-bad simply wouldn't have caught on at all. As well, our existing social immune systems are carrying a lot of the slack. Each of us that sees someone else in the herd get caught up in a gambling addiction is quick to raise an alarm to the rest of us. The regulation can be better, but it's more than fast enough to settle in and spend multiple decades working out what that regulation should look like.

It strikes me that generative AI can be thought of the same way. Lots of things will be really good, and we don't want to wait one year to get them, much less 5, 10, or, based on U.S. history, 60 years to get the benefits. Lots of things will be super-bad, and almost everyone will stop using them immediately. Lots of things will seem somewhat bad, and most people will stop using them, but a few will carry on; that small fraction is actually good for humanity as a whole, because it increases diversity in the gene pool. Then in some tiny remaining pie slice, after all the previous cases are covered, there will be a harmful technology that wasn't weeded out by itself and that wasn't stopped by our more general-purpose defense mechanisms. That tiny pie slice is where regulation can help, and it won't be machines shooting laser beams. It will be something we haven't thought of, that hardly anyone is talking about, because we haven't even tried yet and don't have any information to go on.

The case for IPv4.1

In 2003, Dan Bernstein wrote about the IPv6 mess, focusing primarily on the lack of viable migration plan. Avery Pennarun wrote again about the problem in 2011, also with a focus on the lack of viable migration plan. Now it's been another decade, for a total of around 30 years on this project.

With the benefit of 20/20 hindsight, we now know that the IPv6 plan was never a very good one. If IPv6 had been accurately estimated to take 30+ years of migration, it would not have achieved the buy-in that it did, and most likely the designers would have kept going. The individuals involved are all top of class from what I can tell, but sometimes good people produce bad results, especially in a group context. I'm reminded, in all of this, about the Ada programming language, with the unusual way it was designed. I'm even more reminded of a certain saying about committees: "none of us is dumb as all of us".

Big migrations do not always go this way. The World Wide Web only needed nine years, if we count from Tim Berners-Lee's memo to the launch of the Google search engine. Under the right conditions, the whole Internet can upgrade in less than ten years.

The difference in these projects is easy to understand. Avery describes it as follows:

In short, any IPv6 transition plan involves everyone having an IPv4 address, right up until everyone has an IPv6 address, at which point we can start dropping IPv4, which means IPv6 will start being useful. This is a classic chicken-and-egg problem, and it's unsolvable by brute force; it needs some kind of non-obvious insight. djb apparently hadn't seen any such insight by 2002, and I haven't seen much new since then.

Here in 2024, it's not really too late to start on a backup plan. It's anyone's guess whether a more modest IPv4.1 proposal would finish faster than IPv6, but we can all be sure of one thing. Our chances are better if we have two dogs in the race.

How to do it

Here's the general idea of IPv4.1, just to show what I mean. It's been posted about before, so consider this to be my explanation rather than anything fundamentally new.

The core change for IPv4.1 is that the IP packet header needs room for a larger address. The packet format can be the one from the SIPP proposal, but with version "7" instead of "6", since 6 is now taken. When an address is written down, for example in a router's configuration UI, these 8-byte addresses can be written down the same way we are used to, just with 8 digits instead of 4. Existing addresses can be zero-padded on the left, e.g. 0.0.0.0.127.0.0.1.

Whenever two machines talk to each other, they use the existing IPv4 packet format if they both have 4-byte addresses. If either of them has an 8-byte address, then they switch to the IPv4.1 format. In this case, though, the packets still mean the same thing they always did! Part of the key to achieving a large migration is not to saddle it with additional ones.

8-byte addresses should be sufficient for this problem. The reason that IPv6 addresses use 16 bytes is that IPv6 plans for the machines on people's private internal networks to avoid reusing an address from anyone else's private network. This property will not be pursued for IPv4.1, and it would not do anything useful it were, because per the previous paragraph, all the packets in IPv4.1 are supposed to mean the same thing they did in IPv4.

Rollout

During the initial rollout of IPv4.1, many machines will not yet be able to talk to each other with the extended addresses. This will be much like the case with IPv6 right now.

From there, it will be a race. It's speculation which race will go faster, but IPv4.1 has some big advantages. It doesn't take any configuration file changes, and it doesn't require designing IP blocks. It definitely doesn't require thinking about link-local addresses, updates to how DHCP works, or any other fundamental aspects of networking. All that it takes for IPv4.1 to win is for software providers to add it into their latest version, and for the major nodes on the Internet to upgrade their software some time over the next 5-10 years. Major Internet nodes usually have to upgrade their software once in a while, already, due to security and compliance requirements.

We don't have to know for sure that IPv4.1 will win the race before we can make a decision about trying. We can simply observe that the problem is important, and we can observe that our chances are better with two ways to win.

Task-based build systems

Today I ran into a great book chapter by Erik Kuefler and Lisa Carey about Google's build system philosophy. It's all very good, but one part that stood out to me is the distinction between a task-based build system and an artifact-based build system. Ant is a task-based build system and lets you define a task in terms of other tasks. Google's build system, partially open sourced as the "Bazel" project, is based on artifacts. An important part of an artifact-based system, they explain, is that it honors build cache semantics. They write specifically the following:

For a remote caching system to work, the build system must guarantee that builds are completely reproducible. That is, for any build target, it must be possible to determine the set of inputs to that target such that the same set of inputs will produce exactly the same output on any machine. This is the only way to ensure that the results of downloading an artifact are the same as the results of building it oneself. Fortunately, Bazel provides this guarantee and so supports remote caching. Note that this requires that each artifact in the cache be keyed on both its target and a hash of its inputs—that way, different engineers could make different modifications to the same target at the same time, and the remote cache would store all of the resulting artifacts and serve them appropriately without conflict.

Why the M4 macro syntax goes so wrong

I've been hacking in M4 recently, and I ran across a great article about M4 by Michael Breen. I highly recommend it for anyone either using M4 (e.g. for Autotools or for Bison), or considering M4 for a new project (probably a bad idea). This observation caught my eye:

While m4's textual rescanning approach is conceptually elegant, it can be confusing in practice and demands careful attention to layers of nested quotes.

Christman Brown writes something similar, in his review of M4:

It is difficult to debug. I quickly found my default behavior when encountering something unexpected is to just increase the escape quotes around something. That often fixed it.

What's going on, here? A macro language is supposed to let you write normal text and then sprinkle some macro expansions here and there. It's supposed to save you from the tedium of dealing with a full-fledged general purpose programming language. Also, sometimes this strategy works out. With the C preprocessor, you write ordinary C code most of the time, and it works totally fine to occassionally call a macro. Why does this approach work in C but not in M4?

I think Michael Breen is onto something. Macros are a form of subroutine, and with a well designed syntax for subroutine call, you want it to be straightforward and thoughtless to invoke a subroutine and pass it some arguments. You want the arguments to feel just like text in any other part of your system. Think how it feels to write HTML and to put a div around part of your page. You don't have to do any special encoding to the stuff you put inside the div; you can, any time you want, take any chunk of your code with balanced tags and put that inside another pair of tags. With M4, the basic facility of a subroutine call, which you use all over the place, is somehow tricky to use.

M4 is not wrong to have a quoting mechanism, but where it goes wrong is to require quoting on the majority of subroutine calls. Here's what that looks like in practice. M4 uses function-call syntax to invoke macros, so a call looks like foo(a, b, c). That's a great syntax to try, because function calls are a ubiquitious syntax that users will recognize, but it has a problem for a text macro language in that the comma is already a common character to use in the a, b, and c arguments. Having observed that, M4 could and should have moved away from the function call notation and looked for something else. Instead, the designers stuck with the function calls and augmented it with an additional kind of syntax, quotation. In practice, you usually quote all of your arguments when you pass them to an M4 macro, like this: foo([a], [b], [c]). Only usually, however, and you have to think about it every time. The quotation and the macro call are two different kinds of syntax, and the user has to control them individually.

