A few times, I've seen the following take on the relative mertis of Scala and Erlang: Erlang is great for concurrent programming and it has a great track record in its niche, but it's unlikely to become mainstream because it's foreign and it doesn't have as many libraries as Java. Scala, on the hand, has the best of both worlds. Its has functional semantics, its Actors library provides Erlang style concurrency, and it runs on the JVM and it has access to all the Java libraries. This combination makes Scala it a better choice for building concurrent applications, especially for companies that are invested in Java.
I haven't coded in Scala, but I did a good amount of research on it and it looks like a great language. Some of the best programmers I know rave about it. I think that Scala can be a great replacement for Java. Function objects, type inference, mixins and pattern matching are all great language features that Scala has and that are sorely missing from Java.
Although I believe Scala is a great language that is clearly superior to Java, Scala doesn't supersede Erlang as my language of choice for building high-availability, low latency, massively concurrent applications. Scala's Actors library is a big improvement over what Java has to offer in terms of concurrency, but it doesn't provide all the benefits of Erlang-style concurrency that make Erlang such a great tool for the job. I did a good amount of research into the matter and these are the important differences I think one should consider when choosing between Scala and Erlang. (If I missed something or got something wrong, please let me know. I don't profess to be a Scala expert by any means.)
Scala's Actor library does a good job at emulating Erlang style message passing. Similar to Erlang processes, Scala actors send and receive messages through mailboxes. Like Erlang, Scala has pattern matching sematics for receiving messages, which results in elegant, concise code (although I think Erlang's simpler type system makes pattern matching easier in Erlang).
Scala's Actors library goes pretty far, but it doesn't (well, it can't) provide an important feature that makes concurrent programming so easy in Erlang: immutability. In Erlang, multiple processes can share the same data within the same VM, and the language guarantees that race conditions won't happen because this data is immutable. In Scala, though, you can send between actors pointers to mutable objects. This is the classic recipe for race conditions, and it leaves you just where you started: having to ensure synchronized access to shared memory.
If you're careful, you may be able to avoid this problem by copying all messages or by treating all sent objects as immutable, but the Scala language doesn't guarantee safe access to shared objects. Erlang does.
Hot code swapping
Hot code swapping it a killer feature. Not only does it (mostly) eliminates the downtime required to do code upgrades, it also makes a language much more productive because it allows for true interactive programming. With hot code swapping, you can immediately test the effects of code changes without stopping your server, recompiling your code, restarting your server (and losing the application's state), and going back to where you had been before the code change. Hot code swapping is one of the main reasons I like coding in Erlang.
The JVM has limited support for hot code swapping during development -- I believe it only lets you change a method's body at runtime (an improvement for this feature is in Sun's top 25 RFE's for Java). This capability is not as robust as Erlang's hot code swapping, which works for any code modification at any time.
A great aspect of Erlang's hot code swapping is that when you load new code, the VM keeps around the previous version of the code. This gives running processes an opportunity to receive a message to perform a code swap before the old version of the code is finally removed (which kills processes that didn't perform a code upgrade). This feature is unique to Erlang as far as I know.
Hot code swapping is even more important for real-time applications that enable synchronous communications between users. Restarting such servers would cause user sessions to disconnect, which would lead to poor user experience. Imagine playing World of Warcraft and, in the middle of a major battle, losing your connection because the developers wanted to add a log line somewhere in the code. It would be pretty upsetting.
A common argument against GC'd languages is that they are unsuitable for low latency applications due to potential long GC sweeps that freeze the VM. Modern GC optimizations such as generational collection alleviate the problem somewhat, but not entirely. Occasionally, the old generation needs to be collected, which can trigger long sweeps.
Erlang was designed for building applications that have (soft) real-time performance, and Erlang's garbage collection is optimized for this end. In Erlang, processes have separate heaps that are GC'd separately, which minimizes the time a process could freeze for garbage collection. Erlang also has ets, an in-memory storage facility for storing large amounts of data without any garbage collection (you can find more information on Erlang GC at http://prog21.dadgum.com/16.html).
Erlang might not have a decisive advantage here. The JVM has a new concurrent garbage collector designed to minimize freeze times. This article and this whitepaper (PDF warning) have some information about how it works. This collector trades performance and memory overhead for shorter freezes. I haven't found any benchmarks that show how well it works in production apps, though, and if it is as effective as Erlang's garbage collector for low-latency apps.
