Showing posts with label scala. Show all posts
Showing posts with label scala. Show all posts
Wednesday, February 4, 2015
Play Scala and Iteratees vs Node js JavaScript and Socket io
On the JavaScript side of the world, node.js and socket.io take an imperative programming approach for stream processing by using the EventEmitter API. Ive found this imperative approach easy to learn, so for this blog post, Ill show a side by side comparison of the imperative and functional techniques for each example. The focus of the post will be on WebSockets, including a basic "hello world" example, an echo server, and a simple chat application.
Note: Ill only implement server-side code in this post. You can test your server without writing any client-side code by using the WebSocket echo client on websocket.org and giving it a URL like ws://localhost:9000/some-path.
Producer, consumer, adapter
Terms like Iteratee, Enumerator, and Enumeratee can be confusing, so lets start by trying to define them using normal human terms:
- Enumerator: every time you hear Enumerator, think "producer". This is the thing that pumps out chunks of data.
- Iteratee: every time you hear Iteratee, think "consumer". They are the inverse of the Java Iterator: in an Iterator, you call "next" to request a chunk of data; in an Iteratee, you define a "fold" method that reacts to each chunk of data (inversion of control).
- Enumeratee: every time you hear Enumeratee, think "adapter". Enumeratees can be attached in front of Iteratees to modify the data streaming through them.
Lets start with a simple "hello world" WebSocket example. The first step is accepting a WebSocket connection.
In socket.io, you do this by listening for a connection event with io.sockets.on:
In Play, you create a WebSocket Action and return an Enumerator that is used to send messages to the client and Iteratee that is used to react to messages from the client. For this first version, well just use an Enumerator and Iteratee that do nothing.
Youll need to expose this WebSocket endpoint in your routes file:
You can now use the URL ws://localhost:9000/connect in the WebSocket echo client on websocket.org. You should be able to connect, though you wont see anything interesting quite yet.
Next, lets see how we could send a message to the client. In socket.io, the socket object has an emit method:
To send a message to a client in Play, well replace the empty Enumerator from the first example with an Enumerator that contains our message. Remember, the Enumerator is the "producer", so its responsible for pumping down any data we want to send to the client.
The Enumerator companion object has handy methods to create an Enumerator from a File, InputStream, OutputStream, a Future, and, as well use in this case, an apply method to create an Enumerator from a fixed set of data:
Finally, lets log any messages we get from the client. In socket.io, we can do this by listening for messages with the socket.on method:
In Play, the Iteratee we return in the WebSocket Action is the "consumer", so its responsible for reacting to any messages from the client. You can implement an Iteratee from scratch, but the Iteratee companion object has many methods that simplify the common cases, including "consume" (which consumes and concatenates all the data), "head" (consume just the first chunk of data), and "foreach", which fires a provided callback for each chunk of data. In this case, well use "foreach" and pass it a function that logs each message:
If you reconnect to your WebSocket using the websocket.org test tool, you should now see the message "Hello!". If you send some messages from the webpage, they should show up in your Play console.
Not bad, right? The main take aways are:
- Use the WebSocket.using method to define an Action that can accept WebSocket connections
- Return an Iteratee that knows how to consume messages from the client
- Return an Enumerator that knows how to produce messages for the client
An echo server with some filtering
Lets go a tiny bit further and create an echo server: that is, a server that echoes back any messages it gets from the client.
This is easy to do in socket.io by combining socket.on and socket.emit:
Creating an Enumerator and Iteratee that are "joined" is strangely complicated to do from scratch in Play. Fortunately, Play has an object called Concurrent that contains several methods to create Enumerators and Iteratees that are connected in various ways. For example, the Concurrent.joined function gives us an echo WebSocket in one line:
Try it out in the browser by reconnecting from websocket.org. Any message you send should be echoed back and show up in the log on the webpage.
Lets make this example a little more interesting and filter the data we echo back. First, lets just modify each message by adding some text to it before echoing it back.
The imperative approach used in socket.io means you have to modify the actual listener function (the callback for socket.on) to accomplish this:
In Play, we can accomplish the same thing without having to modify the internals of the Concurrent.joined code. Since we have an actual value (the Iteratee) instead of just some side-effect function (the callback passed to socket.on), we can use functional composition to modify the Iteratee before returning it. To do that, well create an Enumeratee, which, as you may remember from above, is an adapter: well attach it in front of our Iteratee to modify the data before the Iteratee consumes it.
The Enumeratee companion object has several convenience methods for creating Enumeratees: well first use Enumeratee.map, which lets you transform each chunk of data:
The advantage of composition becomes even more apparent if we want to do some fancier filtering. For example, lets only echo back messages that arent empty and contain non-whitespace characters.