The reason it works out better for C's preprocessor is that the C language already has a way to quote any commas that occur in the arguments, and the preprocessor understands those quoting mechanisms. For example, with sqrt(foo(x, y)), the preprocessor understands that the comma inside the (x, y) part should not count as separating the parameters to sqrt. Programmers already write those parentheses without thinking about it, because the function-call notation for foo(x, y) already requires parentheses. Unfortunately, C does not do the right thing for an example like swap(a[i,j], a[i,j+1]), because it does not treat square brackets the way it treats parenthesis. It could and it should, however. None of this maps over to M4 very well, because usually the arguments to an M4 macro are not code, and so the author isn't going to naturally escape any commas that occur. The function-call syntax just doesn't work well for the situation M4 is intended to be used in.

Fixing the local problem

If we wanted to write a next-generation M4, here are some observations to start from:

• It is better if the quoting syntax is built into the subroutine call syntax. That way, users don't have to independently reason about both calls and quotes, and instead can just think about call-and-quote as a single thing that they do.
• Look to markup languages for inspiration, for example XML or Latex. The general feel we are going for is that you mostly write text, and then occasionally you sprinkle in some macro calls. That's a markup language!

Based on these, a good thing to try for an M4-like system would be to use XML tags for macro invocation. XML is a culmination of a line of tag-based markup languages starting with SGML and HTML, and it is generally state of the art for that kind of language. Among other advantages, XML is a small, minimal language that you can learn quickly, and it has explicit syntax for self-closing tags rather than some tags being self-closing and others not, in a context-dependent way depending on the schema that is currently in effect for a given file.

Latex's macro syntax is also very interesting, and it has a big advantage in usually saying each tag name just once (\section{foo}) rather than twice (<section>foo</section>). However, my experience with Latex is that I am in constant doubt how much lookahead text a macro invocation will really look at; the braces-based syntax is just a convention, and you never know for sure which macros really look at those conventions or not. That said, the general syntax looks like a promising idea to me if it were locked down a little more rather than being based on the Tex macro language. A similar approach was used in Scribe, a markup language designed by Brian Reid in the 70s.

What to use for now

As things stand, I don't think M4 really has a sweet spot. Old projects that want to have an ongoing roadmap should probably move away from M4. New projects should never use it to begin with. What are the options right now, without having to build a new textual macro language?

It's not a bad option to use a general-purpose language like Python or Java. If you follow the links from the PP generic preprocessor that is used in Pandoc, they tell you they are replacing their templating by more and more usage of Lua, a general purpose language. When you use a general-purpose language to generate text, you can use the normal library routines your language already supports, plus a mature library of lists, maps, structs, and iteration routines on top of them. An example of this direction is the Jooq library for generating SQL code.

Another strong approach is to use XML, possibly augmented by XSLT. An example would be the query help format of the GitHub Code Scanner, a format that I designed many years ago at a startup called Semmle. We had an existing syntax based on HTML with regex-based rewrites applied to the HTML file, and we had a problem that people were typo-ing the syntax without realizing it, resulting in help files that were sometimes unreadable and were often missing a large portion of the text. I explored a few options for getting us a more rigorous format with tool support for common errors, and I landed on XML, which I feel like worked out pretty well. In addition to the format itself being nice to work with, we got to tap into the existing XML ecosystem, for example to use Eclipse's excellent XML editor.

I briefly explored JSON as well, which is another minimal syntax that is easy to learn, but I quickly realized why they call XML a markup language. Unlike with JSON, XML lets you mostly write normal text, and then as a secondary thing, add special syntax--hence, "marking up" your text. XML is also a very mature system in general, so for example we could configure Eclipse (which was a viable tool back then!) to auto-complete tags and give you errors within the editor if you used tags that aren't allowed. If I were to rewrite how Bison's skeletons work, I think something based on XML would be tempting to try. Bison already uses this approach for its debug output. I'm not sure, though; XSLT looks pretty voluminous in practice.

Some of the best options are embedded in an existing general-purpose language. JSX is embedded in JavaScript, and Scribe is embedded in Scheme. I'm not sure how practical these are if you aren't already working in those environments, but if you are, look for one that works with your current source language.

The larger lesson

An effective tool has to be evaluated in the context it will be used in. Both C and M4 use function-call notation for macro invocation, but in C it works well, while with M4 it becomes a nightmare. Achieving an effective tool, therefore, requires design thinking. You need to learn about the situation you are targeting, you need to adjust your solution based on that, and above all you need to be ready to critique your solutions and iterate on improvements. The critique can take many forms, and a really important one is to watch how your users are doing and to really reflect on why it's happening and how it could be different.

Surprises can happen anywhere, and you'll support your users more if you can act on those surprises and try something different.

A Sampling Trick

Here's a sampling trick that comes up all the time in industry, but that took me a while to work out. The problem goes like this:

Suppose a server is receiving a sequence of events, for any kind of event you might be interested in, and you want to save off a sample of the events for later analysis. How do you select which events to sample?

Let me show some solutions that I see people try, and then describe a way that seems to beat all those out.

Solutions that don't work well

Choose a fixed percentage. One approach is to choose a percentage, say 5%, and to log that percentage of incoming events. For each event that arrives, you check a random number generator against your percentage, and independently decide for each event whether to log a sample or not. This approach is pretty good except that it's often hard to choose a percentage ahead of time. If you get thousands of events per second, then 5% is going to be way too many. If you get one event per minute, then 5% is going to be way too few. In many cases, you are going to use the same sampling code to run against multiple streams of events, for example to sample access to multiple databases, or to sample requests to multiple RPCs on the same server. As another example, you may run the code in both staging and production environments, and you'll need a different percentage for each environment.

Look at historical data rates. Another way is to use a percentage, but to dynamically adapt the percentage based on historical traffic rates. This way works in theory, but in my experience, systems like this tend to be unstable. For example, the system will dial in a low percentage based on overnight traffic, or because the system paused due to a garbage collection or a spate of disk I/O on the underlying machine. Then a burst of activity will happen, and the system will use its old percentage for a while, thus sampling way too much data. Then the percentage updates, but the burst of activity disappears. Now the server is sampling at a very low percentage and throwing everything away, because the burst of traffic disappeared. In general, I can't say that this method can't work, but in a world where software always has bugs and surprises, it's much better to use algorithms that are more clearly stable and predictable.

Use time-based sampling. Instead of sampling a percentage of traffic, the system can select one event for each window of time, for example one event every 10 seconds. This method adapts to different rates of incoming events, including bursty traffic where the rate isn't steady. Now there's a new problem, though: the events are not randomly selected any more! This method will work fine if the server is receiving a homogenous series of events that are all spaced out the same from each other, but that may or may not be the case. A lot of times, there is some external process, perhaps generated by an end user, where one external event will create a clump of events that arrive at your server at almost the same time as each other. If you select one event every 10 seconds, you will tend to sample events that are at the beginning of the clumps, thus making some of the events more likely to be sampled than others.

A dumb way that works

Here's an approach you could follow that will reliably work, but at the expense of a lot of memory usage. What you could do is collect all of the events for a 10-second window into an in-memory buffer. Each time a new event arrives, you don't think twice, you simply add it to the buffer to be looked at later on. Once the ten-second window expires, you look through the buffer and randomly select one of the events in the buffer. Then you reset the buffer and start again for the next 10-second window.

This way meets all of the constraints, but now it uses a lot of memory. It's close, though, except for the memory problem. Is there a way to do this, but with a smaller buffer? It would need to be some kind of algorithm that discards events as it goes, rather than waiting to the end of the 10-second window to do a whole lot of discarding all at once.

The sampling trick

Here's a way that works. Keep the following two pieces of information in memory.

• The number of events that have arrived, in the current 10-second window.
• One single event from the window. This is the event that will be sampled for sure if no other events arrive.

Given this setup, here's what you do when a new event arrives.

1. Increase the count of events by one.
2. Select a random number from 1 to the new, updated count of events.
3. If the random number is 1, then replace the event in the buffer by the new event; the new event is the one you need, and you know for sure you don't need the old event.
4. If the random number is anything other than 1, then discard the new event; you're done with that event for good and aren't going to use it.

When the 10-second window elapses, emit your current sample and then reset the buffer to its original state: 0 events that have arrived, and null as the currently stored event.

An example walk-through

To see why this works, let's start with a scenario that exactly three events arrive during the window. Here's what happens.

Event 1 arrives. When the first event arrives, increase the count to 1. Roll a number from 1 to 1, which will always result in 1, so store the first event into the buffer. Here's what is in the buffer now:

• Event 1, for sure.

