Scala Exchange 2015 is on! I will be updating this entry with summaries on the talks I attended. Typos a plenty, be warned, but I hope this proves useful to somebody while we wait for the videos to be made available. You can also read my summary of day one.
Keynote: Spark+Hadoop and how it relates to Scala
By Sean Owen
Intends to give a different perspective on Scala, a less technical talk, more about his path into Scala coming from Hadoop and Java.
Sean started with EJB and Struts. Some of these ideas faded after the dot-com bust. Moved to Google, built Barcode scanner in Android. He chose to move towards Java to avoid C++. After Google got involved with open-source in a project that ended up being Apache Mahout. There were ML systems before Hadoop, but Hadoop attracted all those Java developers interested in the area. He worked with it until 2 years ago where he moved to Spark, and Scala.
The common thread on most of this path is the JVM. For enterprise people, everything looks like Spring/Maven/etc, the standard stack. Your world view is defined by it, you mold problems to fit that.
Why Hadoop in Java, instead of C++ which was used in Google for Map-Reduce? Officially due to the usual reasons: type-safe, garbage colector, easier to debug, libraries available, and platform independence. But Sean thinks there are other reasons: lots of people fled the dot-com burst to the safety of tech giants. Enterprise architecture meets web scale and Hadoop is born. A colleague thinks the real reason is Lucene used Java as an experiment, and Hadoop inherited that decision along the way.
Hadoop has some functional ideas from the start. Map-reduce, HDFS (immutable file system), no side effects in mappers/reducers. They probably were not taken directly from FP, they are just good ideas and assumptions that work well in a distributed environment, almost essential.
However, Hadoop 'feels' enterprise-y. For example, a Mapper is a heavy-weight object instead of a simple function. It allows you to store state, exposes threading to you, makes you manage serialisation, etc. There may be reasons behind (performance, etc) but it's not very functional.
In a parallel world, Python grew in the 'big-data' environment. Python is different: has an interpreter, sometype inference, has lambdas, libraries for ML, and pip to manage dependencies. Not at all like enterprise software. There are some nice things in the Python environment we wish we could have in the Java world (at that time). But if you are used to Java's world you won't realise you may want these things.
Then we arrive to 2014, where functional invades the Hadoop ecosystem via Spark, Kafka, and Java 8 lambdas. Mentions to Storm and other similars projects that reinforce the trend. So suddenly the JVM has access to some of the features it was lacking.
But this is not about Spark replacing Map-reduce unexpectedly, we already had projects (like Crunch) that had improved in that area so the progression is 'natural', it has been coming for a while. This is just a next step in the evolution.
Why do Scala/Spark fit Hadoop? Many reasons:
- Immutable data at the foundation, perfect match between FP and Hadoop
- Functional paradigm from the start
- Unlocks caching due to immutability, at all levels from Spark to HDFS
- Scala naturally from Java, which in turn replaced C++. Natural progression
- Hybrid imperative/functional wins, easier to adapt to and quickly grab benefits
- Memory caching allows for fast iteration, which benefits data scientists as iteration is at the core of ML
- The provided Shell allows for exploratory analytics, makes it easier to prototype
- It's more familiar to Python devs and the environment/features they are used to
What do I like about Scala as a former Java developer?
- Love collections, say what you will but they give you a lot of expressiveness
- Love case classes and tuples, save you a lot of boilerplate
- Love language constructs like val, lazy and match. They simplify and clarify what you intend to write
- Love closures, which are much better than Java 8 lambdas, cleaner and slicker
What do I dislike about Scala as a former Java developer?
- Dislike Option, not sure it's way better than null
- Dislikes incompatible minor releases, has been very painful in Spark. A big deal for projects
- Dislikes Sbt imperative style, still prefers Maven declarative style. Writing code in your build script is a step backwards when you try to understand somebody else Sbt build script.
- Dislike the fact scalac is incredibly heavy. Yes, he understands the Scala compiler does so much more for you, but compiling park takes so long it seems crazy on 2015.
The last complain about Scala community it's the excessive fascination with syntax, sometimes using complex functional chain of operations to express simple ideas which could be expressed in a simpler way in Java.
Scala is, no doubt, the future for Hadoop ecosystem. That is good.
Streams: reactive? functional? Or: akka- & scalaz- streams side-by-side
By Adam Warski
The aim of today is to see the differences between Akka and Scalaz streams. Most of the talk will be live coding, so please watch the video.
The problem we want to solve is to process data as it comes. We can't hold all the data into memory for some reason, irrelevant why. We focus on scenarios with a single node which cover a lot of common scenarios, for multi-node computations please check Spark.
Akka Streams introduces itself as a library to process and transfer a sequence of elements in a bounded buffer space. Scalaz Streams is a streaming I/O library focussed on compositionally, expressiveness and resource safety.
In both cases we want to define a pipeline, in a type-safe way, and then run data through it. Scalaz-stream is slightly more typesafe than Akka Streams. In both libraries the first step is to create a 'blue-print' (graph for Akka, process for Scalaz) that defines how to transform the data. After that is defined we can execute it (materialise in Akka, run in Scalaz).
