Since there are actor libraries like akka targeting the JVM and claiming to offer similar benefits, why should someone prefer erlang?
And because the JVM does global stop-the-world garbage collection, which makes soft real-time implausible because of the unpredictability of GC affecting your actors. Erlang has per-process heaps.
Basically the Erlang VM was created for this use case while the JVM was not, and its not something you can just add with a library.
edit: Also the lightweightness of Erlang processes compared to Java threads[1] and hot code upgrades.
Not quite. Every environment needs some shared memory semantics. In Erlang that's done with ETS tables, which don't undergo GC at all. Java's GCs are so good now, that, if shared data structures are used sparingly, would still give you better performance than BEAM. Plus, you have commercial pauseless GCs, as well as hard realtime JVMs.
> Basically the Erlang VM was created for this use case while the JVM was not, and its not something you can just add with a library. edit: Also the lightweightness of Erlang processes compared to Java threads[1] and hot code upgrades.
The JVM is so versatile, though, that all this can actually be added as a library, including true lightweight threads and hot code swapping (see Quasar[1]).
Nevertheless, BEAM was indeed designed with process isolation in mind, and on the JVM, actors might interfere with one another's execution more so than on BEAM, but even on BEAM you get the occasional full VM crashes. If total process isolation is not your main concern, you might find that Java offers more than Erlang, all things considered.
"Assume developers are non-malicious and will only pass immutable objects across actor/future boundaries."
Also, I'm not that familiar with Erlang's memory model, so I might be wrong on this. But as far as I'm aware the memory for a message in Erlang is shared between threads - it's only local variables that use private memory. This means Erlang will also need some sort of concurrent garbage collector - does Erlang's version not stop the world, or at least the messaging subsystem?
Yes and no. Some large binaries (a specific Erlang data type, that can say represent a packet or block of data from disk), will be shared and reference counted when passed between processes instead of copied. They are immutable just like most datatypes in Erlang. These binaries have a specific GC algorithm that it just might take longer sometime for them to be reclaimed. But it seems all that could presumably be done via atomic updates to counters and references.
In general most messages are copied on send. So implementation wise GC is very simple then. On another level because data in Erlang is immutable, the fact that messages get copied is also an implementation detail! One could conceive another implementation of a VM that only passes references and immutable data on message send (well minus when it sends it to another machine, of course). But that would make GC a bit more tricky just like in case of those binaries.
Another common issues is taking references to a small part of a large shared binary.
No, the messages are truly copied: http://jlouisramblings.blogspot.dk/2013/10/embrace-copying.h...
edit: the exception being large binaries apparently
Does it? Back when I was in HFT, we were definitely running a JVM with background thread GC.
It is really a fantastic piece of technology:
http://www.azulsystems.com/zing/pgc
Even just marveling at the complexity and how they got it working.
Otherwise, besides those tricks, how would you do it when you have multiple threads accessing objects on a shared heap?
Erlang's VM is another even wonderful piece of engineering. Each little process lives in its own memory heap. Then pauseless garbage collection become trivial. It has many other really cool and unique features (hot code reloading, inter-node distribution, ability to load C code, etc etc...)
How hard are the limits of "soft" realtime?
Do any of the library actor models offer something like that?
We found that the difficulty didn't lay in getting Akka to work in a clustered fashion—that was simple—but rather in architecting our backend's work distribution mechanics so as not to overload any given node. I blogged about our experience with Akka: http://blog.goconspire.com/post/64130417462/akka-at-conspire...
The closer abstraction is probably OS processes + IPC. Then you get closer to the spirit of it. And Chrome browser and other software take that approach. It isolates faults. But well, you have to do a lot more work around it and those are not exactly light-weight. Erlang processes are only a few K of memory each.
The closer abstraction is probably more like a distributed fault tolerant OS + processes using network transparent IPC.
Erlang's process are just threads pretending to be process which will spawn much faster than Akka.
I believe there are a few articles of Akka actor limitation versus Erlang's. I haven't delve deep into this but there are caveat with receiving message and how to handle it versus Erlang not having such caveat.
Coding in a language that isn't built with Actor/Concurrency in mind is a huge pain in the butt. Think of Javascript and Node.js and callback hells, which of course have push Javascript to adopt things such as future and etc..
Of course you can say Scala is built with concurrency in mind same with Clojure. But the underlying gears, the JVM was not compares to Erlang.
There are trade off between Erlang's VM and Java's VM. And if your requirement is a perfect match for either Erlang and Java you mind as well pick the best because coding against what the tool was intended for is just for people who enjoy pain and frustration.