Event 2 arrives. When the second event arrives, increase the count to 2. Roll a number from 1 to 2. This roll has a probability of 1/2 of being a 1, in which case event 2 will replace the first one. At this point, there is a 1/2 probability that event 2 took over the buffer, and a 1/2 probability that event 1 is still in there.

• Event 1, at probability 1/2.
• Event 2, at probability 1/2.

Event 3 arrives. Increase the count to 3, and roll a number from 1 to 3. This roll has a probability of 1/3 of being a 1 and causing event 3 to take over the buffer. At probability 2/3, some other number is rolled, and the existing event is brought forward. Therefore, the probability for event 1 to still be in the buffer is the 1/2 chance that the event already had, times the 2/3 chance that it survives this latest roll of the die. For event 2, the same reasoning applies as for event 1.

• Event 1, at probability 1/2 * 2/3 = 1/3.
• Event 2, at probability 1/2 * 2/3 = 1/3.
• Event 3, at probability 1/3.

Proof of the general case

Property. When using The Sampling Trick described above, after the arrival of N events, each event has 1/N probability of being the one in the sample buffer.

The world's smallest proof by induction. In the base case, suppose N=1. Then there is just one event, and it is always selected. The one event has probability 1/1, which is what we wanted to prove.

For the inductive case, assume the property is correct for N-1, and prove it for N.

For the Nth event, it was the last event that arrived, and the rules of the algorithm are that it will replace all prior events at probability 1/N. So the theorem is true for the Nth event.

For any other event, we know by the inductive assumption that it has a 1/(N-1) chance of being in the buffer when the Nth event arrives. The Nth event will replace the buffered event at probability 1/N, so the probability of the existing event being left alone is (N-1)/N. Therefore, the probability that the prior event is both selected from the first N-1 rounds, and also stays in the buffer after the Nth round, is 1/(N-1) x (N-1)/N, which is 1/N.

Q.E.D.

How do you use this in practice?

Sampling is helpful for always-on computer systems that can never have any down time. In a case like that, you need to not only make the server internally robust for every kind of input it might receive, but also be externally compatible with all the other software that's around it in the current production ecosystem.

There are a lot of tricks and know-how for how to use the sample data once you have it, but the first step is to capture the data at all.

Supporting C in a build system

In a previous post, I showed that the standard build rules for C code are unreliable. Let me describe two ways to do better.

In the interest of brevity, I will describe the build rules using a toy build engine called Blcache (short for "build cache"). I initially tried to write this post using standard tools like Make, Ninja, Bazel, and Nix, but they all have one limitation or another that distracts from the main point of this post.

Example problem

Here is an example problem this post will work against. There is a test.c file, and it includes two separate header files:

// File test.c
#include <stdio.h>
#include "syslimits.h"
#include "messages.h"

int main() {
  printf("%s: %d\n", MSG_LIMIT_THREADS, LIMIT_THREADS);
}

// File localhdrs/messages.h
#define MSG_LIMIT_THREADS "Limit on threads:"

// File hdrs/syslimits.h
#define LIMIT_THREADS 10

After compiling this code, a third header file is added as follows:

// File localhdrs/syslimits.h
#define LIMIT_THREADS 500

The challenge is for the addition of this header file to trigger a rebuild, while still making the build as incremental as possible.

Method 1: Compile entire components

The simplest way to get correct C compiles is to compile entire components at a time, rather than to set up build rules to compile individual C files. I've posted before on this strategy for Java, and it applies equally well to C.

This approach seems to be unusual, but my current feel is that it should work well in practice. It seems to me that when you are actively working on a given component, you should almost always use an IDE or other specialized tool for compiling that given component. The build system should therefore not concern itself with fine-grained incremental rebuilds of individual C files. Rebuilds of whole components--executables, libraries, and shared libraries--should be plenty, and even when using a system like Gyp, there are advantages to having the low-level build graph be simple enough to read through and debug by hand.

Using such an approach, you would set up a single build rule that goes all the way from C and H files to the output. Here it is in JSON syntax, using Blcache:

{
    "environ": [
        "PATH"
    ],
    "rules": [
        {
            "commands": [
                "gcc -Ilocalhdrs -Ihdrs  -o test test.c"
            ], 
            "inputs": [
                "test.c", 
                "localhdrs",
                "hdrs"
            ], 
            "name": "c/test", 
            "outputs": [
                "test"
            ]
        }
    ]
}

The "environ" part of this build file declares which environment variables are passed through to the underlying commands. In this case, only PATH is passed through.

There is just one build rule, and it's named c/test in this example. The inputs include the one C file (test.c), as well as two entire directories of header files (localhdrs and hdrs). The build command for this rule is very simple: it invokes gcc with all of the supplied input files, and has it build the final executable directly.

With the build rules set up like this, any change to any of the declared inputs will cause a rebuild to happen. For example, here is what happens in an initial build of the tool:

$ blcache c/test
Started building c/test.
Output from building c/test:
gcc -Ilocalhdrs -Ihdrs  -o test test.c

$ ./target/c/test
Limit on threads: 10

After adding syslimits.h to the localhdrs directory, the entire component gets rebuilt, because the localhdrs input is considered to have changed:

$ blcache c/test
Started building c/test.
Output from building c/test:
gcc -Ilocalhdrs -Ihdrs  -o test test.c

$ ./target/c/test
Limit on threads: 500

As a weakness of this approach, though, any change to any C file or any header file will trigger a rebuild of the entire component.

Method 2: Preprocess as a separate build step

Reasonable people disagree about how fine-grained of build rules to use for C, so let me describe the fine-grained version as well. This version can rebuild more incrementally in certain scenarios, but that benefit comes at the expense of a substantially more complicated build graph. Then again, most developers will never look at the build graph directly, so there is some argument for increasing the complexity here to improve overall productivity.

The key idea with the finer-grained dependencies is to include a separate build step for preprocessing. Here's a build file to show how it can be done:

{
    "environ": [
        "PATH"
    ],
    "rules": [
        {
            "commands": [
                "gcc -c -o test.o target/c/preproc/test.i"
            ], 
            "inputs": [
                "c/preproc/test:test.i"
            ], 
            "name": "c/object/test",
            "outputs": [
                "test.o"
            ]
        },

        {
            "commands": [
                "gcc -Ilocalhdrs -Ihdrs -E -o test.i test.c"
            ],
            "inputs": [
                "test.c", 
                "localhdrs",
                "hdrs"
            ], 
            "name": "c/preproc/test",
            "outputs": [
                "test.i"
            ]
        },

        {
            "commands": [
                "gcc -o test target/c/object/test.o"
            ],
            "inputs": [
                "c/object/test:test.o"
            ],
            "name": "c/test",
            "outputs": [
                "test"
            ]
        }
    ]
}

This file has three rules in it that chain together to produce the final output file. When you build the test executable for the first time, all three rules will be executed:

$ blcache c/test
Started building c/preproc/test.
Output from building c/preproc/test:
gcc -Ilocalhdrs -Ihdrs -E -o test.i test.c

Started building c/object/test.
Output from building c/object/test:
gcc -c -o test.o target/c/preproc/test.i

Started building c/test.
Output from building c/test:
gcc -o test target/c/object/test.o

$ ./target/c/test
Limit on threads: 10

First, the test.c file is preprocessed, yielding test.i. Second, the test.i file is compiled to test.o. Finally, test.o is linked into the final test executable.

Adding the new syslimits.h file behaves as expected, causing the full chain of recompiles.

$ blcache c/test
Started building c/preproc/test.
Output from building c/preproc/test:
gcc -Ilocalhdrs -Ihdrs -E -o test.i test.c

Started building c/object/test.
Output from building c/object/test:
gcc -c -o test.o target/c/preproc/test.i

Started building c/test.
Output from building c/test:
gcc -o test target/c/object/test.o

$ target/c/test
Limit on threads: 500

Modifying an irrelevant header file, on the other hand, only causes the precompilation step to run. Since the precompilation yields the same result as before, rebuilding stops at that point.

$ touch localhdrs/irrelevant.h
$ blcache c/test
Started building c/preproc/test.
Output from building c/preproc/test:
gcc -Ilocalhdrs -Ihdrs -E -o test.i test.c

Using cached results for c/object/test.
Using cached results for c/test.

