When should I use Actors vs. Remote Actors in Akka?
I understand that both can scale a machine up, but only remote actors can scale out, so is there any practical production use of the normal Actor?
If a remote actor only has a minor initial setup overhead and does not have any other major overhead to that of a normal Actor, then I would think that using a Remote Actor would be the standard, since it can scale up and out with ease. Even if there is never a need to scale production code out, it would be nice to have the option (if it doesn't come with baggage).
Any insight on when to use an Actor vs. Remote Actor would be much appreciated.
Remote Actors cannot scale up, they are only remote references to a local actor on another machine.
For Akka 2.0 we will introduce clustered actors, which will allow you to write an Akka application and scale it up only using config.
Regular Actors can be used in sending out messages in local project.
As for the Remote Actors, you can used it in sending out messages to dependent projects that are connected to the project sending out the message.
Please refer here for the Remote Akka Actors
http://doc.akka.io/docs/akka/snapshot/scala/remoting.html
The question asks "If a remote actor only has a minor initial setup overhead and does not have any other major overhead then I would think that using a Remote Actor would be the standard". Yet the Fallacies of distributed computing make the point that it is a design error to assume that remoting with any technology has no overhead. You have the overhead of copying the messages to bytes and transmitting it across the network interface. You also have all the complexity of different processes being up, down, stalled or unreachable and of the network having hiccups leading to lost, duplicated or reordered messages.
This great article has real world examples of weird network errors which make remoting hard to make bullet proof. The Akka project lead Roland Kuln in his free video course about akka says that in his experience for every 1T of network messages being sent he sees a corruption. Notes on Distributed Systems for Young Bloods says "distributed systems tend to need actual, not simulated, distribution to flush out their bugs" so even good unit tests wont make for a perfect system. There is lots of advice that remoting is not "free" but hard work to get perfect.
If you need to use remoting for availability, or to move to huge scale, then note that akka does at-least-once delivery with possible duplication. So you must ensure that duplicated messages don't create bad results.
The moment you start to use remoting you have a distributed system which creates challenges which are discussed in Distributed systems for fun and profit. Unless you are doing very simply things like stateless calculators that are idempotent to duplicated messages things get tricky. One of assignments on that akka video course at the link above is to make a replicated key-value store which can deal with lost messages by writing the logic yourself. Its far from being an easy assignment. State distributed across different processes gets very hard, actors encapsulate state, therefore distributing actors can get very hard, depending on the consistency and availability requirements of the system you are building.
This all implies that if you can avoid remoting and achieve what you need to achieve then you would be wise to avoid it. If you do need remoting then Akka makes it easy due to its location transparency. So whilst its a great toolbox to take with you on the job; you should double check if the job needs all the tools or only the simplest ones in the box.
Related
I am currently working on a Play! project that has the following architecture:
Controllers -> Services (actors) -> Models (Regular case classes)
For each request that comes in, we will issue a call to the service layers like so:
Service ? DoSomething(request, context)
We have a set number of these service actors behind an akka router that are created during app initialization, and is expandable on demand.
And in the service we mostly do modest data manipulation or database calls:
receive = {
case DoSomething(x, y) => {
...
Model.doSometing(...)
sender ! result
}
}
I am having second thoughts on whether we should be using actors for our services or just use Futures only.
We do not have any internal state that needs to be modified in the service actors, whatever message comes in goes to a function and spits out the result. Isn't this the big strength of the actor model?
We are doing a lot of tasks which seem to take a lot away from the actor model
We aren't doing heavy computation and remoting doesn't make sense because most of the work is for the database and roundtriping to a remote actor to make some db call is unnecessary
We do use reactivemongo, so every db call is non blocking. We can make a lot of these calls
It seems to me that removing akka and just use Futures makes our life a lot easier, and we don't really lose anything.