With socket.io, we have to make an even larger modification to the contents of the socket.on callback function. This is easy to do in single, simple example, but this would get more complicated across many examples in a large application.
In Play, we can simply add another Enumeratee to validate the content and compose it with the previous one. Well create this Enumeratee using the "collect" method, which only allows data to pass through it if it matches one of the cases for the PartialFunction you pass in:
Youll often run across Enumeratee code that uses operators instead of the function names. The ><> operator (fish?) composes two Enumeratees and the &> applies an Enumeratee to an Iteratee:
The operators do allow you to drop lots of parens in your code, but personally, I find the operators hard to remember and impossible to google. Therefore, I believe that in this case, they lead to code that is optimized for writing instead of reading, which is a bad trade off. I recommend sticking with the full function names.
Main take aways from this section:
A chat server
As a final example, lets put together a simple WebSocket chat server that just handles a few basic tasks:
And here is the Play version:
There are two new things to learn from this code:
If you open the websocket.org echo page in two separate tabs and start sending messages, you should see the messages showing up in the log in both tabs, prefixed with "[user1]: " or "[user2]: ". You should also see messages when a client connects or disconnects.
Final thoughts
Im still learning my way around Iteratees, so I welcome any corrections, suggestions, or other feedback. If youve come across other good tutorials and examples, please share them!
I must admit that I have some lingering doubts about Iteratees as the primary means of working with streams of data in Play. They seem to be a powerful abstraction - interleaving and modifying streams on the fly is especially cool - but the degree of complexity is very high; the sparse documentation and large vocabulary of strange method names (unfoldM, fromCallback1, joinConcatI, ><>) only make it worse.
Just about anyone can get rolling with socket.io in a matter of minutes, whereas most people take several tries across several days to grok Iteratees. Functional programming often offers a high cost to high reward trade-off, but the ratio with Iteratees is extreme. Moreover, it seems like you end up with an imperative approach for many use cases anyway, such as the Channel object returned by Concurrent.broadcast.
Id also argue that the vocabulary of Iteratees does not clearly convey the intent of the code. Look at the two chat server examples above: which one reads more clearly? In my opinion, socket.on and socket.emit make it crystal clear what the code is trying to do. On the other hand, Iteratee.foreach and Iteratee.map do not. The latter feels more like Im trying to make the tools do what I want, whereas in the former, Im clearly dealing with the actual problem space.
Iteratees are a fairly general tool, so perhaps they should be reserved for low-level, advanced functionality, while some simpler, use-case specific abstractions are available for most common use cases? For example, maybe we just need a good socket.io equivalent for Play built on top of Iteratees? Ive seen socket.io.play and atmosphere-play, but Im not sure if they are stable, maintained, tested, documented, or high quality. If anyone has experience with these libraries, let me know!
This is easy to do in socket.io by combining socket.on and socket.emit:
Creating an Enumerator and Iteratee that are "joined" is strangely complicated to do from scratch in Play. Fortunately, Play has an object called Concurrent that contains several methods to create Enumerators and Iteratees that are connected in various ways. For example, the Concurrent.joined function gives us an echo WebSocket in one line:
Try it out in the browser by reconnecting from websocket.org. Any message you send should be echoed back and show up in the log on the webpage.
Lets make this example a little more interesting and filter the data we echo back. First, lets just modify each message by adding some text to it before echoing it back.
The imperative approach used in socket.io means you have to modify the actual listener function (the callback for socket.on) to accomplish this:
In Play, we can accomplish the same thing without having to modify the internals of the Concurrent.joined code. Since we have an actual value (the Iteratee) instead of just some side-effect function (the callback passed to socket.on), we can use functional composition to modify the Iteratee before returning it. To do that, well create an Enumeratee, which, as you may remember from above, is an adapter: well attach it in front of our Iteratee to modify the data before the Iteratee consumes it.
The Enumeratee companion object has several convenience methods for creating Enumeratees: well first use Enumeratee.map, which lets you transform each chunk of data:
The advantage of composition becomes even more apparent if we want to do some fancier filtering. For example, lets only echo back messages that arent empty and contain non-whitespace characters.
With socket.io, we have to make an even larger modification to the contents of the socket.on callback function. This is easy to do in single, simple example, but this would get more complicated across many examples in a large application.