It's not shown in this example, but since each C file is compiled individually, a change to a C file will only trigger a rebuild of that one file. Thus, the technique here is fine-grained in two different ways. First, changes to one C file only trigger a recompile of that one file. Second, changes to the H files only trigger preprocessing of all C files, and then only compilation of those C files that turn out to be affected by the H files that were changed.

By the way, there's a trick here that generalizes to a variety of cached computations. If you want to add a cache for a complicated operation like a C compile, then don't try to have the operation itself be directly incremental. It's too error prone. Instead, add a fast pre-processing step that accumulates all of the relevant inputs, and introduce the caching after that pre-processing step. In the case of this example, the fast pre-processing step is, well, the actual C preprocessor.

Coda

Before realizing the problem with C compilation, I used C as an example of why you might want to break the rules a little bit about the most strict and simplistic version of a build cache. However, now it seems to me that you find the best set of build rules if you strictly adhere to a build-cache discipline. I'm sorry, build cache. I should never have doubted you.

Standard build rules for C are unreliable

The standard way of integrating C into a build system is to use automatic dependencies generated from the compiler. Gcc and Clang can emit a list of the header files they read if you run them with the -M option. Visual Studio can do it as well, using the /showIncludes option. What I will call the "standard approach" in this post is to use the dependencies the user explicitly declared, and then to augment them with automatic dependencies generated by options like -M or /showIncludes.

Until a few years ago, I just took this approach as received wisdom and didn't think further about it. It's a neat trick, and it works correctly in the most obvious scenarios. Unfortunately, I have learned that the technique is not completely reliable. Let me share the problem, because I figure that other people will be interested as well, especially anyone else who ends up responsible for setting up a build system.

The root problem with the standard approach is that sometimes a C compile depends on the absence of a file. Such a dependency cannot be represented and indeed goes unnoticed in the standard approach to automatic dependencies. The standard approach involves an "automatic dependency list", which is a file listing out the automatically determined dependencies for a given C file. By its nature, a list of files only includes files that exist. If you change the status of a given file from not existing, to existing, then the standard approach will overlook the change and skip a rebuild that depends on it.

To look at it another way, the job of a incremental build system is to skip a compile if running it again would produce the same results. Take a moment to consider what a compiler does as it runs. It does a number of in-memory operations such as AST walks, and it does a number of IO operations including reading files into memory. Among those IO operations are things like "list a directory" and "check if a file exists". If you want to prove that a compiler is going to do the same thing on a second run as it did on the first, then you want to prove that those IO operations are going to do the same thing on a second run. That means all of the IO operations, though, not just the ones that read a file into memory.

Such a situation may seem exotic. At least one prominent source has declared that the standard approach is "correct" up to changes in the build command, which suggests to me that the author did not consider this scenario at all. It's not just a theoretical problem, though. Let me show a concrete example of how it can arise in practice.

Suppose you are compiling the following collection of files, including a single C file and two H files:

// File test.c
#include <stdio.h>
#include "syslimits.h"
#include "messages.h"

int main() {
  printf("%s: %d\n", MSG_LIMIT_THREADS, LIMIT_THREADS);
}

// File localhdrs/messages.h
#define MSG_LIMIT_THREADS "Limit on threads:"

// File hdrs/syslimits.h
#define LIMIT_THREADS 10

Using automatic dependencies, you set up a Makefile that looks like this:

CFLAGS=-Ilocalhdrs -Ihdrs

test.o test.d : test.c
 gcc $(CFLAGS) -M test.c > test.d
 gcc $(CFLAGS) -c test.c

test: test.o
 gcc -o test test.o

-include test.d

You compile it and everything looks good:

$ make test
gcc -Ilocalhdrs -Ihdrs -M test.c > test.d
gcc -Ilocalhdrs -Ihdrs -c test.c
gcc -o test test.o
$ ./test
Limit on threads: 10

Moreover, if you change any of the input files, including either of the H files, then invoking make test will trigger a rebuild as desired.

$ touch localhdrs/messages.h
$ make test
gcc -Ilocalhdrs -Ihdrs -M test.c > test.d
gcc -Ilocalhdrs -Ihdrs -c test.c
gcc -o test test.o

What doesn't work so well is if you create a new version of syslimits.h that shadows the existing one. Suppose you next create a new syslimits.h file that shadows the default one:

// File localhdrs/syslimits.h
#define LIMIT_THREADS 500

Make should now recompile the executable, but it doesn't:

$ make test
make: 'test' is up to date.
$ ./test
Limit on threads: 10

If you force a recompile, you can see that the behavior changed, so Make really should have recompiled it:

$ rm test.o
$ make test
gcc -Ilocalhdrs -Ihdrs -M test.c > test.d
gcc -Ilocalhdrs -Ihdrs -c test.c
gcc -o test test.o
$ ./test
Limit on threads: 500

It may seem picky to discuss such a tricky scenario as this one, with header files shadowing other header files. Imagine a developer in the above scenario, though. They are doing something tricky, yes, but it's a tricky thing that is fully supported by the C language. If this test executable is part of a larger build, the developer can be in for a really difficult debugging exercise to try and understand why their built executable is not behaving the way that's consistent with the source code. I dare say, it is precisely such tricky situations where people rely the most on their tools behaving in an intuitive way.

I will describe how to set up better build rules for this scenario in a followup post.

Two little things I wish Java would add

When geeking out about language design, it's tempting to focus on the things that require learning something new to even understand how it works. SAM types require understanding target typing, and type members require understanding path-dependent types. Fun stuff.

Aside from these things that are fun to talk about over beers, I really wish Java would pick up a few things from Scala that are just plain more convenient.

Multi-line string literals

A great way to structure a unit test is to feed in a chunk of text, run some processing that you want to verify, convert the actual output to text, and then compare it against another chunk of text that's included in the test case. Compared to a dense string of assertEquals calls, this testing pattern tends to be much easier to read and understand at a glance. When such a test fails, you can read a text diff at a glance and possibly see multiple different kinds of failure that happened with the test, rather than stare into the thicket of assertEquals calls and try to deduce what is being tested by the particular one that failed.

The biggest weakness of this style is very mundane: it's hard to encode a multi-line chunk of text in Java. You have to choose between putting the text in an external file, or suffering through strings that have a lot of "\n" escapes in them. Both choices have problems, although the latter option could be mitigated with a little bit of IDE support.

In Scala, Python, and many other languages, you can write a multi-line string by opening it with triple quotes (""") rather than a single quote mark ("). It's a trivial feature that adds a lot to the day to day convenience of using the language.

As one trick to be aware of, it's important to help people out with indentation when using triple quotes. In Scala, I lobbied for the stripMargin approach to dealing with indentation, where you put a pipe on each continuation line, and anything up to the pipe is considered leading indentation and removed. In retrospect, I wish I had pushed for that to simply be the default behavior. If you need to insert a literal continuation character, you can always write it twice. Making people write stripMargin on almost every multi-line string is a form of boilerplate.

Case classes

There are philosophers who disagree, but I find them a little too philosophical for my taste. Sometimes you really want to write a class that has no hidden internal state. Sometimes it would be a breach of the API to retain any internal state, or to implement the public API as anything other than plain old final fields. Some motivating examples are: tiny types, data structure nodes such as links in a linked list, and data-transfer objects.
In such a case, it takes a tremendous amount of code in Java to implement all the odds and ends you would really like for such a class. You would really like all of the following, and they are all completely mechanical:

• Constructors that copy their parameters to a series of final fields.
• A toString() implementation.
• Comparison operations: equals(), hashCode(), and compareTo(). Ideally also helpers such as isLessThan().
• Copy constructors that make a new version by replacing just one of the fields with a new value.

The equals() method is particularly painful in Java because there is a lot of advice going around about how to write them that is not consistent. I've been drug into multi-day debates on equals() methods where people cite things I published in the past to try and use against me; I'm pretty sure I meant what I said then and mean what I say now. Above all, though, I'd rather just have a reasonable equals() method and not spend time talking about it.

