Providence: Machine Learning At Stack Exchange

At Stack Exchange, we’ve historically been pretty loose with our data analysis.  You can see this in the “answered questions” definition (has an accepted answer or an answer with score > 0), “question quality” (measured by ad hoc heuristics based on votes, length, and character classes), “interesting tab” homepage algorithm (backed by a series of experimentally determined weights), and rather naïve question search function.

Note: not all nodes are picturedThis approach has worked for a long time (turns out your brain is a great tool for data analysis), but as our community grows and we tackle more difficult problems we’ve needed to become more sophisticated.  For example, we didn’t have to worry about matching users to questions when we only had 30 questions a day but 3,500 a day is a completely different story.  Some of our efforts to address these problems have already shipped, such as a more sophisticated homepage algorithm, while others are still ongoing, such as improvements to our search and quality scoring.

One of our other efforts is to better understand our users, which led to the Providence project.  Providence analyzes our traffic logs to predict some simple labels (like “is a web developer” or “uses the Java technology stack”) for each person who visits our site.  In its early incarnation, we only have a few labels but we’re planning to continue adding new labels in order to build new features and improve old ones.

While we can’t release the Stack Overflow traffic logs for privacy reasons, we believe it’s in the best interest of the community for us to document the ways we’re using it.  Accordingly, this is the first post in a series on the Providence project.  We’re going to cover each of the individual predictions made, as well as architecture, testing, and all the little (and not-so-little) problems we had shipping version 1.0.

We have also added a way for any user to download their current Providence prediction data because it’s theirs and they should be able to see and use it as they like.  Users can also prevent other systems (Careers, the Stack Overflow homepage, etc.) from querying their Providence data if they want to.

First up: What kind of developer are you?

One of the first questions we wanted Providence to answer was ‘What “kind” of developer are you?’.  This larger question also encompassed sub-questions:

  • What are the different “developer-kinds”?
  • How much, if at all, do people specialize in a single “kind” of development?
  • Among these different kinds of developers, do they use Stack Overflow differently?

We answered the first sub-question by looking at a lots and lots of résumés and job postings.  While there is definitely a fair amount of fuzziness in job titles, there’s a loose consensus on the sorts of developers out there.  After filtering out some labels for which we just didn’t have much data (more on that later), we came up with this list of developer-kinds:

  • Full Stack Web Developers
  • Front End Web Developers
  • Back End Web Developers
  • Android Developers
  • iOS Developers
  • Windows Phone Developers
  • Database Administrators
  • System Administrators
  • Desktop Developers
  • Math/Statistics Focused Developers
  • Graphics Developers

The second sub-question we answered by looking at typical users of Stack Overflow.  Our conclusion was that although many jobs are fairly specialized, few developers focus on a single role to the exclusion of all else.  This matched our intuition, because it’s pretty hard to avoid exposure to at least some web technologies, not to mention developers love to tinker with new things for the heck of it.

Answering the final sub-question was nothing short of a leap of faith.  We assumed that different kinds of developers viewed different sets of questions; and, as all we had to use were traffic logs, we couldn’t really test any other assumptions anyway.  Having moved forward regardless, we now know that we were correct, but at the time we were taking a gamble.

The Data

Engage!A prerequisite for any useful analysis is data, and for our developer-kind predictions we needed labeled data.  Seeing that Providence did not yet exist, this data had not been gathered.  This is a chicken and egg problem that frequently popped up during the Providence project.

Our solution was an activity we’ve taken to calling “labeling parties.”  Every developer at Stack Exchange was asked to go and categorize several randomly chosen users based on their Stack Overflow Careers profile, and we used this to build a data set.  For the developer-kinds problem, our labeling party hand classified 1,237 people.

The Classifier

In our experience, naïvely rubbing standard machine learning algorithms against our data rarely works.  The same goes for developer-kinds.  We attacked this problem in three different steps: structure, features, algorithms.

Looking over the different developer-kinds, it’s readily apparent that there’s an implicit hierarchy.  Many kinds are some flavor of “web developer,” while others are “mobile developer,” and the remainder are fairly niche; we’ve taken to calling “web,” “mobile,” and “other” major developer-kinds.  This observation led us to first classify the major developer-kind, and then proceed to the final labels.

Yes, the designer call out is by design

Since we only really have question tag view data to use in the initial version of Providence, all of our features are naturally tag focused.  The breakdowns of the groups of tags used in each classifier are:

  • Major Developer-Kinds
    • Web programming languages (java, c#, javascript, php, etc.)
    • Mobile programming languages (java, objective-c, etc.)
    • Non-web, non-mobile programming languages
    • iDevices
    • Web technologies (html, css, etc.)
  • Mobile Developer-Kinds
    • iDevice related (ios, objective-c, etc.)
    • Android related (android, listview, etc.)
    • Windows Phone related (window-phone, etc.)
  • Other Developer-Kinds
    • Each of the top 100 used tags on Stack Overflow
    • Pairs of each of the top 100 used tags on Stack Overflow
    • SQL related (sql, tsql, etc.)
    • Database related (mysql, postgressql, etc.)
    • Linux/Unix related (shell, bash, etc.)
    • Math related (matlab, numpy, etc.)