In Play, we can simply add another Enumeratee to validate the content and compose it with the previous one. Well create this Enumeratee using the "collect" method, which only allows data to pass through it if it matches one of the cases for the PartialFunction you pass in:
Youll often run across Enumeratee code that uses operators instead of the function names. The ><> operator (fish?) composes two Enumeratees and the &> applies an Enumeratee to an Iteratee:
The operators do allow you to drop lots of parens in your code, but personally, I find the operators hard to remember and impossible to google. Therefore, I believe that in this case, they lead to code that is optimized for writing instead of reading, which is a bad trade off. I recommend sticking with the full function names.
Main take aways from this section:
- Use the methods in the Concurrent object to create Iteratees and Enumerators that are connected. Well see another example of this in the next section.
- Create and compose Enumeratees when you need to modify the stream of data going into your Iteratees.
A chat server
As a final example, lets put together a simple WebSocket chat server that just handles a few basic tasks:
- Send a message to all clients when a user connects or disconnects
- Each time a message is received from a client, send it out to all other clients
And here is the Play version:
There are two new things to learn from this code:
- We are using Concurrent.broadcast to create an Enumerator and Channel that can be shared across all WebSocket connections. Channel is a new class specifically built to push messages to multiple Iteratees (ie, multiple clients). Interestingly, it is an imperative abstraction, where you just call a "push" method for its side-effect.
- The "map" method on the Iteratee will fire when the Iteratee is done; for WebSockets, this happens when the client disconnects, an event we capture and broadcast to all clients.
If you open the websocket.org echo page in two separate tabs and start sending messages, you should see the messages showing up in the log in both tabs, prefixed with "[user1]: " or "[user2]: ". You should also see messages when a client connects or disconnects.
Final thoughts
Im still learning my way around Iteratees, so I welcome any corrections, suggestions, or other feedback. If youve come across other good tutorials and examples, please share them!
I must admit that I have some lingering doubts about Iteratees as the primary means of working with streams of data in Play. They seem to be a powerful abstraction - interleaving and modifying streams on the fly is especially cool - but the degree of complexity is very high; the sparse documentation and large vocabulary of strange method names (unfoldM, fromCallback1, joinConcatI, ><>) only make it worse.
Just about anyone can get rolling with socket.io in a matter of minutes, whereas most people take several tries across several days to grok Iteratees. Functional programming often offers a high cost to high reward trade-off, but the ratio with Iteratees is extreme. Moreover, it seems like you end up with an imperative approach for many use cases anyway, such as the Channel object returned by Concurrent.broadcast.
Id also argue that the vocabulary of Iteratees does not clearly convey the intent of the code. Look at the two chat server examples above: which one reads more clearly? In my opinion, socket.on and socket.emit make it crystal clear what the code is trying to do. On the other hand, Iteratee.foreach and Iteratee.map do not. The latter feels more like Im trying to make the tools do what I want, whereas in the former, Im clearly dealing with the actual problem space.
Iteratees are a fairly general tool, so perhaps they should be reserved for low-level, advanced functionality, while some simpler, use-case specific abstractions are available for most common use cases? For example, maybe we just need a good socket.io equivalent for Play built on top of Iteratees? Ive seen socket.io.play and atmosphere-play, but Im not sure if they are stable, maintained, tested, documented, or high quality. If anyone has experience with these libraries, let me know!
Tuesday, January 27, 2015
Seven Languages in Seven Weeks Scala Day 3
After some functional programming on day two, its time for the third and final day of Scala in Seven Languages in Seven Weeks.
Scala, Day 3: Thoughts
After two lengthy chapters on the object oriented and functional programming syntax/options in Scala, the third day rushes through some of the most intriguing features, including pattern matching and concurrency via actors. I would have preferred to spend a bit more time on these complicated topics.
In fact, I had the same complaint on Day 3 of IO, where we also blasted through a discussion of concurrency in just a few pages. I respect the difficulties of plowing through seven different languages in a single book and dont expect deep discussions of any one of them, but I think the book wouldve been stronger if it had a greater bias towards the more advanced "day 3" topics of each language instead of basic syntax discussions in day 1 or 2.
Scala, Day 3: Problems
Extend the "sizer" application to count and size links
Take the sizer application (code, output) and add a message to count the number of links on a page. Follow these links and calculate their size as well, so you get the total size for each page.
The code:
The output:
This problem was a great way to experiment with actors in Scala. The sequential solution is self explanatory, so heres an outline of the concurrent one:
The sequential code takes nearly 20 seconds to run while the concurrent code takes less than 3 seconds, a 7x improvement. The concurrent code is definitely more complicated, but not unreasonably so. Even though it was my first time using Scala actors, the code took less than 30 minutes to get working, much of it spent learning about the self keyword. In fact, I find it very easy to reason about Scalas actors, which is not something I can say for Javas synchronized keyword, Executors, Runnable, and various other concurrency constructs.