The mystics are coming out

Discussion on the Scala collections revamp is starting to get mystical. It really bugs me: good language design makes a huge difference, but it's hard to see unless you actually deal with thousands or more developers on millions or more lines of code. Casual observers of language design discussions don't see it themselves, so they don't realize what these problems look like to the other people in the discussion. So they think everyone is just goofing around and throwing out random ideas just because they are possible.

I started to follow up on the issue itself, but I fear making the problem worse. So I'll post a couple of rebuttals here. I don't really think the people actually involved in the redesign will be distracted by the mystics, anyway. They are like gold-sellers in MMOs or beggars in San Francisco; unless you are specifically trying to engage with them, you just learn to tune them out. Well, okay, they are like really nerdy gold sellers who like to talk about higher math. Okay I better just drop this attempt at an analogy.

First, Python's indexing operator has a lot of practical problems. Here are a few concrete examples: https://plus.google.com/+MattMight/posts/BVSmNadKni4 . Thinking backward from those puzzlers, I think the root of the problem is that the [a:b] slicing operator means something different depending on whether a and b are positive, and whether a is smaller or larger than b. This is foundational syntax used everywhere, and if you want code to be readable, people need to know whether it's doing a forward or reverse slice without having to do mental data flow analysis on the parameters. Java avoids this trap, and Scala should, too. The sublist operation should only work on non-negative arguments, and only when the first argument is smaller than the second.

The other thing I will say is that exceptions, null, and -1 are also all fine, when used by a tasteful library designer. We tried at Semmle to get people using more Options, and we found that it harmed our reputation as a quality lint tool. I can only speak publicly about open-source code, but to give an example, Apache Spark has over a thousand occurrences where they use null but, with only local changes, they could use Option instead. It's too many. It means that the basic premise of the programming advice has some sort of problem.

As one stab at it--though it's really a huge topic--you have to think about what you want the callers to do to defend against a missing value. If Scala's basic get/apply methods starts returning an Option, then people will just litter their code with calls to .get, so the net result is that the code is more bloated but otherwise behaves exactly the same. Even in pure math, people will write notation like f'(x) as the derivative of f, but you know, derivative isn't always defined. So should smart mathematicians instead write get(f')(x)? Or (f') match { ... }?

That's my try, but you don't even have to understand this if you are willing to ape mature libraries in areas that they are working okay. It's not a practical problem in Java that the various .get() methods return null or throw exceptions; even if you say it's not perfect, it's certainly not bad. It is, however, a very big problem that Java collections are overly mutable. For example, see this Stack Overflow question: https://stackoverflow.com/questions/2842169/why-are-public-static-final-array-a-security-hole. Scala will be more attractive to more developers if it focuses on these widely recognized pain points. Which is why the mystics drive me crazy--if they get their way they will find that their playground is gradually becoming a ghost town, but only after it's too late.

Initial input on the Scala collections redesign

It looks like Martin Odersky is considering an overhaul of Scala's collections:

A redesign of the standard library is on the roadmap for one of the next Scala versions (could be as early as 2.13). So I think now is the time to start thinking about what we want to change, in particular in what concerns collections!

Some more details are available on the Dotty issue tracker. Paul Phillips has weighed in on this issue with a very interesting slide deck.

I think it will be a very positive change if done carefully. It's painful to change such a fundamental library, but the current state is a real pain point for practical Scala code. If no other contender can get full support, I would even go so far as to suggest going back to Matthias Zenger's original version and working forward from there more conservatively this time. It was really a gem of useful functionality, in particular the "persistent" collections which used to be contraversial back then. Since that initial version, there have been several waves of changes that added complexity while only supporting minority use cases: lazily evaluated views, concurrent collections, and the CanBuildFrom magic. These are all valuable but should all be done in a separate library that is opted into when you need it.

I cannot put together a complete straw man implementation given my professional duties right now. I can make a few high-level suggestions, though, based on my experience in developer tools at Google and LogicBlox. Mind you, these are just initial reactions. I may well be overlooking important implementation details. Also, I haven't had the pleasure of using Scala for the new "big data" growth area, so I may be overlooking some concerns compared to the kinds of code I am used to.

At a high level, I'd want to see the following in an updated collection library:

• A focus on just the core collection types: maps, sets, linked lists, and some form of array-like sequences (maybe Array itself). Other collection types do come up in practice, but in a minority of contexts, and for those cases it should be fine to use a concrete collection type that does not necessarily inherit from any standard-library collection traits.
• Simple old-fashioned classes and methods, without any implicit parameters or type members. The status quo is bad in numerous ways, including problems with all of the following: IDE auto-completion, single-step debugging, and and compiler error messages.

Non-goals include the following:

• User extension of the library with new collection types. This is a small enough use case that it doesn't merit giving up on any other design goals of the library, e.g. performance.
• Supporting the majority of the collection types that are available now. I wrote a 1-2 page description of each of them for the latest Programming in Scala, and I found it quite the slog. The library has simply grown very large over time. It would be better if the less common collection types were more obviously in a side library. They're all useful sometimes, but the grand total is just overwhelming.
• Infinite collection types, in particular the lazy Stream type.

Some other things are desirable but might be unreasonable:

• High performance in common use cases involving a for() loop. This is a complicated problem, but it is a persistent problem area that drives people away from writing idiomatic Scala code. If nothing else, perhaps for() should translate to iterators instead of to higher-order functions. Another approach would be to have the compiler recognize and optimize standard types like List, Set, and Map; this will work better if those types are sealed rather than user extensible.
• Rename the three versions of Set and Map to have different names. For example, maybe AbstractSet, HashSet, and Set. There are numerous reasons for this; for example, automatic IDE import doesn't work as well if you have multiple classes with the same name. I put this in the "maybe" list, though, because it would require a lot of code to be rewritten. Even taking care to phase the change in over a period of months, it's a lot of code that would need to be turned over.

On a minor note, type Seq seems pretty useless. I know that Java has an interface for this, but maybe take a stand on this. For well-written code, it doesn't help anything to know that a type is a Seq and not just an Iterable. So delete the useless intermediate trait.

Ken Clark on shame mobs

Ken Clark has posted the top seven things he likes about shame mobs. Here's a taste:

5) Internet shame mobs weigh the evidence carefully and deliberately before attacking, so they only happen to people who deserve them. [...] 3) Internet shame mobs always make sure that the punishment is proportional to the crime.

There's a larger phenomenon here where problematic information spreads faster than the correction to it. If it spreads fast enough, then it can even pass a tipping point where it becomes very hard to get a hearing to object to any part of it. Everyone has already heard the idea from, well, everyone else, so they quickly dismiss anyone who objects without even really considering it.

The key to stopping such memetic chain reactions is to apply some filtering before propagating information that you read. It's still early days for the Internet, though, and we are all still learning to inoculate ourselves from being the wrong kind carrier.

There is some reason to have hope. Chain emails used to flourish, but are now mostly stamped out. In their heyday, 15-20 years ago, it was fairly common to open your mail program and see numerous messages that said something very exciting, and furthermore that the message should be passed on to everyone you know as fast as possible. Nowadays, the people I interact with just delete any email such emails. If an email explicitly says that it should be forwarded to everyone you know, then it triggers something like an antibody response. Such an email starts looking very bogus, and it gets deleted quickly and possible even flagged for followup by the email provider.

Intriguingly, people would likely not have developed that response had they not gone through the misery of dealing with chain emails earlier on. There are clear parallels to viruses and to building up antibodies!

Shame mobs are a place where it still goes on, though. I'm not entirely sure why it happens. In part, people just want to defend an idea, and they are happy to use real people as an example no matter the consequences. In part, people just enjoy being part of a mob. I hope that shame mobs go the same way as the chain email. We shall see.

Finding and fixing complicated code

It's hard to beat lines of code as a complexity metric: the subjective complexity of a piece of code seems to correlate very well with its length. Over the years, many other metrics have been devised and explored, but they tend to be more complicated yet less effective than the simple old lines-of-code.

One exception to the rule is McCabe Cyclomatic Complexity, a metric almost exactly as old as I am. McCabe can often find methods that aren't that many lines of long, but that nonetheless a typical maintainer would identify as being complicated and difficult to maintain. Small wonder that GNU Complexity puts so much emphasis on it.

