Merge pull request #247 from NeowayLabs/addConcurrencyDesign

Concurrency design

proposal/2-concurrency.md

@@ -0,0 +1,297 @@
+# Proposal: Concurrency on Nash
+
+There has been some discussion on how to provide concurrency to nash.
+There is a [discussion here](https://github.com/NeowayLabs/nash/issues/224) 
+on how concurrency could be added as a set of built-in functions.
+
+As we progressed discussing it seemed desirable to have a concurrency
+that enforced no sharing between concurrent functions. It eliminates
+races and forces all communication to happen explicitly, and the
+performance overhead would not be a problem to a high level language
+as nash.
+
+## Lightweight Processes
+
+This idea is inspired on Erlang concurrency model. Since Nash does
+not aspire to do everything that Erlang does (like distributed programming)
+so this is not a copy, we just take some things as inspiration.
+
+Why call this a process ?
+On the [Erlang docs](http://erlang.org/doc/getting_started/conc_prog.html)
+there is a interesting definition of process:
+
+```
+the term "process" is usually used when the threads of execution share no
+data with each other and the term "thread" when they share data in some way.
+Threads of execution in Erlang share no data,
+that is why they are called processes
+```
+
+In this context the process word is used to mean a concurrent thread of
+execution that does not share any data. The only means of communication
+are through message passing. Since these processes are lightweight
+creating a lot of them will be cheap (at least must cheaper than
+OS processes).
+
+Instead of using channel instances in this model you send messages
+to processes (actor model), it works pretty much like a networking
+model using UDP datagrams.
+
+The idea is to leverage this as a syntactic construction of the language
+to make it as explicit and easy as possible to use.
+
+This idea introduces 4 new concepts, 3 built-in functions and one
+new keyword.
+
+The keyword **spawn** is used to spawn a function as a new process.
+The function **send** is used to send messages to a process.
+The function **receive** is used to receive messages from a process.
+The function **self** returns the pid of the process calling it.
+
+An example of a simple ping/pong:
+
+```
+pid <= spawn fn () {
+    ping, senderpid <= receive()
+    echo $ping
+    send($senderpid, "pong")
+}()
+
+send($pid, "ping", self())
+pong <= receive()
+
+echo $pong
+```
+
+Spawned functions can also receive parameters (always deep copies):
+
+```
+pid <= spawn fn (answerpid) {
+    send($answerpid, "pong")
+}(self())
+
+pong <= receive()
+echo $pong
+```
+
+A simple fan-out/fan-in implementation (N jobs <-> N processes):
+
+```
+jobs = ("1" "2" "3" "4" "5")
+
+for job in $jobs {
+    spawn fn (job, answerpid) {
+        import io
+
+        io_println("job[%s] done", $job)
+        send($answerpid, format("result [%s]", $job))
+    }($job, self())
+}
+
+for job in $jobs {
+    result <= receive()
+    echo $result
+}
+```
+
+All output (stdout and stderr) of processes go to their
+parent until the root (main) process, so printing inside
+a child process will print on the stdout of the main process.
+
+### Advanced Fan-out Fan-in
+
+Here is an example of a more elaborated fan-out/fan-in.
+On this case we have much more jobs to execute than
+workers, so it requires more coordination than the previous example.
+
+For brevity this example does not handle timeouts.
+
+Lets suppose an script that tries different passwords on a host:
+
+```
+var passwords_feed <= spawn fn() {
+
+    fn sendpassword(password) {
+        var worker <= receive()
+        if !send($worker, $password) {
+            sendpassword($password)
+        }
+    }
+
+    for password in generate_passwords() {
+        sendpassword($password)
+    }
+}
+
+fn login(output, passwords_feed, done) {
+
+    for send($passwords_feed, self()) {
+        var password = receive()
+        var result <= login "someuser" $password
+        send($output, $result)
+    }
+
+    send($done, "done")
+}
+
+fn outputhandler() {
+    for {
+        var result = receive()
+        if $result == "0" {
+            echo "success"
+        }
+    }
+}
+
+var workers = 10