For many features, rather than use the total tag views, we calculate an average and then use the deviation from that.  With some features, we calculate this deviation for each developer-kind in the training set; for example, we calculate deviation from average web programming language tag views for each of the web, mobile, and other developer-kinds in the Major Developer Predictor.

Turning these features into final predictions requires an actual machine learning algorithm, but in my opinion, this is the least interesting bit of Providence.  For these predictors we found that support vector machines, with a variety of kernels, produce acceptably accurate predictions; however, the choice of algorithm mattered little, various flavors of neural networks performed reasonably well, and the largest gains always came from introducing new features.

So how well did this classifier perform?  Performance was determined with a split test of job listing ads with the control group being served with our existing algorithm which only considered geography, we’ll be covering our testing methodology in more depth in a future post.  In the end we saw an improvement for 10-30% over the control algorithm, with the largest gains being seen in the Mobile Developer-Kinds and the smallest in the Web Developer-Kinds.

Next up: What technologies do you know?

Jil: Version 2.x, Dynamic Serialization, And More

Jil, a fast JSON serializer/deserializer for .NET, has hit 2.x!  It’s gained a lot more perf work, some improved functionality, and dynamic serialization.

What’s Dynamic Serialization?

Jil has always supported static serialization via a generic type parameter.  This would serialize exactly the provided type; things like inheritance, System.Object members, and the DLR were unsupported.

In .NET there are, broadly speaking, three “kinds” of dynamic behavior: polymorphism, reflection, and the dynamic language runtime.  Jil now supports all three.


Polymorphism comes up when you use subclasses.  Consider the following.

class Foo
  public string Fizz;
class Bar : Foo
  public string Buzz;


Since Jil can’t know the Bar class exists when Serialize<Foo>(…) is first called, the Buzz member would never be serialized.  SerializeDynamic(…), however, knows that the Foo class isn’t sealed and that the possibility of new members needs to be accounted for on every invocation.  A similar situation exist with virtual members, which is also handled in SerializeDynamic.


Reflection matters when the objects being serialized have normal .NET types (ints, strings, lists, user defined classes , and so on) at runtime but that type isn’t known at compile time.  Consider the following.

object a = 123;
object b = "hello world";

Calling Serialize(a) or Serialize(b) would infer a type of System.Object, since at compile time that’s the only information we have.  Using SerializeDynamic, Jil knows to do a runtime lookup on the actual type of the passed in object.

Dynamic Runtime

The dynamic runtime backs the dynamic keyword in C#, but for Jil’s purposes it can be thought of as special casing types that implement IDynamicMetaObjectProvider.

While it’s rare to directly implement IDynamicMetaObjectProvider, code using ExpandoObject or DynamicObject isn’t unheard of.  For example, Dapper uses ExpandoObject to implement its dynamic returns.

Speed Tricks

As usual, Jil focuses on speed although dynamic serialization is necessarily slower than static serialization.

Jil builds a custom serializer at each point in the type graph where the type could vary on subsequent calls.  For example, if Jil sees a “MyAbstractClassBase” abstract class as a member it will do an extra lookup on each call to SerializeDynamic(…) to find out what the type really is for that invocation.  If instead Jil sees a “string” or a “MyValueType” struct as a member, it knows that they cannot vary on subsequent calls and so will not do the extra lookup.  This makes the first serialization involving a new type slower, but subsequent serializations are much faster.

The most common implementations of IDynamicMetaObjectProvider are special cased, ExpandoObject is treated as an IDictionary<string, object> and DynamicObject’s TryConvert(…) method is called directly.  This avoids some very expensive trial casts that are sometimes necessary when serializing an implementer of IDynamicMetaObjectProvider.

Further Improvements

While dynamic serialization is the main new feature in 2.x, other improvements have been made.

A partial list of improvements:

Relative Speed Improvements

It’s been a while since I’ve posted Jil benchmarks, since most recent work has been on dynamic features that aren’t really comparable.  However, lots of  little improvements have added up to some non-trivial performance gains on the static end.

Overall serialization gains fall around 1% for the Stack Exchange API types, with the larger models and collections gaining a bit more.

In the Stack Exchange API, deserialization of most models has seen speed increases north of 5% with largest models and collections seeing double-digit gains.


These numbers are available in a Google Doc were derived from Jil’s Benchmark project running on a machine with the following specs,:

  • Operating System: Windows 8 Enterprise 64-bit
  • Processor: Intel Core i7-3960X 3.30 GHz
  • Ram: 64 GB
    • DDR
    • Quad Channel
    • 800 MHz

The types used were taken from the Stack Exchange API to reflect a realistic workload, but as with all benchmarks take these numbers with a grain of salt.