There certainly is no shortage of opinion on the topic of what should and shouldn't be an actor. Like these two posts:
http://noelwelsh.com/programming/2013/03/04/why-i-dont-like-akka-actors/
http://www.chrisstucchio.com/blog/2013/actors_vs_futures.html
I don't think you're going to find an absolute answer to this question other then that it's situational and it's really up to your preferences and your problem. What I can do for you is to offer my opinion that is based on us implementing Akka for about 2 years now.
For me, I like to think of Akka really as a platform. We come for the Actor Model but we stay for all of the other goodness that the platform provides like Clustering/Remoting, FSM, Routing, Circuit Breaker, Throttling and the like. We are trying to build an SOA like architecture with our actors acting as services. We are deploying these services across a cluster, so we are taking advantage of things like Location Transparency and Routing to provide the ability for a service consumer (which itself could be another service) to find and use a service no matter where it is deployed, and in a highly available manner. Akka makes this whole process pretty simple based on the platform tools they offer.
Within our system, we have the concept of what I call Foundation Services. These are really simple services (like basic lookup/management services for a particular entity). These services generally don't call any other services, and in some cases, just perform DB lookups. These services are pooled (router) and don't usually have any state. They are pretty similar to what you are describing some of your services to be like. We then start to build more and more complex services on top of these foundation services. Most of these services are short lived (to avoid asking), sometimes FSM based, that collect data from the foundation services and then crunch and do something as a result. Even though these foundation services are themselves pretty simple, and some would say don't require an actor, I like the flexibility in that when I compose them into a higher level service, I can look them up and they can be anywhere (location transparent) in my cluster with any number of instances available (routing) for using.
So for us, it was really a design decision to baseline around an actor as a sort of micro-like service that is available within our cluster for consumption by any other service, no matter how simple that service is. I like communicating with these services, where ever they are, through a coarse grained interface in an async manner. A lot of those principles are aspects of building out a good SOA. If that's your goal, then I think Akka can be very helpful in achieving that goal. If you are not looking to do something like that, then maybe your are right in questioning your decision to use Akka for your services. Like I said earlier, it's really up to you to figure out what you are trying to do from an architecture perspective and then design your services layer to meet those goals.
I think you're on right tracks.
We do not have any internal state that needs to be modified in the service actors, whatever message comes in goes to a function and spits out the result. Isn't this the big strength of the actor model?
I found Chris Stucchio's blog (referred to by #cmbaxter above) instantly delightful. My case was so simple that architectural considerations were not a valid point. Just Spray routing and lots of database access, like you have. No state. Thus Future. So much simpler code.
Actor should be crated when you need some really long living stuff with modifying state. In other cases there are no any benefits from actors, especially from non-typed ones.
- do pattern matching every time
- control actor's lifecycle
- remember the things which should not be passed between threads
Why do all of this when you may have simple Future?
There are tasks where actors fit very well, but not everywhere
I was wondering the same and what we decide to do was to use Akka for our data access and it works very well, it's very testable (and tested), and very portable.
We created repositories, long living actors, that we bootstrapped in our App : (FYI, we are using slick for our DB Access, but also have a similar design for our MongoDB needs)
val subscriptionRepo = context.actorOf(Props(new SubscriptionRepository(appConfig.db)), "repository-subscription")
Now we are able to send a "Request" Message for data, ex:
case class SubscriptionsRequested(atDate: ZeroMillisDateTime)
that the actor will respond with
case class SubscriptionsFound(users: Seq[UserSubscription])
or Failure(exception)
In our situation (spray apps but also CLI), we wrapped those calls in a short living actor that take the context and complete on reception and closes itself. (You could handle domain specific logic in those actor, have it to extend another actor that manages its lifecycle and exception so you would only have to specify a partial function for your needs and leave the abstract actor to deal with timeouts, common exceptions etc.
We also have situations where we needed more work to be done in the initiating actor, and it very convenient to fire x messages to your repositories and have your actor storing those messages as they arrive, doing something once they are all there, firing back for completion to the sender( for instance) and close itself.