What McCabe does is count the number of independent paths are needed to cover a method from beginning to end. The way McCabe described it, in his delightfully readable paper on the subject, is that it's the number of test cases you need to get statement coverage for a given method. It sounds like something that would be impractically hard to calculate--you are minimizing over arbitrary test cases!--but it's not complicated at all. He proved that, with a few extra assumptions, all you really have to do is count the number of branching options in a given body of code.

McCabe complexity is very effective in my experience. A method with complexity 10 is usually a dense, gnarly tangle of code, even if it's only 20 lines long. Yet, 20 lines of code is not enough to reliably be complicated just by its number of lines of code. If it's a simple linear sequence of non-branching code, it's usually reasonable to leave it alone. McCabe Complexity is therefore a real coup.

Even better, there's usually a straightforward way to act on code that has a poor McCabe complexity measure. If you see a method with a lot of branches in it, you can usually make the code more maintainable by factoring out some of the clusters of branches to their own smaller methods. The main method gets easier to understand, and if you do it tastefully, the smaller factored-out methods will also be easier to understand.

One place McCabe complexity falls down, though, is large switch statements. If you write network code that receives a message and dispatches it, it's perfectly fine style to write a switch statement for the dispatch with one case for each kind of message. This is especially true if each case simply calls a method that does the real work. Such a switch statement will easily have a dozen cases, one for each message type, but what alternative would be any more maintainable? Factoring out methods for subsets of the cases doesn't seem obviously better, and reducing the number of cases is often impossible.

McCabe is even worse for switch statements that consider pairs of types that are each drawn from some enum. To see some typical examples, look at Cast.scala in the Apache Spark source code. All the big switch statements in there are enumerating pairs of types, and they strike me as a very reasonable way to manage a problem that is just fundamentally complicated. This organizational style reminds me of how people organize definitions on paper when they are studying and explaining language semantics. Such people will typically define their relations and functions using a dozen or so cases, each of which has a number of assumptions (x is a class type, y is a class type, and x inherits from y), and a conclusion you can draw if all of the assumptions hold (x is a subtype of y). If this style is reasonable for theory papers, where people are expending enormous time and energy to bring a complicated subject within the reach of the human mind, it seems only reasonable to write such definitions in code when possible.

Overall, McCabe complexity is excellent and has stood the test of time. I wonder, though, if "number of branching statements" might be even more useful for anyone trying to simplify their code. "Number of branching statements" is the same as McCabe for if statements and while loops, but for switch statements, it only adds 1. Such a measure seems like it would better focus on code that is not merely complicated, but also possible to simplify.

Every build is special

Since Semmle's business involves integrating with our customers' builds, I'm very sympathetic to Tim Boudreau's perspective on simple plain-jane builds that just work:

But with time and enough consulting customers, you grow a real appreciation for projects you can simply check out from version control and build.

It's always a relief when someone wants to, say, build a stock iPhone app using xbuild. It's often fine when they want to build a C program using msbuild. Third place is Maven, for a simple Java program; I say third place, because hardly anyone uses it without loading it with a bunch of plugins that fundamentally change how it works. But still, at least the overall project structure is somewhat standardized.

In my modest experience, I only really see the above for relatively small and simple projects. As "Scott" puts it in the first comment at the above link:

The idea that your build is "not as special as you think" more often false than it is true.

I just skimmed a code base I have handy, and here are some of the things I see in there that break the mold of any off-the-shelf build tool:

• It uses parser generators, such as ANTLR.
• It's mainly a Java code base, but there is a little bit of code in all of: C, Objective C, JavaScript, Scala, C#, and Python.
• It uses a home-grown stack trace obfuscator and de-obfuscator that involves bytecode rewrites on the deployed jars.
• It includes its own version of zipmerge, with options not available in the open-source version.
• It uses the JIT support in mini-Lua to generate machine code that is then compiled into string literals in a C program.
• It includes a proprietary query language and numerous queries in that language. Those queries get compiled during the build, using a query engine that is also compiled during the same build.

Surveillance states are possible

While clamping down on private encryption is bad policy, both for the economy and for privacy, I don't think it's technically impossible to implement. Let me draw a couple of comparisons to show why.

---

As background, here is Cory Doctorow explaining, like many other commenters, that the Internet is too wild and wooly for major governments to possibly implement widespread surveillance:

For David Cameron's proposal to work, he will need to stop Britons from installing software that comes from software creators who are out of his jurisdiction. The very best in secure communications are already free/open source projects, maintained by thousands of independent programmers around the world. They are widely available, and thanks to things like cryptographic signing, it is possible to download these packages from any server in the world (not just big ones like Github) and verify, with a very high degree of confidence, that the software you've downloaded hasn't been tampered with.

With cellular phones, any phone that uses the relevant chunks of bandwidth is legally required to use certain protocols that are registered with the government. This has been bad economically, in that the telephone network has developed much more slowly than the relatively unregulated Internet. However, being bad economically has never exactly stopped rules from being put into place.

Yes, you can rig up a wireless network in your garage that breaks the rules. However, as soon as you try to use it over a wide geographic region, you're going to be relatively easy to catch. You will have to either broadcast a strong signal, or make use of the existing telephone backbone, or both.

To draw another comparison, consider the income tax. Income tax is easy to avoid with small operations, because you can just pay cash under the table. However, larger operations have to file a variety of paperwork, and the interlocking paperwork is what will get you. The more you take part in the above-ground economy, the harder it is to spin a big enough web of lies to get out of your taxes.

To get back to Internet protocols, it will certainly always be possible to break the rules on an isolated darknet you assemble in your garage. However, as soon as you send packets across the Internet backbone, any use of unregistered protocols is going to be very easy to detect.

To rub the point in further, don't forget that the authorities have no requirement to go after everyone who they detect doing something fishy. If they are anything like the American tax service, they'll randomly (or politically....) select people to target, and those people will then be required to undergo an audit at their own expense. If they survive the audit, the tax service just says "I'm sorry" and moves on to the next victim. Because of selective enforcement, law enforcement has no requirement to go after everyone using illegal encryption.

Of course all this is bad for the economy and for humanity's development at large. Don't oppose a cryptography clampdown because it's technically impossible, or you will look just as silly as the people that say DNS takedowns are technically impossible. Rather, oppose a cryptography clampdown because we don't want to live like that. We want to have private communications, and we want to allow innovation on the Internet. It's brand new, and if we clamp down on it, it will ossify in its current state the same way that the telephone network did.

Goodbye, Spark Plug

Don't take me wrong. I know a dog is just a dog, and a pet is just a pet. There are people reading this who have cancer, and there are some who have outlived their human children. On the scale of life challenges, I've just had maybe a 3/10.

Still, I would like to write a few words. It's a way to organize my thoughts, and a way to say goodbye. I promise the next post will be about programming or law or identity or the web, but that all seems rather dry to me today.

As all you pet owners know, you get a little Pavlovian jolt each time you help out your little ward and they reward you for it. For example, when they greet you at the door and run in circles. Or, when they learn your gait well enough to jump on each leg in time and then jump to the other one before you pick that foot up. When they're cold, and you blanket them up, and they let out a deep sigh of contentment. When there's a burr in their foot, and they plaintively hold it out so that someone with thumbs can do something about it.

Over time it really adds up. You become attuned to where you expect them to be when you walk into a room. You gain a second sense about which ruffles in a couch blanket have a dog under them. You expect if you plink through a few measures of a waltz, that you'll see two brown eyes peek around the corner to see what you're doing. After 18 years of that and then nothing, you are left with a lot of little voids that add up to one great big void.

Some animals go and hide when they become deathly sick, but this one did not. In his last hours he came to me to fix it. Dog or no, it was crushing to see such hope and confusion, yet so little ability to do anything about it.

To anyone out there facing this kind of question, let me tell you that I feel no unease at all about the decision to eschew blood samples, fluid IVs, antibiotics, and I didn't even ask what else to try and give him a little more time. I keep thinking, he was 18, and kidneys don't get better, and he had multiple other problems, anyway. Indeed, what I really wish I could go back in time on is delaying euthanasia for too long. I even had my mind made up, and I went to a 24-hour vet to do it, but I backed down when confronted with a thorough vet that wanted to rehash the dog's entire medical history. I thought I could just take him to our regular vet the next day, but the sun never rose for him again. Yes, hindsight is 20/20, but I wish I had insisted.

