Parapet - purely functional library to develop distributed and event-driven systems
Motivation
It's not a secret that writing distributed systems is a challenging task that can be logically broken into two main aspects: implementing distributed algorithms and running them. Parapet plays the role of execution framework for distributed algorithms - it can be viewed as an intermediate layer between a low-level effect library and high-level operations exposed in the form of DSL. Distributed engineers who mainly focused on designing and implementing distributed algorithms don't need to be worried about low-level abstractions such as IO
or have a piece of deep knowledge in certain computer science subjects, for instance, Concurrency. All they need to know is what properties the library satisfies and what guarantees it provides. On the other hand, engineers who are specializing in writing low-level libraries can concentrate on implementing core abstractions such as IO
or Task
, working on performance optimizations and implementing new features.
Parapet is the modular library where almost any component can be replaced with a custom implementation.
Distributed engineers unite!
NOTE - Github documentation may be out of date. It's preferable to use the official documentation.
Contents
- Key Features
- DSL
- Process
- Channel
- Error Handling and DeadLetterProcess
- EventLog
- Configuration
- Correctness Properties
- Distributed Algorithms in Parapet
- Performance Analysis
- Contribution
Key Features
- Purely functional library written in scala using Tagless-Final Style and Free Monads; thoughtfully designed for people who prefer functional style over imperative
- Modular - almost any component can be replaced with a custom implementation
- DSL provides a set of operations sufficient to write distributed algorithms
- Lightweight and Performant. The library utilizes resources (CPU and Memory) smartly, the code is optimized to reduce CPU consumption when your application in idle state
- Built-in support for the following effect libraries: Cats Effect, Monix, and Scalaz ZIO. The library can be extended to support other effect libraries
Getting started
The first thing you need to do is to add two dependencies into your project: parapet-core
and interop-{effect_library}
for a specific effect library. You can find the latest version in maven central.
libraryDependencies += "io.parapet" %% "core" % "0.0.1-RC1"
For Cats Effect add libraryDependencies += "io.parapet" %% "interop-cats" % "0.0.1-RC1"
For Monix add libraryDependencies += "io.parapet" %% "interop-monix" % "0.0.1-RC1"
For Scalaz ZIO add libraryDependencies += "io.parapet" %% "interop-scalaz-zio" % "0.0.1-RC1"
Once you added the library, you can start writing your first program. However, it's worth taking a few minutes and getting familiar with two main approaches to write processes: generic and effect specific. I'll describe both in a minute. For those who aren't familiar with effect systems like Cats Effect, I'd strongly recommend you to read some articles about IO monad. Fortunately, you don't need to be an expert in Cats Effect to use Parapet.
The first approach we'll consider is Generic. It's recommended to stick to this style when writing processes. Let's develop a simple printer process that will print users requests to the system output.
import io.parapet.core.{Event, Process}
class Printer[F[_]] extends Process[F] {
import Printer._ // import Printer API
import dsl._ // import DSL operations
override def handle: Receive = {
case Print(data) => eval(println(data))
}
}
object Printer {
case class Print(data: Any) extends Event
}
Let's walk through this code. You start writing your processes by extending Process
trait and parameterizing it with an effect type. In this example, we left so-called hole F[_]
in our Printer
type which can be any type constructor with a single argument, e.g. F[_]
is a generic type constructor, cats effect IO
is a specific type constructor and IO[Unit]
is a concrete type. Starting from this moment, it should become clear what it means for a process to be generic. Simply speaking, it means that a process doesn't depend on any specific effect type e.g. IO
. Thus we can claim that our Printer
process is surely generic. The next step is to define a process API or contract that defines a set of events that it can send and receive. Process contract is an important part of any process specification that should be taken seriously. API defines a protocol that other processes will use to communicate with your process. Please remember that it's a very important aspect of any process definition and take it seriously. The next step would be importing DSL
, Parapet DSL is a small set of operations that we will consider in detail in the next chapters. In this example, we need only eval
operator that suspends a side effect in F
, in our Printer process we suspend println
effectful computation. Finally, every process should override handle
function defined in Process
trait. handle
function is a partial function that matches input events and produces an executable flows
. If you ever tried Akka framework you may find this approach familiar (for the curious, Receive
is simply a type alias for PartialFunction[Event, DslF[F, Unit]]
). In our Printer process, we match on Print
event using a well known pattern-matching feature in Scala language. If you are new in functional programming, I'd strongly recommend to read about pattern-matching - it's a very powerful instrument.