Thanks to this design, we have a very reactive repository living outside our app, completely tested with Akka TestKit and H2, completely DB agnostic, and it's dead easy, to access data from our DBs (and we never do any ASK, only Tell : Tell to repo, tell to sender, complete or x Tells to repos, pattern match on expected results until completion, tell to sender).
If I have a server running on my machine, and several clients running on other networks, what are some concepts of testing for synchronicity between them? How would I know when a client goes out-of-sync?
I'm particularly interested in how network programmers in the field of game design do this (or just any continuous network exchange application), where realtime synchronicity would be a commonly vital aspect of success.
I can see how this may be easily achieved on LAN via side-by-side comparisons on separate machines... but once you branch out the scenario to include clients from foreign networks, I'm just not sure how it can be done without clogging up your messaging system with debug information, and therefore effectively changing the way that synchronicity would result without that debug info being passed over the network.
So what are some ways that people get around this issue?
For example, do they simply induce/simulate latency on the local network before launching to foreign networks, and then hope for the best? I'm hoping there are some more concrete solutions, but this is what I'm doing in the meantime...
When you say synchronized, I believe you are talking about network latency. Meaning, that a client on a local network may get its gaming information sooner than a client on the other side of the country. Correct?
If so, then I'm sure you can look for books or papers that cover this kind of topic, but I can give you at least one way to detect this latency and provide a way to manage it.
To detect latency, your server can use a type of trace route program to determine how long it takes for data to reach each client. A common Linux program example can be found here http://linux.about.com/library/cmd/blcmdl8_traceroute.htm. While the server is handling client data, it can also continuously collect the latency statistics and provide the data to the clients. For example, the server can update each client on its own network latency and what the longest latency is for the group of clients that are playing each other in a game.
The clients can then use the latency differences to determine when they should process the data they receive from the server. For example, a client is told by the server that its network latency is 50 milliseconds and the maximum latency for its group it 300 milliseconds. The client then knows to wait 250 milliseconds before processing game data from the server. That way, each client processes game data from the server at approximately the same time.
There are many other (and probably better) ways to handle this situation, but that should get you started in the right direction.
I've been tasked with creating an intermediate layer which needs to exchange data (over HTTP) between two independent systems (e.g. Receiver <=> Intermediate Layer (IL) <=> Sender). Receiver and Sender both expose a set of API's via Web Services. Everytime a transaction occurs in the Sender system, the IL should know about it (I'm thinking of creating a Windows Service which constantly pings the Sender), massage the data, then deliver it to the Receiver. The IL can temporarily store the data in a SQL database until it is transferred to the Receiver. I have the following questions -
Can WCF (haven't used it a lot) be used to talk to the Sender and Receiver (both expose web services)?
How do I ensure guaranteed delivery?
How do I ensure security of the messages over the Internet?
What are best practices for handling concurrency issues?
What are best practices for error handling?
How do I ensure reliability of the data (data is not tampered along the way)
How do I ensure the receipt of the data back to the Sender?
What are the constraints that I need to be aware of?
I need to implement this on MS platform using a custom .NET solution. I was told not to use any middleware like BizTalk. The receiver is an SDFC instance, if that matters.
Any pointers are greatly appreciated. Thank you.
A Windows Service that orchestras the exchange sounds fine.
Yes WCF can deal with traditional Web Services.
How do I ensure guaranteed delivery?
To ensure delivery you can use TransactionScope to handle the passing of data between the
Receiver <=> Intermediate Layer and Intermediate Layer <=> Sender but I wouldn't try and do them together.
You might want to consider some sort of queuing mechanism to send the data to the receiver; I guess I'm thinking more of a logical queue rather than an actual queuing component. A workflow framework could also be an option.
make sure you have good logging / auditing in place; make sure it's rock solid, has the right information and is easy to read. Assuming you write a service it will execute without supervision so the operational / support aspects are more demanding.