I would also like to call out Paul Westcott’s (manofstick on Github) contributions to the Jil project, which have made some of these recent performance gains possible.

As always, you can

Browse the code on  GitHub or Get Jil on NuGet

Jil: Dynamic Deserialization

With version 1.3.0 Jil now supports dynamic deserialization of JSON, with the appropriately named JSON.DeserializeDynamic(…) methods.

What Types Are Supported?

Jil’s dynamic deserializer parses the same set of types that its static deserializer does.  Supported types are:

  • Strings and Chars
  • Booleans
  • Integers
  • Floating point numbers
  • DateTimes
  • Nullables
  • Enumerations
  • Guids, in the “D” format
  • Arrays
  • Objects

DateTime formats must be configured through an Options object, and includes four popular JSON date formats.

How Dynamic Is It?

Jil returns a dynamic object, introduced in C# 4, rather than some custom “JSON” class.  Using the parsed object is done with natural feeling C# instead of method or indexer calls.

The following are supported operations on a Jil dynamic:

  • Casts
    • ie. (int)JSON.DeserializeDynamic(“123″)
  • Member access
    • ie. JSON.DeserializeDynamic(@”{“”A””:123}”).A
  • Indexers
    • ie. JSON.DeserializeDynamic(@”{“”A””:123}”)[“A”]
    • or JSON.DeserializeDynamic(“[0, 1, 2]”)[0]
  • Foreach loops
    • ie. foreach(var keyValue in JSON.DeserializeDynamic(@”{“”A””:123}”)) { … }
      • in this example, keyValue is a dynamic with Key and Value properties
    • or foreach(var item in JSON.DeserializeDynamic(“[0, 1, 2]”)) { … }
      • in this example, item is a dynamic and will have values 0, 1, and 2
  • Common unary operators (+, -, and !)
  • Common binary operators (&&, ||, +, -, *, /, ==, !=, <, <=, >, and >=)
  • .Length & .Count on arrays
  • .ContainsKey(string) on objects

Jil also allows you to cast JSON arrays to IEnumerable<T>, JSON objects to IDictionary<string, T>, and either to plain IEnumerables.

Dynamic objects returned by Jil are immutable.  Setting properties, setting through an indexer, .Add, +=, and so on are not supported and will throw exceptions at runtime.

Performance Tricks

As always, Jil works quite hard to be quite fast.  While dynamic deserialization is necessarily somewhat slower than its static counterpart, it’s still quite fast.

Custom Number Parsers

.NET’s built in number parsers are much more general purpose than what Jil needs, so a custom parser that only does what is demanded by the JSON spec is used instead.

Custom Number Format

The JSON spec explicitly doesn’t specify the precision of numbers, and Jil exploits that to implement a non-IEEE794 floating point representation while parsing.  Jil uses a less compact, but easier to parse, “Fast Number” of 169-bits for most purposes internally.  Converting from this format to .NET’s built-in floating point values requires some extra work, but is still a net speed up for a small number of accesses.  In addition, converting from integer “Fast Numbers” to .NET’s built-in integer types is quite fast.

 Minimal Allocations

While dynamic dispatch implies a great many more allocations than with Jil’s static deserializer, the dynamic deserializer still strives to minimize them.  You can see this in the methods for parsing strings, where a char[] is used if possible rather than allocating strings via a StringBuilder.


Benchmarking dynamic code is tricky.  Comparisons to other libraries may not be apples-to-apples if the same dynamic behavior hasn’t been implemented, or if parsing or conversion is delayed to different points.

Rather than put together a, quite possibly misleading, comparison to other JSON libraries.  The following compares Jil’s dynamic deserializer to its static deserializer.

Comparison of Jil's static and dynamic deserializers

Comparison of Jil’s static and dynamic deserializers

The code for this benchmark can be found on Github.

While the dynamic deserializer is quite fast, it’s worth remember that dynamic dispatch is considerably slower than static dispatch and that some type conversions are deferred when deserializing dynamically.  In practice this means that using the static deserializer is still noticeably faster than using the dynamic deserializer, but the dynamic deserializer is still usable for most purposes.

Grab the latest Jil on Nuget or Checkout the code on Github

After the inevitable bug fixes, I’ll probably be revisiting Jil’s existing SerializeDynamic() method to make it actually dynamic.  Currently it only dynamically dispatches on the first type it sees, which isn’t nearly as useful.

Jil: Doing JSON Really, Really Quickly

After about three months of work, and some time in a production environment, the second half of Jil (a fast JSON library built on Sigil) is ready for release.

Jil now supports both serializing and deserializing JSON.

As with serialization, Jil deserialization supports very few configuration options and requires static typing.  The interface is, accordingly, just a simple “JSON.Deserialize<T>(…)” which takes either a string or a TextReader.