+
+var feed <= spawn passwords_feed()
+var outputhandler <= spawn outputhandler()
+
+for i in range(0, $workers) {
+    spawn login($outputhandler, $feed, self())
+}
+
+for i in range(0, $workers) {
+    msg <= receive()
+    if $msg != "done" {
+        echo "dafuck ?"
+    }
+}
+```
+
+### Error Handling
+
+Error handling on this concurrency model is very similar to
+how we do it on a distributed system. If a remote service fails and
+just dies and you are using UDP you will never be informed of it,
+the behavior will be to timeout the request and try again (possibly
+to another service instance through a load balancer).
+
+To implement this idea we can add a timeout to the receive an add
+a new parameter, a boolean, indicating if there is a message or if a
+timeout has occurred.
+
+Example:
+
+```
+msg, ok <= receive(timeout)
+if !ok {
+    echo "oops timeout"
+}
+```
+
+The timeout can be omitted if you wish to just wait forever.
+
+For send operations we need to add just one boolean return
+value indicating if the process pid exists and the message
+has been delivered:
+
+```
+if !send($pid, $msg) {
+    echo "oops message cant be sent"
+}
+```
+
+Since the processes are always local there is no need for a more
+detailed error message (the message would always be the same), the
+error will always involve a pid that has no owner (the process never
+existed or already exited).
+
+We could add a more specific error message if we decide that
+the process message queue can get too big and we start to
+drop messages. The error would help to differentiate
+from a dead process or a overloaded process.
+
+An error indicating a overloaded process could help
+to implement back pressure logic (try again later).
+But if we are sticking with local concurrency only this
+may be unnecessary complexity. You can avoid this by
+always sending N messages and waiting for N responses
+before sending more messages.
+
+### TODO
+
+Spawned functions should have access to imported modules ?
+(seems like no, but some usages of this may seem odd)
+
+If send is never blocking, what if process queue gets too big ?
+just go on until memory exhausts ?
+
+Should send be synchronous how we are going to differentiate
+between a timeout or a invalid pid error ? On the other hand
+synchronous send solves the queueing problem.
+
+## Extend rfork
+
+Converging to a no shared state between concurrent functions initiated
+the idea of using the current rfork built-in as a means to express
+concurrency on Nash. This would already be possible today, the idea
+is just to make it even easier, specially the communication between
+different concurrent processes.
+
+This idea enables an even greater amount of isolation between concurrent
+processes since rfork enables different namespaces isolation (besides memory),
+but it has the obvious fallback of not being very lightweight.
+
+Since the idea of nash is to write simple scripts this does not seem
+to be a problem. If it is on the future we can create lightweight concurrent
+processes (green threads) that works orthogonally with rfork.
+
+The prototype for the new rfork would be something like this:
+
+```sh
+chan <= rfork [ns_param1, ns_param2] (chan) {
+        //some code
+}
+```
+
+The code on the rfork block does not have access to the
+lexical outer scope but it receives as a parameter a channel
+instance.
+
+This channel instance can be used by the forked processes and
+by the creator of the process to communicate. We could use built-in functions:
+
+```sh
+chan <= rfork [ns_param1, ns_param2] (chan) {
+        cwrite($chan, "hi")
+}
+
+a <= cread($chan)
+```
+
+Or some syntactic extension:
+
+```sh
+chan <= rfork [ns_param1, ns_param2] (chan) {
+        $chan <- "hi"
+}
+
+a <= <-$chan
+```
+
+Since this channel is meant only to be used to communicate with
+the created process, it will be closed when the process exit:
+
+```sh
+chan <= rfork [ns_param1, ns_param2] (chan) {
+}
+
+# returns empty string when channel is closed
+<-$chan
+```
+
+Fan out and fan in should be pretty trivial:
+
+```sh
+chan1 <= rfork [ns_param1, ns_param2] (chan) {
+}
+
+chan2 <= rfork [ns_param1, ns_param2] (chan) {
+}
+
+# waiting for both to finish
+<-$chan1
+<-$chan2
+```