The Erlang VM schedules processes preemptively. Each process gets a certain number of reductions (roughly equivalent to function calls) before it's swapped out for another process. Erlang processes can't call blocking operations that freeze the scheduler for long periods. All file IO and communications with native libraries are done in separate OS threads (communications are done using ports). Similar to Erlang's per-process heaps, this design ensures that Erlang's lightweight processes can't block each other. The downside is some communications overhead due to data copying, but it's a worthwhile tradeoff.
Scala has two types of Actors: thread-based and event based. Thread based actors execute in heavyweight OS threads. They never block each other, but they don't scale to more than a few thousand actors per VM. Event-based actors are simple objects. They are very lightweight, and, like Erlang processes, you can spawn millions of them on a modern machine. The difference with Erlang processes is that within each OS thread, event based actors execute sequentially without preemptive scheduling. This makes it possible for an event-based actor to block its OS thread for a long period of time (perhaps indefinitely).
According to the Scala actors paper, the actors library also implements a unified model, by which event-based actors are executed in a thread pool, which the library automatically resizes if all threads are blocked due to long-running operations. This is pretty much the best you can do without runtime support, but it's not as robust as the Erlang implementation, which guarantees low latency and fair use of resources. In a degenerate case, all actors would call blocking operations, which would increase the native thread pool size to the point where it can't grow anymore beyond a few thousand threads.
This can't happen in Erlang. Erlang only allocates a fixed number of OS threads (typically, one per processor core). Idle processes don't impose any overhead on the scheduler. In addition, spawning Erlang processes is always a very cheap operation that happens very fast. I don't think the same applies to Scala when all existing threads are blocked, because this condition first needs to be detected, and then new OS threads need to be spawned to execute pending Actors. This can add significant latency (this is admittedly theoretical: only benchmarks can show the real impact).
Depends on what you're doing, the difference between process scheduling in Erlang and Scala may not impact performance much. However, I personally like knowing with certainty that the Erlang scheduler can gracefully handle pretty much anything I throw at it.
One of Erlang's greatest strengths is that it unifies concurrent and distributed programming. Erlang lets you send a message to a process in the local or on a remote VM using exactly the same semantics (this is sometimes referred to as "location transparency"). Furthermore, Erlang's process spawning and linking/monitoring works seamlessly across nodes. This takes much of the pain out of building distributed, fault-tolerant applications.
The Scala Actors library has a RemoteActor type that apparently provides the similar location-transparency, but I haven't been able to find much information about it. According to this article, it's also possible to distribute Scala actors using Terracotta, which does distributed memory voodoo between nodes in a JVM cluster, but I'm not sure how well it works or how simple it is to set up. In Erlang, everything works out of the box, and it's so simple to get it working it's in the language's Getting Started manual.
Lightweight concurrency with no shared memory and pure message passing semantics is a fantastic toolset for building concurrent applications... until you realize you need shared (transactional) memory. Imagine building a WoW server, where characters can buy and sell items between each other. This would be very hard to build without a transactional DBMS of sorts. This is exactly what Mnesia provides -- with the a number of extra benefits such as distributed storage, table fragmentation, no impedance mismatch, no GC overhead (due to ets), hot updates, live backups, and multiple disc/memory storage options (you can read the Mnesia docs for more info). I don't think Scala/Java has anything quite like Mnesia, so if you use Scala you have to find some alternative. You would probably have to use an external DBMS such as MySQL cluster, which may incur a higher overhead than a native solution that runs in the same VM.
Functional programming and recursion go hand-in-hand. In fact, you could hardly write working Erlang programs without tail recursion because Erlang doesn't have loops -- it uses recursion for *everything* (which I believe is a good thing :) ). Tail recursion serves for more than just style -- it's also facilitates hot code swapping. Erlang gen_servers call their loop() function recursively between calls to 'receive'. When a gen_server receive a code_change message, they can make it a remote call (e.g. Module:loop()) to re-enter its main loop with the new code. Without tail recursion, this style of programming would quickly result in stack overflows.
From my research, I learned that Scala has limited support for tail recursion due to bytecode restrictions in most JVMs. From http://www.scala-lang.org/docu/files/ScalaByExample.pdf:
In principle, tail calls can always re-use the stack frame of the calling function. However, some run-time environments (such as the Java VM) lack the primitives to make stack frame re-use for tail calls efﬁcient. A production quality Scala implementation is therefore only required to re-use the stack frame of a directly tail-recursive function whose last action is a call to itself. Other tail calls might be optimized also, but one should not rely on this across implementations.