The basic building blocks for Akka is a Graph (simple caseL linear pipeline), that has a source (produces elements), Sink (consumes an input) and some Flow pieces that transforms elements. At runtime each component is materialised into an Actor, and each actor does the operation defined in the relevant component. Each component also materialise into an additional value (example: Source into a Future you can use to get the result of the processing).
In Scalaz the approach is different. We initially create a data type
T is the output type,
F[_] describes side-effects that can occur while processing. In simplest case we have
Process[Nothing, T] that emits
T without any change. We also have aliases like
Sink[F[_],0] = Process[F, O => F[Unit]]. The Process is in fact akin to a state machine.
Akka implements the reactive-streams standard, which provides back-pressure via a back-channel, via dynamic push-pull. Source only produces data when the back-channel indicates it is needed, to avoid overwhelming components downstream. Everything is actor based, things happen concurrently.
Scalaz has back-pressure for free as it is entirely pull-based. Elements are evaluated on by one, in a functional approach by which we only process one element once the state machine says we can do so. Not dependent on Scalaz except for
Task. To be clear, Scalaz is slower than Akka version (2-3x).
(Code examples via live-coding start, please watch video. I'll try to summarise any key ideas mentioned)
In Akka operations over the Source (like
filter) are converted into actors behind the scenes. The first example to manipulate a text file (read, filter, map and save) already creates at least 8 actors. You can't influence the concurrency.
Code in Scalaz version looks similar, just using different types as Source. Scalaz calls
run twice, first once compiles process into a
Task, the second runs the
Task. If you want parallelism you need to be explicit about it, there is none by default.
In the first test Akka took 3.84s while Scalaz took 6.93s. Take in account Scalaz was using a single thread (no parallelism defined).
The second example wants to transform Int to Int, processing odd and even elements in parallel. For Akka this means creating a FlowGrap that defines the data flow by first creating components of the Graph and then declaring the connections. Beware the graph correctness is checked at runtime, not at compile time! Adam comments that creating a component,
SplitStage, was complex and more so due to the use of a mutable API.
Scalaz version of the Int processing, all purely functional. Adam confesses it was hard to write the first time. Scalaz requires us to be explicit on parallelism, so we need to create bounded queues (for memory protection) and we use streams to add/get elements to/from them. We also need to close the queues once we are done. In the code we define several components, please watch the video to see the full structure (hard to describe without showing the code). All the code is purely functional with full control on side-effects. All the connections are checked at compile time.
In the second test Akka took 3.43s and Scalaz 3.25s, as both were waiting for 1s during processing.
Fast automatic type class derivation with shapeless
Modelisation: in Scala you use a lot of case classes and sealed traits to represent your domain. Hundreds or more on big projects. The usual tasks include conversion to/from JSON, persist in binary format, render as CSV, pretty-print, etc. You do this a lot, so you want to automate these tasks.
You could use macros, but the Scala type system is quite complex and that means you need to consider a lot of things. Also, macros have some surprises within the API (side effects). All in all, using macros to automate these tasks becomes unwieldy.
A solution is to use Shapeless instead.
A first example is a
Printer class that takes a value of type
T and prints it. To explain how to do it, first let's talk about
HLists are a sequence of types/values, similar to Scala
List except its elements may have different types inside the same list (
String :: Int :: HNil as an example). We can identify a
HList to a case class of the same types.
As we can build
HLists inductively, by prepending elements (
::), we can define our
Printer class by recursively iterating over the
HList and printing the values. We use support
Printer for basic types (
Printer[Int], etc) and as we can decompose the
HList into these basic components, we know we will print as expected.
As we mentioned, we can map case classes to
HList. Given that, defining
Printer in terms of
HList (which is easy) provides us a way to print any existing case class in our code, including any case class we create in the future.
Sometimes there are issues due to wrong divergences or recursive types. We can make our derivation more robust by using Shapeless'
Lazy type, which helps the compiler resolve the issues mentioned before and handles real recursion in types properly.
A caveat to consider is that implicit priorities may cause issues in automatic generation. There are workarounds, like using
export-hook from Miles Sabin, or break implicits into an object hierarchy.
There are ways to speed up the process, for example using upickle, a project that uses macros to help with type generation.
Keynote: Typelevel - the benefits of collaboration
By Miles Sabin
This is a different talk from the ones Miles usually gives, where he codes live. This one has slides! And no talking about Shapeless either.
The title was chosen months ago. This is about Typelevel and how collaboration is good. But, following Jessica's talk, I feel the subtitle should be 'The benefits of community'.
A year ago, Lars gave a keynote in this very stage. I watched that video and was astonished to how much has happened in the last year. Some very good, some not quite so great.
Let's travel to September 2014, where Typelevel announced a fork of the Scala compiler. We wanted to push some prevalent Typelevel ideas into it, to fix some particular things. It's something that can cause a lot of friction and disruption, but luckily the reaction was mostly positive. This shows that there is a big vibrant OS community around Scala. Shows that Typesafe is happy to let people help. It's better to have people collaborate instead of moaning. EPFL also showed a generally positive feeling about the possible outcomes of experimentation and prototyping in Scala compiler.