That's it. We have considered every important aspect of our Printer
process. Let's move forward and write a simple client process that will talk to our Printer
.
import io.parapet.core.Event.Start
import io.parapet.core.{Process, ProcessRef}
import io.parapet.examples.Printer._ // import Printer API
class PrinterClient[F[_]](printer: ProcessRef) extends Process[F] {
override def handle: Receive = {
// Start is a lifecycle event that gets delivered when a process started
case Start => Print("hello world") ~> printer
}
}
As you already might have noticed, we are repeating the same steps we made when were writing our Printer
process:
- Create a new Process with a hole
F[_]
in its type definition - Extend
io.parapet.core.Process
trait and parametrizing it with generic effect typeF
- Implement
handle
partial function
Let's consider some new types and operators we have used to write our client: ProcessRef
, Start
lifecycle event and ~>
(send) infix operator. Let's start from ProcessRef
. ProcessRef
is a unique process identifier (UUID by default). It represents a process address in Parapet system and must be unique - it's recommended to use ProcessRef
instead of a Process
object directly unless you are sure you want otherwise. It's not prohibited to use Process
object directly, however using a process reference may be useful in some scenarios. Let's consider one such case. Imagine we want to dynamically change the current Printer
process in our client so that it will store data in a file on disk instead of printing it to the console. We can add a new event ChangePrinter
:
case class ChangePrinter(printer: ProcessRef) extends Event
Then our client will look like this:
class PrinterClient[F[_]](private var printer: ProcessRef) extends Process[F] {
import PrinterClient._
import dsl._
override def handle: Receive = {
case Start => Print("hello world") ~> printer
case ChangePrinter(newPrinter) => eval(printer = newPrinter)
}
}
object PrinterClient {
case class ChangePrinter(printer: ProcessRef) extends Event
}
This design cannot be achieved when using direct processes b/c it's not possible to send Process
objects, processes are not serializable in general. One more thing, you can override a Process#ref
field, only make sure it's unique otherwise Parapet system will return an error during the startup.
Ok, we are almost done! There are a few more things left we need to cover: Start
lifecycle event and ~>
operator and there is nothing special about these two. Parapet has two lifecycle events:
io.parapet.core.Event.Start
is sent to a process once it's created in Parapet systemio.parapet.core.Event.Stop
is sent to a process when an application is interrupted withCtrl-C
or when some other process sentStop
orKill
event to that process. The main difference betweenStop
andKill
is that in the former case a process can finish processing all pending events before it will receiveStop
event, whereasKill
will interrupt a process and then deliverStop
event, all pending events will be discarded. If you familiar with JavaExecutorService
then you can think ofStop
asshutdown
andKill
asshutdownNow
.
Finally ~>
is the most frequently used operator that is defined for any type that extends io.parapet.core.Event
trait. ~>
is just a symbolic name for send(event, processRef)
operator.
By this moment we have two processes: Printer
and PrinterClient
, nice! But wait, we need to run them somehow, right?
Fortunately, it's extremely easy to do so, all we need is to create PrinterApp
object which represents our application and extend it from CatsApp
abstract class. CatsApp
extends ParApp by specifying concrete effect type IO
:
abstract class CatsApp extends ParApp[IO]
CatsApp
is provided by the library.
import cats.effect.IO
import io.parapet.CatsApp
import io.parapet.core.Process
object PrinterApp extends CatsApp {
override def processes: IO[Seq[Process[IO]]] = IO {
val printer = new Printer[IO]
val printerClient = new PrinterClient[IO](printer.ref)
Seq(printer, printerClient)
}
}
This is Cats Effect specific application, meaning it uses Cats IO type under the hood. If you run your program you should see hello world
printed to the console. Also notice that we are using concrete effect type IO to fill the hole in our Printer
type, e.g.: new Printer[IO]
in practice it can be any other effect type like Task
, although it requires some extra work in the library.
