Skip to content

thejerf/reign

Repository files navigation

reign 0.9.2

Build Status

Reign, "Rewrite Erlang In Go Nicely", is designed to make it easy to port Erlang or Erlang-style programs into Go by providing a framework for Erlang-style message passing, including the creation of clusters.

That is, this is not intended to be a "hey, look, if I wrap a mutex around a slice I get something that looks like a PID" four hours of screwing around in Go... this is intended to be a real, production-quality replacement for Erlang-style message passing functionality, suitable for porting existing Erlang programs out of Erlang without significant architecture overhauls.

This module is fully covered with godoc, but test coverage is not yet 100%. But it isn't bad.

If you do not need exactly Erlang-style semantics, note there are other libraries out there that may be better starting points. But I don't know of anything else for Go explicitly written with this goal in mind.

Current status: This is used in a production system in a 4-node cluster, and has been through some trial-by-fire. But as is the nature of this code, I'm sure there are more bugs we haven't tickled yet. I think if you want to convert an existing Erlang code base to Go, this is likely to be very, very helpful to you, but I don't promise it'll do everything you need perfectly or that you won't need to work with me to make some things work.

Getting Started

In order to address security concerns, reign uses TLS certificates signed by a central CA, and all nodes mutually verify the CA's signature with each other. To run reign at all requires creating a CA and certificates for each node.

Since it's a bit of a pain to do this in openssl, especially if you don't have it installed (Windows), a pure-go program to create some certificates is provided, as well as a sample program. If you have where ever "go get" puts binaries for you in your path, you can run:

# go get github.com/thejerf/reign/cmd/reign_init
# go get github.com/thejerf/reign/cmd/reign_sample
# reign_init
Signing certificate created
Constructed certificate for node 1
Constructed certificate for node 2

Note this will plop some certificates into your current directory. You may want to start in a temporary directory or something.

Now, in two separate shells, you can run:

# reign_sample 1
# reign_sample 2

This starts up a simple chat server, as documented in the sample.go program. I have chosen to create an example that focuses very specifically on the message-passing; consequently it is terribly, terribly ugly on the console. I assume you will prefer this to some slick presentation that obscures the message-passing under the pile of code it takes to make a halfway decent interface of any kind.

To get started, I suggest simply copying and pasting that example into a new workspace and start working with the code. That way you start with a working system the whole time.

Reign can use any certificates and certificate authorities that the Go TLS library can handle. The certificates created by this script may not entirely meet your needs. But they can get you started.

What Is "Erlang-like clustering"?

There are some common misconceptions about what "Erlang clustering" is. Erlang gives you two things:

  • A PID which can be sent messages from anywhere within a cluster. The message itself may include PIDs, which remain live and useful through the transfer.
  • Mnesia, a shared dabatase, which can contain PIDs.

The misconception is that Erlang gives you some sort of magic automatic clustering. It does not. It is still possible and even a bit easy to build applications that are locked to a single OS process. It is your job to take the Erlang primitives and build clusterable code.

What Is Reign?

Reign is a library to allow you to Rewrite Erlang In Go Nicely... that is, by allowing you to do fairly direct ports of your existing Erlang code without having to rearchitect it, or give up on clustering support.

This package provides a PID-like data structure and the code to use them across a cluster of Go processes, that may live on other machines. It is expected that even code porting from Erlang will use other database solutions, so there is no Mnesia equivalent, nor any desire to implement such a thing.

The goal is NOT a precise translation; for one thing, that's just plain impossible as both languages can easily do things the other can not. In particular, Erlang's native concurrency is asynchronous (that is, an Erlang message can be sent to a process whether or not it is receiving), and Go can use user-defined types, both of which have major impacts on properly structuring your program. The goal of this library is to give you a scaffolding that you can use to rewrite your Erlang code without having to entirely restructure it. In some sense, ideally over time you should use less and less of this library, however the goal is that this is production-quality and you should not be forced to do so, especially as there aren't necessarily pure-Go implementations of some of this functionality.

If you're using this to start a new Go project, you're probably doing it wrong. Probably. Maybe. To be honest while I have not yet pulled the trigger, I often find myself eyeing this library for my own projects even so.

Mailboxes And Addresses

Addresses are the equvialent of Erlang PIDs. Mailboxes are the equivalent of the implicit mailbox that an Erlang process has. Since goroutines can not carry implicit data, they are explicit in Reign. To avoid confusion about this unitary idea in Erlang being split in half in Reign, neither is called a "PID", especially since it isn't a "process ID" anyhow.

To get the equivalent of a PID in Erlang, once you have the cluster running, use NewMailbox from the connection service to receive a paired Address and Mailbox. The Address may be moved across the cluster transparently. The Mailbox is bound to the OS process that created it. Mailboxes provide both the ability to simply receive the next message and to pick a particular message out of the queue, with the same caveats in Go as there are in Erlang about the dangers of the queue backing up if you do not empty it.

In addition to asynchronous message queues and selective receive, reign implements an equivalent to "linking" in Erlang called via OnCloseNotify. When used like Erlang linking, this allows for some relatively safe RPC calling that will handle the mailbox (or the relevant node) being entirely unattached while you are trying to communicate with it. Exactly as with Erlang, this still isn't a guarantee of delivery, but it helps a lot.