The Numbers

Jil’s entire raison d’être is ridiculous optimizations in the name of speed, so let’s get right to the benchmarks.

Let’s start with the (recently updated) SimpleSpeedTester:

Performance Graph

The numbers show Jil’s deserializers are about 50% faster than the next fastest JSON library.  Protobuf-net is included as a baseline, it isn’t a JSON serializer (it does Protocol Buffers) but it’s the fastest .NET serializer (of anything) that I’m aware of.

The Tricks

Just like with serializing, Jil does runtime code generation using Sigil to produce tight, monolithic deserialization methods.  That right there accounts for quite a lot of performance.

To speed up deserializing numbers, Jil has a large collection of methods tailored to each built-in number type.  Deserializing bytes, for example, can omit sign checks and all  but one overflow check.  This specialization avoids unnecessary allocations, and sidesteps all the overhead in .NET imposed by support for culture-specific number formatting.

When deserializing JSON objects, Jil tries to avoid allocations when reading member names.  It does this by checking to see if a hash function (Jil uses variants of the Fowler–Noll–Vo hash function) can be truncated into a jump table, ie. Jil calculates hash(memberName) % X for many Xs to see if a collision free option exists.  If there’s an option that works, Jil uses it instead of actually storing member names which saves an allocation and the cost of a lookup in a Dictionary.

A fair amount of performance is gained by being very parsimonious with memory on the heap.  Jil typically allocates a single character buffer, and may allocate a StringBuilder if certain conditions are met by the deserialized data.  Additional string allocations will be made if Jil cannot use hash functions member name lookups.

Interested in Jil?

Grab It On Nuget or Checkout The Code

Next up is probably dynamic deserialization, though naturally I’d love to make the existing functionality even faster.  Don’t hesitate to ping me with more performance tricks.

Jil: Serializing JSON Really, Really Quickly

Several months ago I more or less finished up my “just for fun” coding side project Sigil (which I’ve written about a few times before), and started looking for something new to hack on.  I settled on a JSON serializer for .NET that pulled out all the stops to be absolutely as fast as I could make it.

About three months later, I’m ready to start showing off…


a fast JSON serializer built on Sigil.

This is still a very early release (arbitrarily numbers 0.5.5, available on Nuget [but not in search results]), so there are a lot of caveats:

  1. Jil has very few configuration options
  2. Jil has very limited support for dynamic serialization
  3. Jil only serializes, no deserialization
  4. Jil only supports a subset of types
  5. Jil has seen very little production use
    • At time of writing, Jil has been pulled into the Stack Exchange API for about a week without issue

Some Numbers

Benchmarks are most the exciting thing about Jil, so let’s get right to them.  The following are from a fork of theburningmonk’s SimpleSpeedTester project.

In tabular form, the same data:

Raw numbers are available in a google document.  There are also further benchmarks I created while making Jil on Github.

The take away from these benchmarks is that Jil is about twice as fast as the next fastest .NET JSON serializer.  Protobuf-net is still faster (as you’d expect from an efficient binary protocol from Google and a library written by Marc Gravell) but Jil’s closer to it than to the next JSON serializer.

Some Tricks

I could write a whole series of how Jil shaves microseconds, and may yet do so.  I’ll briefly go over some of the highlights right now, though.

The first one’s right there in the name, Jil is built on Sigil.  That means a type’s serializer gets created in nice tight CIL, which becomes nice tight machine code, and no reflection at all occurs after the first call.

Second, Jil has a number of very specialized methods for converting data types into strings.  Rather than relying on Object.ToString() or similar, Jil has a separate dedicated methods for shoving Int32s, UInt64s, Guids, and so on into character arrays.  These specialized methods avoid extra allocations, sidestep all the culture-specific infrastructure .NET makes available, and let me do crazy things like divide exactly 14 times to print a DateTime.

As you’d expect of performance focused code in a garbage collected environment, the third thing Jil focuses on is trying not allocate anything, ever.  In practice, Jil can keep allocations to a single small charactar array except when dealing with Guids and IDictionary<TKey, TValue>s.  For Guids Jil must allocate an array for each since Guid.ToByteArray() doesn’t take a buffer, while serializing Dictionaries still allocates an enumerator.

If you’ve clicked through to Jil’s source by now, you might have noticed some MethodImpl attributes.  That’s a part of the Jil’s fourth big trick, trading a fair amount of memory for more speed.  Aggressively inlining code saves a few instructions spent branching, and even more time if instruction prefetching isn’t perfect in the face of method calls.

Last but not least, Jil avoids branches whenever possible; your CPU’s branch predictor can ruin your day.  This means everything from using jump tables, to skipping negative checks on unsigned types, to only doing JSONP specific escaping when requested, and even baking configuration options into serializers to avoid the runtime checks.  This does mean that Jil can create up to 64 different serializers for the same type, though in practice only a few different configurations are used within a single program.