In our example, we created PrinterClient
which does nothing but sending Print
event at the startup. In my opinion, it doesn't deserve to be a standalone process, would be better if we create a process in place:
object PrinterApp extends CatsApp {
override def processes: IO[Seq[Process[IO]]] = IO {
val printer = new Printer[IO]
val start = Process[IO](_ => {
case Start => Printer.Print("hello world") ~> printer.ref
})
Seq(start, printer)
}
}
Although it's a matter of taste, there is no hard rule.
DSL
This chapter describes each DSL operator in details. Let's get started.
Contents
unit
unit
- semantically this operator is equivalent with Monad.unit
and obeys the same laws. Having said that the following expressions are equivalent:
event ~> process <-> unit ++ event ~> process
event ~> process <-> event ~> process ++ unit
This operator can be used in fold
operator to combine multiple flows. Example:
processes.map(event ~> _).fold(unit)(_ ++ _)
It also can be used to represent an empty flow:
{
case Start => unit // do nothing
case Stop => unit // do nothing
}
flow
flow
- suspends the thunk that produces flow. Semantically this operator is equivalent with suspend
for effects however it's strongly not recommended to perform any side effects within flow
.
Not recommended:
def print(str: String) = flow {
println(str)
unit
}
Recommended:
def print(str: String) = flow {
eval(println(str))
}
flow
may be useful to implement recursive flows. Example:
def times[F[_]](n: Int) = {
def step(remaining: Int): DslF[F, Unit] = flow {
if (remaining == 0) unit
else eval(print(remaining)) ++ step(remaining - 1)
}
step(n)
}
If you try to remove flow
you will get StackOverflowError
Another useful application is using lazy values inside flow
. Example:
lazy val lazyValue: String = {
println("evaluated")
"hello"
}
val useLazyValue = flow {
val tmp = lazyValue + " world"
eval(println(tmp))
}
send
send
- sends an event to one or more receivers. Event will be delivered to all receivers in the specified order.
Parapet provides a symbolic name for this operator ~>
although in the current implementation it doesn't allow to send an event to multiple receivers. It will be added in the future releases.
Examples:
send(Ping, processA, processB, processC)
Ping
event will be sent to the processA
then processB
and finaly processC
. It's not guaranteed that processA
will receive Ping
event before processC
as it depends on it's processing speed and current workload.
Ping ~> processA
Not supported:
Ping ~> Seq(processA, processB, processC)
Possible workaround:
Seq(processA, processB, processC).map(Ping ~> _).fold(unit)(_ ++ _)
Send multiple events to a process:
Seq(e1, e2, e3) ~> process
forward
forward
- sends an event to the receiver using original sender reference. This may be useful for implementing a proxy process.
Example:
val server = Process[IO](_ => {
case Request(body) => withSender(sender => eval(println(s"$sender-$body")))
})
val proxy = Process[IO](_ => {
case Request(body) => forward(Request(s"proxy-$body"), server.ref)
})
val client = Process.builder[IO](_ => {
case Start => Request("ping") ~> proxy
}).ref(ProcessRef("client")).build
The code above will print: client-proxy-ping
par
par
- executes operations from the given flow in parallel. Example:
par(eval(print(1)) ++ eval(print(2)))
Possible outputs: 12 or 21
delay
delay
- delays every operation in the given flow for the given duration.