Notes On Translation From Erlang

Go channels are synchronous. Erlang-style messages are asynchronous. In particular, this means that code written like this:

receive {msg1, Msg} -> {echo, Msg, self()} ! AnotherPID, receive {echoResponse, Msg2} -> ... end end

while idiomatic Erlang, is not generally safe in Go with the standard channels if you try to naively replace the receive statements with channel operations, because while you are processing a message, you are not able to receive any others. Should you end up sending a message that is a query to another such process, you'll deadlock on their finishing processing their message. The best case is weird latencies, the worst case is huge chunks of the system will end up deadlocked.

Of course, the correct Go answer is "don't do that", and when writing code from scratch it is a solvable problem. However, it is very easy for Erlang code to have accidentally embedded the asynchronous nature of messages quite deeply into its architecture, especially given the number of ways there are of using "receive" without realizing it. For instance, any gen_server that makes a call to another gen_server is exhibiting this pattern.

If you do not know for sure that the only possible next message for a mailbox is coming from the reply, you need to be sure to use .Receive() with a proper selection function in such cases.

Known Issues

In Erlang, a PID secretly contains more information that identifies something about when the node that originated the PID starts, or something like that, which prevents the "same" PID from being used from two different executions. This is easy to fix, but it isn't fixed yet.

I haven't yet benchmarked this very throughly, but generally speaking it seems to perform comparably to Erlang at worst, and better in a lot of common cases. There are obvious cases that are harder to compare, such as the impact of the Go GC on your runtime. This code does inevitably result in allocations.

Pre-1.0 Possible Changes

  • The NodeID type may be unexported. I'm not sure it's ever directly useful to external code.
  • I'd like to change the logging code to log with JSON objects that happen to render to strings, allowing transparent use with logrus.

Contribution Opportunities

Of course, it's open source. Contribute whatever you like. But in addition to resolving the known issues or the pre-1.0 possible changes, I would certainly be interested in:

  • More test coverage.

  • Adding an element to the Address that includes the node's start-up time or some other element that will be different between each run, so an address can't be obtained from one OS process, then that process dies and another starts in its place, and then someone re-uses that ID. Erlang has this.

  • Under the hood, we use channels for communication. The initial design did not. If a user is willing to commit to saying "the users of this mailbox only ever process the messages in order, so we do not need ReceiveNext", would it be possible to have what is probably a separate type of mailbox that can safely expose a channel?

    (Note it will never be possible to send messages directly to a channel. Go channels do not work over networks. It would be no different than you starting up a goroutine on your own that accepts messages and sends them somewhere. The previous paragraph proposes something that should not need a new goroutine spun up by Reign. (I want reign to spin up a relatively constant number of goroutines, O(nodes) in the worst case. O(mailboxes) is unacceptably expensive and impactful.)

  • Dynamically modifying the cluster at run time. Conceptually it's not too hard, "just work", but there's some i's to dot and t's to cross. I have some questions around the UI for this.

  • Is it possible to efficiently avoid the need to register message types? I'm vaguely nervous about looking at the gob error message and trying to infer whether I should call "register", or similar things, but it would make this easier to use.

  • Pluggable serialization types. Again, conceptually not that difficult, "just work".

  • Any additional similar Erlang-like services that are not covered by other packages, that are useful for Erlang code porting. e.g., I'm not really that interested in an Mnesia-like service for reign, because there are so many good database choices nowadays. (If you want, you can create such a thing, but make it a separate project. I'll happily link it.)

  • A coworker was doing some experimentation on having the mailboxes use channels instead of sync.Cond. It is not clear to either of whether this really clarified the code, or whether it would offer any performance advantages. (Unfortunately, it still wouldn't allow you to put a mailbox in a select statement. Mailbox semantics are fundamentally incompatible with select, as they are fundamentally asynchronous.) It certainly added a lot of lines of code.

  • Additional debugging capabilities. It would be interesting to split the private mailbox stuff off into a "reign/mailbox" package that made a lot more of the internals public for debugging, and then have the current mailbox.go offer the same external interface, but nicely "sealed". Using reign/mailbox would be something you might not want to do in production. But I've found it very useful over the years in Erlang to be able to peek into the mailboxes for diagnostics.

    And there's other capabilities that may need to be exposed, or need to better documented, or something. Another interesting idea would be to add an optional interface like type TracedMsg { ReignSent(); ReignReceived() } that can be called for an implementing type whenever a message is sent, received, transferred across a cluster, etc.

    Erlang's got certain advantages that barring the adoption of a standardized dynamic embedded language for Go, we can't really offer. But we can offer what we can!

  • On that note, any sort of integration with any REPL for Go would be accepted. I would hope it would take the form of implementing certain interfaces or something and not deep integration into the data types, though.

Release Notes

reign uses semantic versioning. As reign has not yet hit 1.0, note that API changes may still occur without a major version release.

  • 0.9.2
    • A somewhat experimental re-write of the internals to use channels instead of Conditions under the hood. This has a few API changes, to include the use of Contexts now.
  • 0.9.1
    • Initial release.

About

Rewrite Erlang In Go Nicely - a library for mimicking Erlang's mailboxes and clustering

Resources

License

Stars

Watchers

Forks

Packages

No packages published