A few months later, Scala Exchange 2014, Lars did his keynote. There was a friendly panel and everybody seemed going along very well.
After that, something came to the spotlight: Typelevel is not only about technology and functional programming, it also wants to promote a good environment for minorities in the community. This caused Typelevel to adopt a code of conduct, seen as a means to an end: encourage good behaviour and deal with bad behaviour. Although most of the people accepted the code of conduct, this lead to a certain amount of drama in a part of the Scalaz community (where Lars was a primary maintainer). This had some cascading effects and shaped Typelevel since them up to now.
One of the first things that became clear is that Scalaz couldn't continue as part of Typelevel. Lars stepped down as a Scalaz maintainer. But there was the need for the abstractions Scalaz provided for Typelevel libraries. As a consequence, Cats was created, to cover the technical needs while agreeing to the code of conduct, thus following Typelevel aims.
As a result of the drama, we realised it wasn't clear to everybody else what is Typelevel about. So we needed to define Typelevel better to help people understand us besides our technical projects like Cats.
So what is Typelevel? (Mind, the following is work-in-progress and Typelevel wants feedback on it). Typelevel is a community of projects and individuals organised around:
- pure, typeful, functional programming in Scala (not in Haskell!)
- independent free and open-source software, not linked to any organisation or company. Scala owes a lot to the community, don't forget it.
- a desire to share ideas and code
- accessible and idiomatic learning resources
- inclusive, welcoming and safe environment
This is evolving, but what is we think Typelevel is about. Please watch Erik Osheim's Scala World talk to see ways you can build open source communities. Very recommended.
Typelevel has always been a collection of projects. Since one year ago we had some changes: Scalaz, Scalaz-streams and Argonaut left, other libraries like Circe, Cats, Dogs, and a few more joined. There's been new tools, tests and laws like Ensime and tut (among others). There have also appeared several integration projects, like shapeless-argonaut, which enhance existing libraries; as well as macro/plugin helpers which facilitate the job of library maintainers. A lot of stuff in one year :)
Typelevel Scala has not moved very much due to all the effort being put into Cats and related projects. But there have been tools developed that proof some concepts we wanted in the forked compiler, so some progress has been achieved. See George Leontiev's Scala World talk for more details. There's been some specific compiler work, minor, but it proves we can work with it.
Some details on some of the libraries:
- well stablished libraries like Scodec (binary data, very powerful and performant), Shapeless (generic programming) and Spire (numeric library for high precision operations). All widely used, directly or indirectly.
- also well stablished are Discipline (law checking for type classes), ScalaCheck (automated property-based testing) and Spec2 (software specification)
- we have integrations, generic extensions of other libraries. For example, argonaut-shapeless.
- we have news things: Algebra, Cats, and Alleycats. Cats ia a Scalaz replacement. Algebra provides a set of algebraic type classes, comprising some fundamental type classes which are then used in Cats. Alleycats are lawless type classes and instances, which are also used in Cats.
There's been a lot of activity lately. For example, since January 2015 there have been 1558 commits to Cats by 57 contributors.
Our guiding principles for Algebra, Cats, and Alleycats:
- approachability: we want people to be able to approach the library much more easily than it was with Scalaz. For example, preference for text-based operators, make sure terminology is expressive enough instead of just focussing on the mathematical origin of the concept.
- modularity: we broke Cats into submodules so it can be used as required instead of being a massive jar
- documentation: we strive to have a massively improved documentation so we can convey useful information in a simple and centralised way
- efficiency: we are willing to compromise and use (under the hood) non-functional constructs to improve performance. Watch Stew O'Connors talk at Scala by the Bay.
Let's talk about Tut. It's a markdown processor which interprets Scala in code blocks. This provides type-checked documentation, which helps us having useful and relevant documentation. It's been even used by Underscore to produce a book, Essential Slick!
Circe is a JSON library for Scala and Scala.js. Derived from Argonaut, uses jawn for parsing. Uses Monocle for lenses, works with Refined. Uses Shapeless for codec derivation.
(Miles is running out of time so it skips some stuff, like explaining other Typelevel libraries)
Simulacrum: provides annotations to automatically generate boilerplate when creating type classes. Machinist is a macro that eliminates overhead associated with Simulacrum.
The experience with all these projects has influenced the way we think about the forked compiler. We have a better picture of the trade offs, and we can tackle small issues in helpers and bigger pieces in the compiler itself.
We want people to get involved. We hope we have created an active and lively community. But we want you in. There is an up-to-date list of projects in our homepage. Most projects exist in Github and Gitter.
News! There will be two Typelevel summits next year! Please join us!
Workshop: Shapeless for Mortals
By Sam Halliday
It is a workshop so there will be no notes on it, sorry!
Don't forget to join us at ScalaXHack tomorrow. Hope you enjoyed the content :)