For sequential flows the flowing expressions are semantically equivalent:
delay(duration, x~>p ++ y~>p) <-> delay(duration, x~>p) ++ delay(duration, y~>p)
delay(duration, x~>p ++ y~>p) <-> delay(duration) ++ x~>p ++ delay(duration) ++ y~>p
For parallel flows:
delay(duration, par(x~>p ++ y~>p)) <-> delay(duration) ++ par(x~>p ++ y~>p)
Note: since the following flow will be executed in parallel the second operation won't be delayed:
par(delay(duration) ++ eval(print(1)))
instead, use:
par(delay(duration, eval(print(1))))
withSender
withSender
- accepts a callback function that takes a sender reference and produces a new flow. Example:
val server = Process[IO](_ => {
case Request(data) => withSender(sender => eval(print(s"$sender says $data")))
})
val client = Process.builder[IO](_ => {
case Start => Request("hello") ~> server
}).ref(ProcessRef("client")).build
The code above will print: client says hello
fork
fork
- does what exactly the name says, executes the given flow concurrently. Example:
val process = Process[IO](_ => {
case Start => fork(eval(print(1))) ++ fork(eval(print(2)))
})
Possible outputs: 12 or 21
register
register
- registers a child process in the Parapet context. It's guaranteed that a child process will receive Stop
event before its parent. Example:
val server = Process[IO](ref => {
case Start => register(ref, Process[IO](_ => {
case Stop => eval(println("stop worker"))
}))
case Stop => eval(println("stop server"))
})
The code above will print:
stop worker
stop server
race
race
- runs two flows concurrently. The loser of the race is canceled.
Example:
val forever = eval(while (true) {})
val process: Process[IO] = Process[IO](_ => {
case Start => race(forever, eval(println("winner")))
})
Output: winner
suspend
suspend
- adds an effect which produces F
to the current flow. Example:
suspend(IO(print("hello world")))
Output: hello world
Not recommended:
suspend {
println("hello world")
IO.unit
}
suspendWith
suspendWith
- suspends an effect which produces F
and then feeds that into a function that takes a normal value and returns a new flow. All operations from produced flow added to the current flow. Example:
suspend(IO.pure(1))) { i => eval(print(i)) }
Output: 1
eval
eval
- suspends a side effect in F
and then adds that to the current flow. Example:
eval(println("hello world"))
Output: hello world
evalWith
evalWith
- Suspends a side effect in F
and then feeds that into a function that takes a normal value and returns a new flow. All operations from a produced flow will be added to the current flow. Example:
evalWith("hello world")(a => eval(println(a)))
Output: hello world
Process
Process
is a key abstraction in Parapet, any application must have a least one process. If you try to run an application w/o processes you will get an error saying that at least one process required. This section covers some useful features that we haven't seen yet, below you will find a shortlist of features:
- Predefined processes and reserved references
- Switching process behavior
- Direct process call
- Process combinators:
and
andor
- Testing your processes
- Basic patterns and tips: implementing timeouts, designing API
Predefined processes and reserved references
Parapet has some reserved process references, e.g.: KernelRef(parapet-kernel)
, SystemRef(parapet-system)
, DeadLetterRef(parapet-deadletter)
, UndefinedRef(parapet-undefined)
. The general rule is that any reference that starts with parapet-
prefix can be used by the platform code for any purpose.
Parapet has a SystemProcess
that cannot be overridden by users. SystemProcess
is a starting point, i.e. it's created before any other process. Lifecycle event Start
is sent by SystemProcess
. Any event sent to the SystemProcess
will be ignored and dropped. Don't try to send any events to SystemProcess
b/c it can lead to unpredictable errors.
DeadLetterProcess
is another process that is created by default, although it can be overridden, for more details check DeadLetterProcess
section under Event Handling
Switching process behavior
Sometimes it might be useful to dynamically switch a process behavior, e.g.: from uninitialized
to ready
state. Thankfully Process
provides switch
method that does exactly that.
Example:
Lazy server:
// for some effect `F[_]`
val server = new Process[F] {
val init = eval(println("acquire resources: create socket and etc."))
def ready: Receive = {
case Request(data) => withSender(Success(data) ~> _)
case Stop => eval(println("release resources: close socket and etc."))