(If I understand the limitation correctly, tail call optimization in Scala only works within the same function (i.e. x() can make a tail recursive call to x(), but if x() calls y(), y() couldn't make a tail recursive call back to x().)
In Erlang, tail recursion Just Works.
Erlang processes are tightly integrated with the Erlang VM's event-driven network IO core. Processes can "own" sockets and send and receive messages to/from sockets. This provides the elegance of concurrency-oriented programming plus the scalability of event-driven IO (the Erlang VM uses epoll/kqueue under the covers). From Googling around, I haven't found similar capabilities in Scala actors, although they may exist.
In Erlang, you can get a remote shell into any running VM. This allows you to analyzing the state of the VM at runtime. For example, you can check how many processes are running, how much memory they consume, what data is stored Mnesia, etc.
The remote shell is also a powerful tool for discovering bugs in your code. When the server is in a bad state, you don't always have to try to reproduce the bug offline somehow to devise a fix. You can log right into it and see what's wrong. If it's not obvious, you can make quick code changes to add more logging and then revert them when you've discovered the problem. I haven't found a similar feature in Scala/Java from some Googling. It probably wouldn't be too hard to implement a remote shell for Scala, but without hot code swapping it would be much less useful.
Scala runs on the JVM, it can easily call any Java library, and it is therefore closer than Erlang to many programmers' comfort zones. However, I think that Erlang is very easy to learn -- definitely easier than Scala, which contains a greater total number of concepts you need to know in order to use the language effectively (especially if you consider the Java foundations on which Scala is built). This is to a large degree due to Erlang's dynamic typing and lack of object orientation. I personally prefer Erlang's more minimalist style, but this is a subjective matter and I don't want to get into religious debates here :)
Java indeed has a lot of libraries -- many more than Erlang. However, this doesn't mean that Erlang has no batteries included. In fact, Erlang's libraries are quite sufficient for many applications (you'll have to decide for yourself if they are sufficient for you). If you really need to use a Java library that doesn't have an Erlang equivalent, you could call it using Jinterface. It may or may not be a suitable option for your application. This can indeed be a deal breaker for some people who are deciding between the two languages.
There's an important difference between Java/Scala and Erlang libraries besides their relative abundance: virtually all "big" Erlang libraries use Erlang's features concurrency and fault tolerance. In the Erlang ecosystem, you can get web servers, database connection pools, XMPP servers, database servers, all of which use Erlang's lightweight concurrency, fault tolerance, etc. Most of Scala's libraries, on the other hand, are written in Java and they don't use Scala actors. It will take Scala some time to catch up to Erlang in the availability of libraries based on Actors.
Reliability and scalability
Erlang has been running massive systems for 20 years. Erlang-powered phone switches have been running with nine nines availability -- only 31ms downtime per year. Erlang also scales. From telcom apps to Facebook Chat we have enough evidence that Erlang works as advertised. Scala on the other hand is a relatively new language and as far as I know its actors implementation hasn't been tested in large-scale real-time systems.
I hope I did justice to Scala and Erlang in this comparison (which, by the way, took me way too much to write!). Regardless of these differences, though, I think that Scala has a good chance of being the more popular language of the two. Steve Yegge explains it better than I can:
Scala might have a chance. There's a guy giving a talk right down the hall about it, the inventor of – one of the inventors of Scala. And I think it's a great language and I wish him all the success in the world. Because it would be nice to have, you know, it would be nice to have that as an alternative to Java.
But when you're out in the industry, you can't. You get lynched for trying to use a language that the other engineers don't know. Trust me. I've tried it. I don't know how many of you guys here have actually been out in the industry, but I was talking about this with my intern. I was, and I think you [(point to audience member)] said this in the beginning: this is 80% politics and 20% technology, right? You know.
And [my intern] is, like, "well I understand the argument" and I'm like "No, no, no! You've never been in a company where there's an engineer with a Computer Science degree and ten years of experience, an architect, who's in your face screaming at you, with spittle flying on you, because you suggested using, you know... D. Or Haskell. Or Lisp, or Erlang, or take your pick."
Well, at least I'm not trying too hard to promote LFE... :)