Think about scenarios:
How do you manage failed deliveries?
What happens if the reciever (or sender) is unavailbale for periods of time (and how long is that?); for example: do you need to "escalate" to an operator via email?
How do I ensure security of the messages over the Internet?
HTTPS. Assuming other existing clients make calls to the Web Services how do they ensure security? (I'm thinking encryption).
What are best practices for handling concurrency issues?
Hmm probably a separate question. You should be able to find information on that easily enough. How much data are we taking? what sort of frequency? How many instances of the Windows Service were you thinking of having - if one is enough why would concurrency be an issue?
What are best practices for error handling?
Same as for concurrency, but I can offer some pointers:
Use an established logging framework, I quite like MS EntLibs but there are others (re-using whatever's currently used is probably going to make more sense - if there is anything).
Remember that execution is unattended so ensure information is complete, clear and unambiguous. I'd be tempted to log more and dial it down once a level of comfort is reached.
use a top level handler to ensure nothing get's lost; but don;t be afraid to log deep in the application where you can still get useful context (like the metadata of the data being sent / recieved).
How do I ensure the receipt of the data back to the Sender?
Include it (sending the receipt) as a step that is part of the transaction.
On a different angle - have a look on CodePlex for ESB type libraries, you might find something useful: http://www.codeplex.com/site/search?query=ESB&ac=8
For example ESBasic which seems to be a class library which you could reuse.
How are Massively Multiplayer Online RPG games built?
What server infrastructure are they built on? especially with so many clients connected and communicating in real time.
Do they manage with scripts that execute on page requests? or installed services that run in the background and manage communication with connected clients?
Do they use other protocols? because HTTP does not allow servers to push data to clients.
How do the "engines" work, to centrally process hundreds of conflicting gameplay events?
Thanks for your time.
Many roads lead to Rome, and many architectures lead to MMORPG's.
Here are some general thoughts to your bullet points:
The server infrastructure needs to support the ability to scale out... add additional servers as load increases. This is well-suited to Cloud Computing by the way. I'm currently running a large financial services app that needs to scale up and down depending on time of day and time of year. We use Amazon AWS to almost instantly add and remove virtual servers.
MMORPG's that I'm familiar with probably don't use web services for communication (since they are stateless) but rather a custom server-side program (e.g. a service that listens for TCP and/or UDP messages).
They probably use a custom TCP and/or UDP based protocol (look into socket communication)
Most games are segmented into "worlds", limiting the number of players that are in the same virtual universe to the number of game events that one server (probably with lots of CPU's and lots of memory) can reasonably process. The exact event processing mechanism depends on the requirements of the game designer, but generally I expect that incoming events go into a priority queue (prioritized by time received and/or time sent and probably other criteria along the lines of "how bad is it if we ignore this event?").
This is a very large subject overall. I would suggest you check over on Amazon.com for books covering this topic.
What server infrastructure are they built on? especially with so many clients connected and communicating in real time.
I'd guess the servers will be running on Linux, BSD or Solaris almost 99% of the time.
Do they manage with scripts that execute on page requests? or installed services that run in the background and manage communication with connected clients?
The server your client talks to will be a server running a daemons or service that sits idle listening for connections. For instances (dungeons), usually a new process is launched for each group, which would mean there is a dispatcher service somewhere mananging this (analogous to a threadpool)
Do they use other protocols? because HTTP does not allow servers to push data to clients.
UDP is the protocol used. It's fast as it makes no guarantees the packet will be received. You don't care if a bit of latency causes the client to lose their world position.
How do the "engines" work, to centrally process hundreds of conflicting gameplay events?
Most MMOs have zones which limit this to a certain amount of people. For those that do have 100s of people in one area, there is usually high latency. The server is having to deal with 100s of spells being sent its way, which it must calculate damage amounts for each one. For the big five MMOs I imagine there are teams of 10-20 very intelligent, mathematically gifted developers working on this daily and there isn't a MMO out there that has got it right yet, most break after 100 players.