}
def uninitialized: Receive = {
case Start => unit // ignore Start event, wait for Init
case Stop => unit // process is not initialized, do nothing
case Init => init ++ switch(ready)
case _ => withSender(Failure("process is not initialized", ErrorCodes.ProcessUninitialized) ~> _)
}
override def handle: Receive = uninitialized
}
// API
object Init extends Event
case class Request(data: Any) extends Event
sealed trait Response extends Event
case class Success(data: Any) extends Event
case class Failure(data: Any, errorCode: Int) extends Event
object ErrorCodes {
val ProcessUninitialized = 0
}
A client which sends Request
event w/o sending Init
:
val impatientClient = Process[F](_ => {
case Start => Request("PING") ~> server
case Success(_) => eval(println("that is not going to happen"))
case f:Failure => eval(println(f))
})
The code above will print: Failure(process is not initialized,0)
A client which sends Init
first and then Request
:
val humbleClient = Process[F](_ => {
case Start => Seq(Init, Request("PING")) ~> server
case Success(data) => eval(println(s"client receive response from server: $data"))
case _:Failure => eval(println("that is not going to happen"))
})
The code above will print:
acquire resources: create socket and etc.
client receive response from server: PING
release resources: close socket and etc.
switch
is NOT an atomic operation, avoid using switch
in concurrent flows because it may result in an error or lead to unpredictable behavior.
Bad:
val process = new Process[F] {
def ready: Receive = _
override def handle: Receive = {
case Init => fork(switch(ready)) // bad, may lead to unpredictable behaviour
}
}
If you need to switch behavior from a concurrent flow just send an event e.g. Swith(State.Ready)
to itself. Process will eventually switch its behavior:
val process = new Process[F] {
def ready: Receive = _
override def handle: Receive = {
case Init => fork {
eval(println("do some work in parallel"))
Switch(Ready) ~> ref // notify the process that it's time to switch it's behaviour
}
case Switch(Ready) => switch(ready)
}
}
sealed trait State
object Ready extends State
case class Switch(next: State) extends Event
Direct process call
Sometimes it may be useful to call a process directly. Especially it's a common case for short living processes. For instance, you may want to create a process, call it and then abandon, garbage collector will do its job. However, if you try to send an event to a process that doesn't exist in the system you will receive Failure
event with UnknownProcessException
. This is where direct call
comes to rescue.
Example:
// API
case class Sum(a: Int, b: Int) extends Event
case class Result(value: Int) extends Event
class Calculator[F[_]] extends Process[F] {
override def handle: Receive = {
case Sum(a, b) => withSender(Result(a + b - 1) ~> _) // yes, very poor calculator
}
}
val student = Process[F](ref => {
case Start => new Calculator().apply(ref, Sum(2, 2))
case Result(value) => eval(println(s"2 + 2 = $value"))
})
Output: 2 + 2 = 3
Note that apply
method doesn't return a normal value rather it returns a program which will be executed as normal flow.
In other words the following expressions are equivalent:
Sum(2, 2) ~> calculator <-> new Calculator().apply(ref, Sum(2, 2)) // where ref belongs to the same process in both cases
Process combinators
Processes can be combined using two logical operators: or
and and
.
and
- combines two processes by producing a new process with ref
of the first process; combines flows iff 'handle' function is defined for the given event in both processes. Sends an error to the sender if either of two processes isn't defined for the given event.
Example:
import cats.effect.IO
import io.parapet.CatsApp
import io.parapet.core.Event.Start
import io.parapet.core.{Event, Process}
object Example extends CatsApp {
import dsl._
case class Print(data: Any) extends Event
override def processes: IO[Seq[Process[IO]]] =
for {
printerA <- IO.pure(Process[IO](_ => {
case Print(data) => eval(println(s"printerA: $data"))
}))
printerB <- IO.pure(Process[IO](_ => {
case Print(data) => eval(println(s"printerB: $data"))
}))
client <- IO.pure(Process[IO](ref => {
case Start => printerA.and(printerB).apply(ref, Print("test"))
}))
} yield Seq(printerA, printerB, client)
}
If you want to register a combined process then you don't need to register printerA
.
Example:
import cats.effect.IO
import io.parapet.CatsApp
import io.parapet.core.Event.Start
import io.parapet.core.{Event, Process}
object Example extends CatsApp {
import dsl._
case class Print(data: Any) extends Event
override def processes: IO[Seq[Process[IO]]] =
for {
printerA <- IO.pure(Process[IO](_ => {
case Print(data) => eval(println(s"printerA: $data"))
}))
printerB <- IO.pure(Process[IO](_ => {
case Print(data) => eval(println(s"printerB: $data"))
}))
combined <- IO.pure(printerA.and(printerB))
client <- IO.pure(Process[IO](_ => {
case Start => Print("test") ~> combined
}))
} yield Seq(combined, printerB, client)
}
or
- creates a new process with ref
of the first process. A combined process refers to the first process if its handle
is defined for the given event, otherwise, to the second process. Sends an error to the sender if neither process is defined for the given event.