Check Out The Code or Grab Jil From Nuget

I’m definitely interested in any crazy code that shaves more time off.  Also faster ways to create a string (rather than write to a TextWriter), my experiment with capacity estimation works… but not for reliably enough speedups to flip it on by default.

Your Future On Stack Overflow

I recently spent a while working on a pretty fun problem over at Stack Exchange: predicting what tags you’re going to be active answering in.

Confirmed some suspicions, learned some lessons, got about a 10% improvement on answer posting from the homepage  (which I’m choosing to interpret as better surfacing of unanswered questions).

Good times.

Why do we care?

Stack Overflow has had the curious problem of being way too popular for a while now.  So many new questions are asked, new answers posted, and old posts updated that the old “what’s active” homepage would cover maybe the last 10 minutes.  We addressed this years ago by replacing the homepage with the interesting tab, which gives everyone a customized view of stuff to answer.

The interesting algorithm (while kind of magic) has worked pretty well, but the bit where we take your top tags has always seemed a bit sub-par.  Intuitively we know that not all tags are equal in volume or scoring potential, and we also know that activity in one tag isn’t really indicative just in that tag.

What we’d really like in there is your future, what you’re going to want to answer rather than what you already have.  They’re related, certainly, but not identical.

Stated more formally: what we wanted was an algorithm that when given a user and their activity on Stack Overflow to date, predicted for each tag what percentage of their future answers would be on questions in that tag.  “Percentage” is a tad mislead since each question on Stack Overflow can have up to five tags, so the percentages don’t sum to anything meaningful.

The immediate use of such an algorithm would be improving the homepage, making the questions shown to you more tailored to your interests and expertise.  With any luck the insights in such an algorithm would let us do similar tailoring elsewhere.

To TL;DR, you can check out what my system thinks it knows about you by going to /users/tag-future/current on any of the older Stack Exchange sites.  The rest of this post is about how I built it, and what I learned doing it.

Unsurprisingly, that’s a pretty good model of how I think about my Stack Overflow participation.

What Do We Know?

A big part of any modeling process is going to be choosing what data to look at.  Cast too wide a net and your iteration time explodes, too narrow and you risk missing some easy gains.  Practicality is also a factor, as data you technically have but never intended to query en masse may lead you to build something you can’t deploy.

What I ended up using is simply the answers on a site (their text, creation dates, and so on), along with the tags the associated questions had when the answer was posted.  This data set has the advantage of being eminently available, after all Stack Exchange has literally been built for the purpose of serving answers, and public knowledge.

At various times I did try using data from the questions themselves and an answerers history of asking, but to no avail.  I’m sure there’s more data we could pull in, and probably will over time; though I intend to focus on our public data.  In part this is because it’s easier to explain and consume the public data but also because intuitively answerers are making decisions based on what they can see, so it makes sense to focus there first.

Except with about 315,000 different parts.

A Model Of A Stack Exchange

The actual process of deriving a model was throwing a lot of assumptions about how Stack Overflow (and other Stack Exchanges) work against the wall, and seeing what actually matched reality.  Painstaking, time consuming, iteration.  The resulting model does work (confirmed by split testing against the then current homepage), and backs up with data a lot of things we only knew intuitively.

Some Tags Don’t Matter

It stands to reason that a tag that only occurs once on Stack Overflow is meaningless, and twice is probably just as meaningless.  Which begs the question, when, exactly does a tag start to matter?  Turns out, before about forty uses a tag on Stack Overflow has no predictive ability; so all these tags aren’t really worth looking at in isolation.

Similarly a single answer isn’t likely to tell us much about a user, what I’d expect is a habit of answering within a tag to be significant.  How many answers before it matters?  Looks like about three.  My two answers in “windows-desktop-gadgets” say about as much about me as my astrological sign (Pisces if you’re curious).

Pictured: the state of programming Q&A before Stack Overflow.

Most People Are Average (That’s Why It’s An Average)

What’s being asked on Stack Overflow is a pretty good indicator of what’s being used in the greater programming world, so it stands to reason that a lot of people’s future answering behavior is going to look like the “average user’s” answering behavior.  In fact, I found that the best naive algorithm for predicting a user’s future was taking the site average and then overlaying their personal activity.

Surprisingly, despite the seemingly breakneck speed of change in software, looking at recent history when calculating the site average is a much worse predictor than considering all-time.  Likewise when looking at user history, even for very highly active users, recent activity is a worse predictor than all time.

One interpretation of those results, which I have no additional evidence for, is that you don’t really get worse at things over time you mostly just learn new things.  That would gel with recent observations about older developers being more skilled than younger ones.

You Transition Into A Tag

As I mentioned above, our best baseline algorithm was predicting the average tags of the site and then plugging in a user’s actual observed history.  An obvious problem with that is that posting a single answer in say “” could get us predicting 10% of your future answers will be in “” even though you’ll probably never want to touch that again.