--
Have a look for Wowemu (there's no official site and I don't want to link to a dodgy site). This is based on ApireCore which is an MMO simulator, or basically a reverse engineer of the WoW protocol. This is what the private WoW servers run off. From what I recall Wowemu is
mySQL
Python
However ApireCore is C++.
The backend for Wowemu is amazingly simple (I tried it in 2005 however) and probably a complete over simplification of the database schema. It does gives you a good idea of what's involved.
Because MMOs by and large require the resources of a business to develop and deploy, at which point they are valuable company IP, there isn't a ton of publicly available information about implementations.
One thing that is fairly certain is that since MMOs by and large use a custom client and 3D renderer they don't use HTTP because they aren't web browsers. Online games are going to have their own protocols built on top of TCP/IP or UDP.
The game simulations themselves will be built using the same techniques as any networked 3D game, so you can look towards resources for that problem domain to learn more.
For the big daddy, World of Warcraft, we can guess that their database is Oracle because Blizzard's job listings frequently cite Oracle experience as a requirement/plus. They use Lua for user interface scripting. C++ and OpenGL (for Mac) and Direct3D (for PC) can be assumed as the implementation languages for the game clients because that's what games are made with.
One company that is cool about discussing their implementation is CCP, creators of Eve online. They have published a number of presentations and articles about Eve's infrastructure, and it is a particularly interesting case because they use Stackless Python for a lot of Eve's implementation.
http://www.disinterest.org/resource/PyCon2006-StacklessInEve.wmv
http://us.pycon.org/2009/conference/schedule/event/91/
There was also a recent Game Developer Magazine article on Eve's architecture:
https://store.cmpgame.com/product/3359/Game-Developer-June%7B47%7DJuly-2009-Issue---Digital-Edition
The Software Engineering radio podcast had an episode with Jim Purbrick about Second Life which discusses servers, worlds, scaling and other MMORPG internals.
Traditionally MMOs have been based on C++ server applications running on Linux communicating with a database for back end storage and fat client applications using OpenGL or DirectX.
In many cases the client and server embed a scripting engine which allows behaviours to be defined in a higher level language. EVE is notable in that it is mostly implemented in Python and runs on top of Stackless rather than being mostly C++ with some high level scripts.
Generally the server sits in a loop reading requests from connected clients, processing them to enforce game mechanics and then sending out updates to the clients. UDP can be used to minimize latency and the retransmission of stale data, but as RPGs generally don't employ twitch gameplay TCP/IP is normally a better choice. Comet or BOSH can be used to allow bi-directional communications over HTTP for web based MMOs and web sockets will soon be a good option there.
If I were building a new MMO today I'd probably use XMPP, BOSH and build the client in JavaScript as that would allow it to work without a fat client download and interoperate with XMPP based IM and voice systems (like gchat). Once WebGL is widely supported this would even allow browser based 3D virtual worlds.
Because the environments are too large to simulate in a single process, they are normally split up geographically between processes each of which simulates a small area of the world. Often there is an optimal population for a world, so multiple copies (shards) are run which different sets of people use.
There's a good presentation about the Second Life architecture by Ian Wilkes who was the Director of Operations here: http://www.infoq.com/presentations/Second-Life-Ian-Wilkes
Most of my talks on Second Life technology are linked to from my blog at: http://jimpurbrick.com
Take a look at Erlang. It's a concurrent programming language and runtime system, and was designed to support distributed, fault-tolerant, soft-real-time, non-stop applications.
I want to know the technical reasons why the lift webframework has high performance and scalability? I know it uses scala, which has an actor library, but according to the install instructions it default configuration is with jetty. So does it use the actor library to scale?
Now is the scalability built right out of the box. Just add additional servers and nodes and it will automatically scale, is that how it works? Can it handle 500000+ concurrent connections with supporting servers.