Example:
import cats.effect.IO
import io.parapet.CatsApp
import io.parapet.core.Event.Start
import io.parapet.core.{Event, Process}
object Example extends CatsApp {
import dsl._
case class Print(data: Any) extends Event
override def processes: IO[Seq[Process[IO]]] =
for {
printerA <- IO.pure(Process[IO](_ => {
case Print(data: Int) => eval(println(s"printerA: $data"))
}))
printerB <- IO.pure(Process[IO](_ => {
case Print(data: String) => eval(println(s"printerB: $data"))
}))
combined <- IO.pure(printerA.or(printerB))
client <- IO.pure(Process[IO](_ => {
case Start => Print("test") ~> combined ++ Print(1) ~> combined
}))
} yield Seq(combined, printerB, client)
}
Testing your processes
Integration tests in parapet written in a generic style that we discussed before so that the same tests can be run against any effect system. Let's try to write a simple test for a proxy process. The first thing you need to do is to add test-utils
library into your project:
libraryDependencies += "io.parapet" %% "test-utils" % version
A simple proxy process that receives requests and forwards them to a service
class Proxy(service: ProcessRef) extends Process[F] {
override def handle: Receive = {
case Request(data) => Request(s"proxy-$data") ~> service
}
}
Test for our Proxy
:
import io.parapet.core.{Event, Process, ProcessRef}
import io.parapet.tests.intg.ProxySpec._
import io.parapet.testutils.{EventStore, IntegrationSpec}
import org.scalatest.FunSuite
import org.scalatest.Matchers._
import org.scalatest.OptionValues._
abstract class ProxySpec[F[_]] extends FunSuite with IntegrationSpec[F] {
import dsl._
test("proxy") {
val eventStore = new EventStore[F, Event]
val testService = Process(ref => {
case req: Request => eval(eventStore.add(ref, req))
})
val proxy = new Proxy[F](testService.ref)
val init = onStart(Request("req") ~> proxy)
unsafeRun(eventStore.await(1, createApp(ct.pure(Seq(init, testService, proxy))).run))
eventStore.get(testService.ref).headOption.value shouldBe Request("proxy-req")
}
}
In order to run this test against Cats Effect IO you need to extend BasicCatsIOSpec
:
import cats.effect.IO
import io.parapet.testutils.BasicCatsIOSpec
class ProxySpec extends io.parapet.tests.intg.ProxySpec[IO] with BasicCatsIOSpec
Basic patterns and tips
TODO
Channel
Channel is a process that implements strictly synchronous request-reply dialog. The channel sends an event to a receiver and then waits for a response in one step, i.e. it blocks asynchronously until it receives a response. Doing any other sequence, e.g., sending two request or reply events in a row will return a failure to the sender.
Example for some F[_]
:
val server = new Process[F] {
override def handle: Receive = {
case Request(data) => withSender(sender => Response(s"echo: $data") ~> sender)
}
}
val client = new Process[F] {
lazy val ch = Channel[F]
override def handle: Receive = {
case Start => register(ref, ch) ++
ch.send(Request("PING"), server.ref, {
case scala.util.Success(Response(data)) => eval(println(data))
case scala.util.Failure(err) => eval(println(s"server failed to process request. err: ${err.getMessage}"))
})
}
}
case class Request(data: Any) extends Event
case class Response(data: Any) extends Event
Error Handling and DeadLetterProcess
There are some scenarios when a process may receive a Failure
event:
- When a target process failed to handle an event sent by another process.