So again I expected there to be some number of uses of a tag after which your history in it is a better predictor than “site average”.  On Stack Overflow, it takes about nine answers before you’re properly “in” the tag.  Of course there needs to be a transition between “site average” and “your average” between three and nine answers, and I found a linear one works pretty well.

We All Kind Of Look The Same

Intuitively we know there are certain “classes” of users on Stack Overflow, but exactly what those classes are is debatable. Tech stack, FOSS vs MS vs Apple vs Google?  Skill level, Junior vs Senior?  Hobbyist vs Professional?  Front-end vs Back-end vs DB?  And on and on.

Instead of trying to guess those lines in the sand, I went with a different intuition which was “users who start off similarly will end up similarly”.  So I clustered users based on some N initial answers, then use what I knew about existing users to make predictions for new users who fall into the cluster.

Turns out you can cut Stack Overflow users into about 440 groups based on about 60 initial tags (or about 30 answers equivalently) using some really naive assumptions about minimum distances in euclidean space.  Eyeballing the clusters, it’s (very approximately) Tech stack + front/back-end that divides users most cleanly.

We’d also expect chocolate in there.

One Tag Implies Another

Spend anytime answering on Stack Overflow and you’ll notice certain tags tend to go together.  Web techs are really good for this like “html” and “css” and “javascript” and “jquery”, but you see it in things like “ios” and “objective-c”.  It stands to reason that answering a few “c#” questions should raise our confidence that you’re going to answer some “linq-to-object” questions then.

Testing that assumption I find that it does, in fact, match reality.  The best approach I found was predicting activity in a tag given activity in commonly co-occurring tags (via a variation on principal component analysis) and making small up or down tweaks to the baseline prediction accordingly.  This approach depends on there being enough data for co-occurrence to be meaningful, which I found to be true for about 12,000 tags on Stack Overflow.

Trust Your Instincts

Using the Force is optional.

One pretty painful lesson I learned doing all this is: don’t put your faith in standard machine learning.  It’s very easy to get the impression online (or in survey courses) that rubbing a neural net or a decision forest against your data is guaranteed to produce improvements.  Perhaps this is true if you’ve done nothing “by hand” to attack the problem or if your problem is perfectly suited to off the shelf algorithms, but what I found over and over again is that the truthiness of my gut (and that of my co-workers) beats the one-size-fits-all solutions.  You know rather a lot about your domain, it makes sense to exploit that expertise.

However you also have to realize your instincts aren’t perfect, and be willing to have the data invalidate your gut.  As an example, I spent about a week trying to find a way to roll title words into the predictor to no avail.  TF-IDF, naive co-occurrence, some neural network approaches, and even our home grown tag suggester never quite did well enough; titles were just too noisy with the tools at my disposal.

Get to testing live as fast as you possibly can, you can’t have any real confidence in your model until it’s actually running against live data.  By necessity much evaluation has to be done offline, especially if you’ve got a whole bunch of gut checks to make, but once you think you’ve got a winner start testing.  The biggest gotcha revealed when my predictor went live is that the way I selected training data made for a really bad predictor for low activity users, effectively shifting everything to the right.  I solved this by training two separate predictors (one for low activity, and one for high).

Finally, as always solving the hard part is 90% of the work, solving the easy part is also 90% of the work.  If you’re coming at a problem indirectly like we were, looking to increase answer rates by improving tag predictions, don’t have a ton of faith in your assumptions about the ease of integration.  It turned out that simply replacing observed history with a better prediction in our homepage algorithm broke some of the magic, and it took about twenty attempts to realize gains in spite of the predictor doing what we’d intended.  The winning approach was considering how unusual a user is when compared to their peers, rather than considering them in isolation.

Again, want to see what we think you’ll be active in?  Hit /users/tag-future/current on your Stack Exchange of choice.

Making Magic: How Sigil Works

Version 3.0.0 of Sigil was just released (grab it on Nuget and check out the source on Github).  The big new feature this release is a disassembler, one which allows for some inspection of the flow of values through a .NET delegate.

But that’s not what I’m writing about.  I figure now’s as good a time as any to write up the “how” of Sigil, given that I covered the “what” and “why” in an earlier post and that recent releases have refactored Sigil’s internals into a state I’m happier with.

Bytecode Verifiers And You

In essence Sigil is a “bytecode verifier”.  If you’ve done JVM or .NET development you should be familiar with the concept, the bytecode verifiers on those platforms make sure that the class files or assemblies you load contain bytecode that can safely be executed.

The definition of “safe” is very particular, a bytecode verifier doesn’t prevent errors from occurring at execution time but rather prevents invariants of the runtime from being violated.  For example, a bytecode verifier would guarantee that invoking an instance method is passed the proper number and types of parameters and that it is invoked against an instance of the appropriate type.

One way to think about bytecode verifiers is that they guarantee that every operation receives the correct types as inputs and every operation leaves the runtime in a well formed state.