Goodbye, Spark Plug. I hope we did well for you.

---

P.S. -- Mom, you are very optimistic to think we can get this plant to bloom every December. We'll give it a try!

Is this the right server?

It's nice to see someone else reach the following conclusion:

"For those familiar with SSH, you should realize that public key pinning is nearly identical to SSH's StrictHostKeyChecking option. SSH had it right the entire time, and the rest of the world is beginning to realize the virtues of directly identifying a host or service by its public key."

Verifying a TLS certificate via the CA hierarchy is better than nothing, but it's not really all that reassuring. Approximately, what it tells you is that there is a chain of certification leading back to one or more root authorities, which for some reason we all try not to think about too much are granted the ultimate authority on the legitimacy web sites. I say "approximately", because fancier TLS verifiers can and do incorporate additional information.

The root authorities are too numerous to really have faith in, and they have been compromised in the past. In general, they and their delegates have little incentive to be careful about what they are certifying, because the entities they certify are also their source of income.

You can get better reliability in key verification if you use information that is based on the interactions of the actual participants, rather than on any form of third-party security databases. Let me describe three examples of that.

---

Pin the key

For many applications, a remotely installed application needs only communicate with a handful of servers back at a central site you control. In such a case, it works well to pin the public keys of those servers.

The page advocates embedding the public key directly in the application. This is an extremely reliable way of obtaining the correct key. You can embed the key in the app's binary as part of your build system, and then ship the whole bundle over the web, the app store, or however else you are transmitting it to the platform it will run on. Given such a high level of reliability, there is little benefit from pulling in the CA hierarchy.

As linked above, you can implement pinning today. It appears to be tricky manual work, though, rather than something that is built into the framework. As well, you don't get to ignore the CA hierarchy by doing this sort of thing. So long as you use standard SSL libraries, you still have to make sure that your key validates in the standard ways required by SSL.

---

Associate keys with links

The Y property deserves wider recognition, given how important hyperlinks are in today's world. Put simply, if someone gives you a hyperlink, and you follow that hyperlink, you want to reliably arrive at the same destination that the sender wanted you to get to. That is not what today's URLs give you.

The key to achieving this property is that whenever you transmit a URL, you also transmit a hash of the expected host key. There are many ways to do this, including the ones described at the above hyperlink (assuming you see the same site I am looking at as I write this!). Just to give a very simple example, it could be as simple as using URLs of the following form:

     https://hash-ABC123.foo.bar/sub/dir/foo.html

This particular example is interesting for being backward compatible with software that doesn't know what the hash means.

I don't fully know why this problem is left languishing. Part of it is probably that people are resting too easy on the bad assumption that the CA hierarchy has us covered. There's a funny mental bias where if we know nothing about a subject, and we see smart people working on it, the more optimistic of us just assume that it works well. Another part of the answer is that the core protocols of the world-wide web are implemented in many disparate code bases; SSH benefits from having an authoritative version of both the client and the server, especially in its early days.

As things stand, you can implement "YURLs" for your own software, but they won't work as desired in standard web browsers. Even with custom software, they will only work among organizations that use the same YURL scheme. This approach looks workable to me, but it requires growing the protocols and adopting them in the major browsers.

---

Repeat visits

One last source of useful information is the user's own previous interactions with a given site. Whenever you visit a site, it's worth caching the key for future reference. If you visit the "same" site again but the key has changed, then you should be extremely suspicious. Either the previous site was wrong, or the new one is. You don't know which one is which, but you know something is wrong.

Think how nice it would be if you try to log into your bank account, and the browser said, "This is a site you've never seen before. Proceed?"

You can get that already if you use pet names, which have been implemented as an experimental browser extension. It would be great if web browsers incorporated functionality like this, for example turning the URL bar and browser frame yellow if they see a a site is a new one based on its certificate. Each browser can add this sort of functionality independently, as an issue of quality of implementation.

In your own software, you can implement key memory using the same techniques as for key pinning, as described above.

---

Key rotation

Any real cryptography system needs to deal with key revocation and with upgrading to new keys. I have intentionally left out such discussion to keep the discussion simple, but I do believe these things can be worked into the above systems. It's important to have a way to sign an official certificate upgrade, so that browsers can correlate new certificates with old ones during a graceful phase-in period. It's also important to have some kind of channel for revoking a certificate, in the case that one has been compromised.

For web applications and for mobile phone applications, you can implement key rotation by forcing the application to upgrade itself. Include the new keys in the newly upgraded version.

FCC inches away from neutrality

The FCC’s latest proposal for network neutrality rules creates space for broadband carriers to offer “paid prioritization” services.[11] While the sale of such prioritization has been characterized as a stark and simple sorting into “fast” and “slow” traffic lanes,[12] the offering is somewhat more subtle: a paid prioritization service allows broadband carriers to charge content providers for priority when allocating the network’s shared resources, including the potentially scarce bandwidth over the last-mile connection between the Internet and an individual broadband subscriber. Such allocation has historically been determined by detached—or “neutral”—algorithms. The Commission’s newly proposed rules, however, would allow carriers to subject this allocation to a content provider’s ability and willingness to pay.

That's from a review on Standard Law Review a few months ago. I think this evolution in the FCC's approach will benefit the public.

It seems important to consider realistic developments of the Internet. Here's a thought experiment I've used for a long time, and that seems to be happening in practice. Try to imagine what goes wrong if a site like YouTube or Netflix pays--with its own money--to install some extra network infrastructure in your neighborhood, but only allows its own packets to go across that infrastructure. Doing so is a flagrant violation of network neutrality, because packets from one site will get to you faster than packets from another site. Yet, I can't see the harm. It seems like a helpful development, and just the sort of thing that might get squashed by an overly idealistic commitment to neutrality.

As a follow-on question, what changes if instead of Netflix building the infrastructure itself, it pays Comcast to do it? It's the same from a consumer's view as before, only now the companies in question are probably saving money. Thus, it's even better for the general public, yet it's an even more flagrant violation of network neutrality. In this scenario, Netflix is straight-up paying for better access.

It seems that the FCC now agrees with that general reasoning. They not only support content delivery networks in general, but now they are going to allow generic ISPs to provide their own prioritized access to sites that pay a higher price for it.

I believe "neutrality" is not the best precise goal to go for. Rather, it's better to think about a more general notion of anti-trust.

Three tiers of classrooms

Via Arnold Kling, I see Jesse Rothstein trying to prove that you can't measure teaching ability, or perhaps even that teaching ability doesn't matter:

Like all quasi-experiments, this one relies on an assumption that the treatment – here, teacher switching – is as good as random. I find that it is not: Teacher switching is correlated with changes in students’ prior-year scores.

It's important to figure out which kind of classroom we are talking about. There are at least three tiers of classroom styles. If you measure only in the middle tier, then I can believe that teacher skill would have only a small effect. However, it's really easy to tell the difference between the tiers if you look, especially for the bottom-most tier compared to the other ones.

At the bottom tier, some classes are just zoos. The teacher is ignored, and the students talk to each other. At best, they work on material for another class. Teacher skill doesn't matter within this tier, from at academic perspective; one zoo teaches students just as much as another zoo. I am sad to say that classrooms like this do exist. It's a potential bright note that such teachers are very easy to identify in an objective way; their students have absolutely terrible results on standardized tests such as Advanced Placement (AP). There's no need for sophisticated statistics if all the students are scoring 1-2 out of 5.

At the middle tier, some classes involve the teacher walking the students through standardized textbooks and other material. Basically, the textbooks are software and the teachers are the hardware that runs it. It's not an inspiring kind of classroom, but at least it is inexpensive. Within this tier, I could see teacher skill not mattering much, because the students spend all their time glued to the course materials. However, you'd certainly like to find out who is in this tier versus in the zoo tier.

Worth a brief mention is that there's an upper tier as well. Maybe "style" is a better word in this case. Sometimes the teacher actually understands the course material, and so is able to respond to the questions with anecdotes and exercises that are tailored for that particular student. For this tier, teacher evaluation is especially important. Among other things, some teachers are fooling themselves, and would be better off staying closer to the book.

Edoardo Vacchi on attribute grammars

I previously wrote that predictable performance is a practical challenge for using attribute grammars on real work. It does little good to quickly write the first version of a compiler pass if you then spend hours debugging oddball performance problems.