Example:
// for some effect F[_]
val faultyServer = Process.builder[F](_ => {
case Request(_) => eval(throw new RuntimeException("server is down"))
}).ref(ProcessRef("server")).build
val client = Process.builder[F](_ => {
case Start => Request("PING") ~> faultyServer
case Failure(Envelope(me, event, receiver), EventHandlingException(errMsg, cause)) => eval {
println(s"self: $me")
println(s"event: $event")
println(s"receiver: $receiver")
println(s"errMsg: $errMsg")
println(s"cause: ${cause.getMessage}")
}
}).ref(ProcessRef("client")).build
The code above will output:
self: client
event: Request(PING)
receiver: server
errMsg: process [name=undefined, ref=server] has failed to handle event: Request(PING)
cause: server is down
EventHandlingException
indicates that a receiver process failed to handle an event.
- When a process event queue is full. It's possible when a process experiencing performance degradation due to heavy load.
Example:
For this example we need to tweak SchedulerConfig:
queueSize = 10000
processQueueSize = 100
// for some effect F[_]
val slowServer = Process.builder[F](_ => {
case Request(_) => eval(while (true) {}) // very slow process...
}).ref(ProcessRef("server")).build
val client = Process.builder[F](_ => {
case Start =>
generateRequests(1000) ~> slowServer
case Failure(Envelope(me, event, receiver), EventDeliveryException(errMsg, cause)) => eval {
println(s"self: $me")
println(s"event: $event")
println(s"receiver: $receiver")
println(s"errMsg: $errMsg")
println(s"cause: ${cause.getMessage}")
println("=====================================================")
}
}).ref(ProcessRef("client")).build
def generateRequests(n: Int): Seq[Event] = {
(0 until n).map(Request)
}
The code above will print a dozens of lines, four lines per Failure
event:
client sent events
self: client
event: Request(101)
receiver: server
errMsg: System failed to deliver an event to process [name=undefined, ref=server]
cause: process [name=undefined, ref=server] event queue is full
=====================================================
self: client
event: Request(102)
receiver: server
errMsg: System failed to deliver an event to process [name=undefined, ref=server]
cause: process [name=undefined, ref=server] event queue is full
=====================================================
self: client
event: Request(103)
receiver: server
errMsg: System failed to deliver an event to process [name=undefined, ref=server]
cause: process [name=undefined, ref=server] event queue is full
=====================================================
EventDeliveryException
indicates that the system failed to deliver an event. Handling such types of errors may be useful for runtime analysis, e.g. a sender process might consider lowering event send rate or even stop sending events to let a target process to finish processing pending events. It's worth noting that you should avoid any long-running computations when processing Failure
events because it could lead to cascading failures.
- A process event handler isn't defined for some events.
Example:
// for some effect F[_]
val uselessService = Process.builder[F](_ => {
case Start => unit
case Stop => unit
}).ref(ProcessRef("server")).build
val client = Process.builder[F](_ => {
case Start =>
Request("PING") ~> uselessService
case Failure(Envelope(me, event, receiver), EventMatchException(errMsg)) => eval {
println(s"self: $me")
println(s"event: $event")
println(s"receiver: $receiver")
println(s"errMsg: $errMsg")
}
}).ref(ProcessRef("client")).build
The code above will print:
self: client
event: Request(PING)
receiver: server
errMsg: process [name=undefined, ref=server] handler is not defined for event: Request(PING)
- A process doesn't exist in Parapet system.
Example:
// for some effect F[_]
val unknownService = Process.builder[F](_ => {
case Start => unit
case Stop => unit
}).ref(ProcessRef("server")).build
val client = Process.builder[F](_ => {
case Start =>
Request("PING") ~> unknownService
case Failure(Envelope(me, event, receiver), UnknownProcessException(errMsg)) => eval {
println(s"self: $me")
println(s"event: $event")
println(s"receiver: $receiver")
println(s"errMsg: $errMsg")
}
}).ref(ProcessRef("client")).build
The code above will print:
self: client
event: Request(PING)
receiver: server
errMsg: there is no such process with id=server registered in the system
Final notes regarding error handling:
- All
Failure
events sent byparapet-system
process (if you are curious you can check it by yourself usingwithSender
). - If a process has no error handling then
Failure
event will be sent toDeadLetterProcess
. More aboutDeadLetterProcess
you will find below
DeadLetterProcess
The library by default provides an implementation of DeadLetterProcess
which just logs failures. Although it might be not very practical, for instance, you may prefer to store failures into a database for further analyses. The library allows providing a custom implementation of DeadLetterProcess
.