Sigil’s Special Features

Where Sigil differs from other bytecode verifiers is that it doesn’t operate on “finished” instruction sequences.  It verifies as you build a sequence, failing as soon as it can be sure the sequence is invalid.

Because Sigil deals with incomplete instruction sequences it also has to do with a lot of unknowns, especially around branches.  It’s quite common to branch to an instruction you haven’t actually emitted yet or emit instructions that aren’t yet provably reachable, both cases a traditional verifier can never encounter.

Sigil also has to explain itself when it fails, so it has to be able to deliver where and why a given sequence became invalid (which can be far removed from the last emitted instruction because of branches).  Similar complications exist when verification is successful, as things like eliding trivial casts and replacing branches with their short forms (which are deferred until an instruction stream is finalized) requires a lot of information about the instruction stream be retained.

Simple Verifying

If you ignore branches, verifying a bytecode sequence is pretty simple.  You can think of it as executing instructions as if they consumed and produced types instead of values.  Since Sigil is a .NET library I’ll be using .NET examples, though the basic idea applies to all similar verifiers.

For example, assume the following is an implementation of a Func<int>:

ldc.i4 1
ldc.i4 2

We know “ldc.i4″ consumes nothing, and produces an int, “add” actually consumes and produces a wide range of types but one of them is “int int -> int”.  The “ret” instruction either consumes nothing or a single type, dependent on the signature of the method it is used in; in this case it consume an “int” which we know because the method is a “Func<int>”.

I’ve written out the state of the stack (the .NET runtime is specified as a stack machine) after each instruction executes:

         // --empty--
ldc.i4 1 // int
ldc.i4 2 // int int
add      // int
ret      // --empty--

We need to add one more rule, which is that control flow ends with a “ret” which leaves the stack empty, and with that it’s clear that this sequence would verify.

How each instruction affects the stack, in terms of types.

The following sequence wouldn’t verify:

ldc.i4 1            // int
ldstr "hello world" // string int
mult                // !ERROR!

Since “mult” doesn’t have a transition from “string int” verification could be shown to fail as soon as “mult” is emitted.

The earliest releases of Sigil were simple verifiers as they didn’t try and trace through.  Through version 1.2.9 you instead had to assert types when marking labels under certain circumstances.

Complex Verifying

As soon as you add branches life gets considerably more complicated.  Now you have to deal with cases where you can’t infer what types are on the stack immediately.

Consider the following:

ldc.i4 1
br end_label


ldc.i4 2
br middle_label

In this example the types passed to “add” are unclear when it is emitted, it’s only when “br middle_label” is encountered that it becomes clear that “add” will verify.

Furthermore, whether or not “ret” verifies depends on the method being built. It would verify if we’re building a Func<int> then it verifies.  If we’re building a Func<string> it should fail when emitted, since there’s no way for “add” to pass a string to “ret”.  If we’re building a Func<double>, then it should fail when “br middle_label” is emitted since then “add” is known to produce and pass an Int32 to “ret”.

How Sigil Copes

Sigil deals with the complexities of verifying partial sequences in two ways: tracking the possible types on the stack, and creating a verifier for each possible code path.

Reconsidering the above example:

                // ** Initial Verifier: V1, with empty stack **
ldc.i4 1        // int
br end_label    // int
                // ** New Verifier: V2 with unknown stack, V1 stashed **
add             // int|long|float|double|etc.
ret             // --empty--
                // ** New Verifier: V3 with unknown stack, V2 stashed **
end_label:      // ** V1 restored (stack is: int) **
ldc.i4 2        // int int
br middle_label // int int (V2 now verified with int int as initial stack)

There’s rather a lot going on now, what with verifiers being stashed and restored.  Also note that when “add” is first encountered it places an “or” clause onto the stack, which allows “ret” to verify if it expects any type in that “or” clause.

How different verifiers see the instruction stream

The logic around creating and restoring verifiers is tricky, but boiled down:

  • At an unconditional branch, store the current stack state and remove the current verifiers
    • If the branched to label has already been marked, take the current stack and check it against the expected initial stack at that label
  • At a conditional branch do the same checks and stores as an unconditional one, but don’t remove the current verifiers
  • When marking a label, if there is no verifier create a new one with an unknown initial stack
    • If there are stacks stored against that label from previous branches, check that those stacks are consistent and make them the new verifiers initial stack
  • Treat “ret” as an unconditional branch which doesn’t store the stack state
  • If at any point an instruction is emitted and there is no current verifier, the newly emitted code is unreachable

This logic is captured in Sigil’s RollingVerifier UnconditionalBranch, ConditionalBranch, Return, and Mark methods.

As for tracking all possible values on the stack (as in the “or” with “add” in the previous example), Sigil considers transitions rather than types on the stack.  So long as there’s at least one transition that could be taken for every instruction a verifier has seen, the instruction stream is considered verifiable.