User experience counts. Not only for products, but for programming languages too.
How many features can you pile into one language before it becomes too cumbersome? How much syntax sugar do you need to understand to be able to read or write code? How many paradigms and mental models do I need to juggle to follow along? Do so many options make a language more flexible or less? Will there be such a thing as "idiomatic Scala" or will it be a free-for-all? Is it better to have a dozen ways to do something or one "proper" and well known way?
I dont know the answers to these questions, but I suspect theyll have a large impact on Scalas adoption. In the meantime, Ill keep hacking away at it to see what I can learn.
On to Erlang
Read on to learn about the next language in the book, Erlang.
Read more »
Scala, Day 3: Thoughts
After two lengthy chapters on the object oriented and functional programming syntax/options in Scala, the third day rushes through some of the most intriguing features, including pattern matching and concurrency via actors. I would have preferred to spend a bit more time on these complicated topics.
In fact, I had the same complaint on Day 3 of IO, where we also blasted through a discussion of concurrency in just a few pages. I respect the difficulties of plowing through seven different languages in a single book and dont expect deep discussions of any one of them, but I think the book wouldve been stronger if it had a greater bias towards the more advanced "day 3" topics of each language instead of basic syntax discussions in day 1 or 2.
Scala, Day 3: Problems
Extend the "sizer" application to count and size links
Take the sizer application (code, output) and add a message to count the number of links on a page. Follow these links and calculate their size as well, so you get the total size for each page.
The code:
The output:
This problem was a great way to experiment with actors in Scala. The sequential solution is self explanatory, so heres an outline of the concurrent one:
- The caller creates B Base Actors, one for each of the B base URLs.
- The caller then calls receive to await a message from each Base Actor.
- Each Base Actor concurrently loads its base URL, finds the links on the page, and creates L Link Actors, one for each of the L links on the page.
- The Base Actor then calls receive to await a message from each of its Link Actors.
- Each Link Actor concurrently loads the page for its given link and sends a message to its parent Base Actor with the size of that page.
- When the Base Actor has received a message form each of his Link Actors, it sends a message to the caller with the total size.
- When the caller has received B messages from the Base Actors, we are done.
The sequential code takes nearly 20 seconds to run while the concurrent code takes less than 3 seconds, a 7x improvement. The concurrent code is definitely more complicated, but not unreasonably so. Even though it was my first time using Scala actors, the code took less than 30 minutes to get working, much of it spent learning about the self keyword. In fact, I find it very easy to reason about Scalas actors, which is not something I can say for Javas synchronized keyword, Executors, Runnable, and various other concurrency constructs.
Wrapping up Scala
Im a bit torn when it comes to Scala. Most of the time, I saw it as a vastly improved version of Java. The support for closures, functional programming, pattern matching, and actors all seem like genuinely useful tools that would dramatically improve productivity, code readability, correctness, and expressiveness. Ive already used Scala in a few of my projects to build some features that wouldve been nearly impossible or incredibly ugly in Java.
However, even in my limited exposure to Scala, Ive already come across a number of hiccups. As I mentioned on day 1, the API docs are not helpful and look like they are written for academics. The IDE support is vastly inferior compared to Java. Ive now tried both Eclipse and IntelliJ, and each one has significant problems: e.g. missing compile errors on some code; finding errors on other code thats actually valid; broken/missing auto complete; issues with the refactor/rename functionality; poor support for running scripts. The compiler is slow. The type hierarchies are complicated. Type inference doesnt always work as well as youd hope.
However, there is one issue that worries me above all else: feature overload. It seems like Scala is trying to be all things to all people. Its object oriented; its functional; it has type inference; it has lots of syntactic sugar; it has actors; its compatible with Java; it has first class support for XML; they are even trying to add macros. While all of these features could lead to an incredibly powerful language, they could also lead to one thats incredibly complicated and difficult to use.
How many features can you pile into one language before it becomes too cumbersome? How much syntax sugar do you need to understand to be able to read or write code? How many paradigms and mental models do I need to juggle to follow along? Do so many options make a language more flexible or less? Will there be such a thing as "idiomatic Scala" or will it be a free-for-all? Is it better to have a dozen ways to do something or one "proper" and well known way?
I dont know the answers to these questions, but I suspect theyll have a large impact on Scalas adoption. In the meantime, Ill keep hacking away at it to see what I can learn.
On to Erlang
Read on to learn about the next language in the book, Erlang.
Subscribe to:
Posts (Atom)