Edoardo Vacchi wrote me the following in response. I agree with him: having an explicit evaluation construct, rather than triggering attribute contributions automatically, is likely to make performance more predictable. UPDATED: edited the first paragraph as suggested by Edoardo.

Hi,

This is Edoardo Vacchi from Università degli Studi di Milano (Italy). For my PhD thesis I’m working on a language development framework called “Neverlang”[1,2]. Neverlang is an ongoing project of Walter Cazzola's ADAPT Lab; I am involved with its latest incarnation "Neverlang 2".

I stumbled upon an (old) blog post of yours about Attribute Grammars [3] and I would be interested to know if you knew some “authoritative” references that I could cite with respect to the points that you raise, with particular attention to point (3) “unpredictable performances” and, in part, to (2) caching.

The Neverlang model resembles that of simple “compiler-compilers” like Yacc, where attributes behave more like variables than functions; thus they are generally computed only once; in Neverlang attributes can also be re-computed using the `eval` construct, which descends into a child and re-evaluates the corresponding semantic action.

On the one hand, the need for an explicit `eval` make it less “convenient” than regular AG-based frameworks; on the other hand, I believe this gives better predictability, and, although the focus for the framework are not performances, but rather modularity, I think that “predictability” would better motivate the reasons for this choice.

Thanks in advance,

[1] http://link.springer.com/chapter/10.1007%2F978-3-642-39614-4_2#page-1
[2] http://dl.acm.org/citation.cfm?id=2584478
[3] http://blog.lexspoon.org/2011/04/practical-challenges-for-attribute.html

Edoardo Vacchi is PhD Student at Walter Cazzola's ADAPT-Lab, a research lab at Università degli Studi di Milano that investigates methods and techniques for programming language development and software adaptation and evolution. Walter Cazzola is associate professor at UniMi and his research is concerned with software and language engineering. More info about Neverlang can be found at the website http://neverlang.di.unimi.it.

My analysis of the Swift language

Apple has put out Swift, which sounds like a nice language overall. Here is my flagrantly non-humble opinion about how its features line up with what I consider modern, well-established aspects of programming language design.

The good

First off, Swift includes range-checked integer arithmetic! Unless you explicitly ask for wraparound, any overflow will cause an exception. I was just commenting yesterday on what a tough problem this is for current programming languages.

It has function types, nested functions, and closures, and it has numerous forms of optimized syntax for closures. This is all heartening, and I hope it will stick going forward, much the way lexical variable binding has stuck. Closures are one of those things that are both helpful and have small down side once your language has garbage collection.

Swift's closures can assign to variables in an outer scope. That's the right way to do things, and I find it painful how much Java's designers struggle with this issue. As a technical detail, I am unclear what happens if a closure captures a variable but does not modify it. What ought to happen is that any read from it will see the latest value, not the value at the time the capture happened. However, the Closures section of the Language Guide suggests that the compiler will capture just the initial value in this case. I believe this is misguided and will cause as many traps as it fixes; for example, suppose the programmer captured a counter, but does not increment that counter itself? The motto here should be: you don't know what the programmer meant, but you know what they wrote.

Type inference is quite welcome. I don't know what more to say than that developers will take advantage of it all the time, especially for local variables.

Tuple types are a small touch that comes up in practical programming all the time. How many times have you wanted to return two values from a function, and had to design a class for it or otherwise to pervert your design?

Enumerations seem good to include in the language. Language designers often seem to think that enums are already handled by other language features, and therefore should not be included. I respect that, but in this case, it's a simple feature that programmers really like to use. Java's enums are baroque, and none of the several Datalog dialects I have wokred on include enums at all. I miss having language support for a closed set of named integers. It's easy to support and will be extremely popular.

As an interesting trick, keyword arguments to functions are supported, but you have to opt in. That's probably a good combination. Keyword arguments are quite useful in cases where you have a lot of parameters, and sometimes this legitimately happens. However, it's unfortunate if you afflict all functions with keyword arguments, because the keyword arguments become part of the API. By making it opt in, the feature is there for the functions which can use it.

Including both structs and classes looks initially redundant, but it's quite helpful to have a value type that encompasses multiple other values. As an example, the boxed Integer type on Java would be much better as a struct than as a class.

Extensions look valuable for programming in the large. They allow you can make an existing class fit into a new framework, and they let you add convenience methods to an existing class. Scala uses its implicit conversions for extensions, but direct support for extensions also makes a lot of sense.

The way option chaining works is a nice improvement on Objective C. In Objective C, any access to nil returns nil. In practice, programmers are likely better off with getting an error when they access nil, as a form of design by contract: when something goes wrong, you want the program to stop at that point, not some indefinite time later. Still, sometimes you want nil propagation, and when you do, Swift lets you just put a "?" after the access.

Weak references are helpful for any language with automatic memory management, but they look especially helpful in a language with reference-counting memory management. I don't follow why there are also the "unowned" references, except that the designers didn't want your code to get polluted with ! dereferences. Even so, I would think this is a case of do or do not do. If you are worried about ! pollution, which is a legitimate concern, then simply don't require the !.

As an aside, this is the reason I am not sure pervasive null is as bad as often claimed. In practical code, there are a lot of cases where a value is sometimes optional but, in a specific context, is known to be present. In such a case, you are just going to deference it, and possibly suffer a null-pointer check if you were wrong. As such, programmers are guided into a style where they just insert dereferences until the compiler shuts up, which makes the code noisey without increasing practical reliability.

The dubious

Swift looks very practical and workable, but there are some issues I think could have been done better.

Single inheritance seems like a step backward. The linearization style of multiple inheritance has proven helpful in practice, and it eliminates the need for a separate "interface" or "protocol" feature. Perhaps designers feel like C++'s multiple inheritance went badly, and are avoiding multiple inheritance like the plague? I used to think that way, but it's been multiple decades since C++'s core design. There are better design for multiple inheritance nowadays.

Swift doesn't appear to include persistent data structures. This is the one feature I miss the most when I don't get to program in Scala, and I don't know why it isn't catching on more widely in other languages. Developers can add their own collection types, but since the new types aren't standard, you end up having to convert to standard types whenever you call into another library.

The automatic immutability of collections assigned to constants looks driven by the lack of persistent collections. It's better to support both features independently: let variables be either mutable or not, and let collections be mutable or not. All four combinations are very useful.

Deinitialization, also known as finalization, looks like a throwback to me. In a system with automatic memory management, you don't want to know precisely when your memory is going to get freed. As such, you can't count on deinitializers running soon enough to be useful. Thus, you always need a fallback plan of deallocating things manually. Once you deallocate manually, though, deinitializers become just a debugging technique. It's better to debug leaks using a tool than with a language feature.

In-out parameters seem like a step backwards. The trouble is that most functions use only in parameters, so when you see a function call, a programmer's default assumption is that the callee will not modify the argument. It can lead to bad surprises if the parameter gets modified at all. Out parameters are so unusual that it's better to be explicit about them, for example by taking a mutable collection as an argument.

Custom precedence (and associativity) is likely to go badly. We discussed this in detail, over the course of days, for X10, because X10 is a scientific language that really benefits from a rich set of operators. One problem with user-defined precedence is that it's hard to scope: you want to scope the operators themselves, not their implementations, because parsing happens before method lookup. It's also tough on programmers if they have to learn a new precedence table for every file of code they read. All in all, we concluded that Scala had a reasonable set of trade offs here: have a built-in precedence table with a huge number of available operators, and make library designers simply choose from the existing operators.

I see no exceptions, which is likely to be a nuisance to programmers if they are truly missing. Sometimes you want to tear down a whole chunk of computation without exiting the whole program. In such cases, exceptions work very well. Maybe I just missed it.

Integer types are hard to get right, and I am not sure Swift has chosen a great approach. It's best to avoid unsigned types, and instead to have untyped operations that can apply to typed integers. It's also best to avoid having low-precision operations, even if you have low-precision storage. Given all of the above, you don't really need explicit conversions any more. Java's integer design is quite good, with the exception of the unnecessary char type that is not even good for holding characters. I suspect many people overlook this about Java, because it's a case where programmers are better off with a language with fewer features.