Take for example the, rather complex, transitions for “add”:

    new StackTransition(new [] { typeof(int), typeof(int) }, new [] { typeof(int) }),
    new StackTransition(new [] { typeof(int), typeof(NativeIntType) }, new [] { typeof(NativeIntType) }),
    new StackTransition(new [] { typeof(long), typeof(long) }, new [] { typeof(long) }),
    new StackTransition(new [] { typeof(NativeIntType), typeof(int) }, new [] { typeof(NativeIntType) }),
    new StackTransition(new [] { typeof(NativeIntType), typeof(NativeIntType) }, new [] { typeof(NativeIntType) }),
    new StackTransition(new [] { typeof(float), typeof(float) }, new [] { typeof(float) }),
    new StackTransition(new [] { typeof(double), typeof(double) }, new [] { typeof(double) }),
    new StackTransition(new [] { typeof(AnyPointerType), typeof(int) }, new [] { typeof(SamePointerType) }),
    new StackTransition(new [] { typeof(AnyPointerType), typeof(NativeIntType) }, new [] { typeof(SamePointerType) }),
    new StackTransition(new [] { typeof(AnyByRefType), typeof(int) }, new [] { typeof(SameByRefType) }),
    new StackTransition(new [] { typeof(AnyByRefType), typeof(NativeIntType) }, new [] { typeof(SameByRefType) })

One more thing to track when dealing with verifiers is whether or not we know their initial stack, what I call “baseless” in the code.  It is not an error for an instruction stream to underflow a baseless verifier, since it’s stack could be anything.  Instead of failing verification, Sigil considers the result of underflowing a baseless stack to be a “wildcard” type which satisfies any transition; this is how “add” can pop a value to continue verification after “middle_label”.

Trickier Things

.NET has a couple hairy CIL opcodes that require special handling: dup, localloc, and leave.

“dup” duplicates the current value on the stack, the difficulty being that we only know the type on the stack if we can verify the preceding instructions which isn’t always.  Sigil handles this by making “dup” place a special type on the stack, which when encountered by the verifier pushes a copy the preceeding transition’s result or a wildcard if underflowing a baseless verifier.

“localloc” is analogous to alloca(), pushing a pointer to memory on the stack, which requires that only a single value be on the stack when executed.  This means the current verifier cannot be baseless to verify.  In this case Sigil uses a special transition which asserts that the size of the stack is one if the verifier is based, and is ignored if it is not.

“leave” is an unconditional branch out of an exception or catch block which empties the stack entirely.  Sigil considers this equivalent to “pop”-ing exactly the number of items currently on the stack, which like “localloc” means the current verifier cannot be baseless.  Like “dup” Sigil uses a special type to indicate that the stack needs to be emptied to the verifier.

Optimizing Emitted Code

There are two kinds of optimizations Sigil can do to emitted CIL: eliding trivial casts, and replacing instructions with their short forms.

Conceptually eliding is straight forward, just keep track of what types are guaranteed to go into a castclass or isinst operation and if those types are assignable to the type encoding with the instruction elide it.  Sigil attaches callbacks, like this one, to “castclass” and “isinst” transitions which are called whenever a verifier processes those operations and passed enough information to decide whether to elide themselves or not.

Some short forms are easy, any of the LoadConstant methods with short forms can be changed at call time.  The trickier ones are branches, as we need to wait until we’re done emitting code and can calculate offsets.  Tracking offsets is handled by BufferedILGenerator (which maintains a cache of byte offsets to each instruction) and a last minute call to Emit.PatchBranches patches all the branch instructions that can fit their offsets into a signed byte.

Optimizing Sigil

Sigil employs some tricks to keep it relatively snappy.

Perhaps most importantly, it doesn’t re-verify operations unless it absolutely has to.  Every currently in scope verifier maintains a cache of its state when it last verified the instruction stream and re-uses that cache when new instructions are added.  The cache does have to be invalidated when the initial stack state changes, which only happens when branching to or marking labels.

Sigil also tries to keep the number of verifiers in scope limited by discarding any non-baseless verifiers past the first one.  Since any verifier that isn’t baseless can be traced back to the start of the method, we know that there are no “or” type clauses on the stack so the verifiers are equivalent going forward even if they took different paths through the code.

Other Tidbits

To wrap this up I’m just going to list some minor things of note found in Sigil.

  • Sigil has most of an implementation of Linq-to-Objects to run on pre-.NET 3.5 runtimes, heavily influenced by Jon Skeet’s Edulinq series
  • Sigil has it’s own Tuple implementation for similar reasons
  • Sigil’s disassembler, while doing a great deal more, started as a replacement for the Mono.Reflection disassembler in our code base
  • Sigil’s exception blocks are slightly different from ILGenerators in that you explicitly attach catches and finallys to them, this makes nested exception handling easier to debug

And again, Sigil can be installed from Nuget and the source is available on Github.


Get every new post delivered to your Inbox.