I am trying to create a web services framework for the enterprise level, that can beat what is out there and is easy to scale, configurable, and maintainable. My definition of scaling is just adding more servers and you should be able to accommodate the extra load.
Thanks
Lift's approach to scalability is within a single machine. Scaling across machines is a larger, tougher topic. The short answer there is: Scala and Lift don't do anything to either help or hinder horizontal scaling.
As far as actors within a single machine, Lift achieves better scalability because a single instance can handle more concurrent requests than most other servers. To explain, I first have to point out the flaws in the classic thread-per-request handling model. Bear with me, this is going to require some explanation.
A typical framework uses a thread to service a page request. When the client connects, the framework assigns a thread out of a pool. That thread then does three things: it reads the request from a socket; it does some computation (potentially involving I/O to the database); and it sends a response out on the socket. At pretty much every step, the thread will end up blocking for some time. When reading the request, it can block while waiting for the network. When doing the computation, it can block on disk or network I/O. It can also block while waiting for the database. Finally, while sending the response, it can block if the client receives data slowly and TCP windows get filled up. Overall, the thread might spend 30 - 90% of it's time blocked. It spends 100% of its time, however, on that one request.
A JVM can only support so many threads before it really slows down. Thread scheduling, contention for shared-memory entities (like connection pools and monitors), and native OS limits all impose restrictions on how many threads a JVM can create.
Well, if the JVM is limited in its maximum number of threads, and the number of threads determines how many concurrent requests a server can handle, then the number of concurrent requests will be determined by the number of threads.
(There are other issues that can impose lower limits---GC thrashing, for example. Threads are a fundamental limiting factor, but not the only one!)
Lift decouples thread from requests. In Lift, a request does not tie up a thread. Rather, a thread does an action (like reading the request), then sends a message to an actor. Actors are an important part of the story, because they are scheduled via "lightweight" threads. A pool of threads gets used to process messages within actors. It's important to avoid blocking operations inside of actors, so these threads get returned to the pool rapidly. (Note that this pool isn't visible to the application, it's part of Scala's support for actors.) A request that's currently blocked on database or disk I/O, for example, doesn't keep a request-handling thread occupied. The request handling thread is available, almost immediately, to receive more connections.
This method for decoupling requests from threads allows a Lift server to have many more concurrent requests than a thread-per-request server. (I'd also like to point out that the Grizzly library supports a similar approach without actors.) More concurrent requests means that a single Lift server can support more users than a regular Java EE server.
at mtnyguard
"Scala and Lift don't do anything to either help or hinder horizontal scaling"
Ain't quite right. Lift is highly statefull framework. For example if a user requests a form, then he can only post the request to the same machine where the form came from, because the form processeing action is saved in the server state.
And this is actualy a thing which hinders scalability in a way, because this behaviour is inconistent to the shared nothing architecture.
No doubt that lift is highly performant but perfomance and scalability are two different things. So if you want to scale horizontaly with lift you have to define sticky sessions on the loadbalancer which will redirect a user during a session to the same machine.
Jetty maybe the point of entry, but the actor ends up servicing the request, I suggest having a look at the twitter-esque example, 'skitter' to see how you would be able to create a very scalable service. IIRC, this is one of the things that made the twitter people take notice.
I really like #dre's reply as he correctly states the statefulness of lift being a potential problem for horizontal scalability.
The problem -
Instead of me describing the whole thing again, check out the discussion (Not the content) on this post. http://javasmith.blogspot.com/2010/02/automagically-cluster-web-sessions-in.html
Solution would be as #dre said sticky session configuration on load balancer on the front and adding more instances. But since request handling in lift is done in thread + actor combination you can expect one instance handle more requests than normal frameworks. This would give an edge over having sticky sessions in other frameworks. i.e. Individual instance's capacity to process more may help you to scale
you have Akka lift integration which would be another advantage in this.