Example using CatsApp
:
import cats.effect.IO
import io.parapet.CatsApp
import io.parapet.core.Event.{DeadLetter, Start}
import io.parapet.core.processes.DeadLetterProcess
import io.parapet.core.{Event, Process, ProcessRef}
object CustomDeadLetterProcessDemo extends CatsApp {
import dsl._
override def deadLetter: IO[DeadLetterProcess[IO]] = IO.pure {
new DeadLetterProcess[IO] {
override def handle: Receive = {
// can be stored in database
case DeadLetter(envelope, error) => eval {
println(s"sender: ${envelope.sender}")
println(s"receiver: ${envelope.receiver}")
println(s"event: ${envelope.event}")
println(s"errorType: ${error.getClass.getSimpleName}")
println(s"errorMsg: ${error.getMessage}")
}
}
}
}
val faultyServer = Process.builder[IO](_ => {
case Request(_) => eval(throw new RuntimeException("server is down"))
}).ref(ProcessRef("server")).build
val client = Process.builder[IO](_ => {
case Start => Request("PING") ~> faultyServer
// no error handling
}).ref(ProcessRef("client")).build
override def processes: IO[Seq[Process[IO]]] = IO {
Seq(client, faultyServer)
}
case class Request(data: Any) extends Event
}
The code above will print:
sender: client
receiver: server
event: Request(PING)
errorType: EventHandlingException
errorMsg: process [name=undefined, ref=server] has failed to handle event: Request(PING)
Configuration
Parapet system can be configured by providing an instance of ParConfig
.
Example:
import cats.effect.IO
import io.parapet.core.Parapet.ParConfig
import io.parapet.{CatsApp, core}
object ConfigExample extends CatsApp{
override def processes: IO[Seq[core.Process[IO]]] = _
override val config: ParConfig = ParConfig(...)
}
ParConfig
has the following properties:
- schedulerConfig:
- queueSize - size of event queue shared by workers
- numberOfWorkers - number of workers; default = availableProcessors
- processQueueSize - size of event queue per individual process,
-1
- unbounded
You should set queueSize
to a value that would match the expected workload. For example, if you are going to send 1M events within the same flow it's recommended to set queueSize
to 1M. However, it depends on how fast your consumer processes and amount of available memory, if that's possible to keep some amount of events in memory - go for it, if not - you will probably need to reconsider your design decisions.
In a case the event queue is full all events will be redirected to EventLog
(see the corresponding section).
EventLog
EventLog
can be used to store events on disk. Latter, events can be retrieved and resubmitted.
In a case, the event queue is full unsubmitted events will be redirected to EventLog
. The default implementation just logs such events. In future releases, more practical implementation will be provided.
Correctness Properties
Safty properties:
- It's guaranteed that events will be delivered to a process in a strictly synchronous request-reply dialog, i.e. a process will receive a new event iff it completed processing the current one.
- All events delivered in send order
Liveness properties:
- Sent events eventually delivered
- A sender eventually receives a response
Distributed Algorithms in Parapet
Please refer to components/algorithms subproject
Performance Analysis
Performance mainly dependents on the underlying effect system. In general, there is always some performance and memory overhead associated with the use of Monads and immutable data structures.
Performance test spec:
- 1M requests + 1M responses = 2M events
- CatsApp based
- Number of workers - 12
- Number of processes - 2 (one publisher, one consumer)
- CPU: Intel(R) Core(TM) i7-8750H CPU @ 2.20GHz, 2208 Mhz, 6 Core(s), 12 Logical Processor(s)
- RAM: 32GB
- OS: Windows 10 x64
Total time: 25206 ms
Contribution
The project in its early stage and many things are subject to change. Now is a good time to join! If you want to become a contributor please send me email or text in gitter channel.
If you'd like to donate in order to help with ongoing development and maintenance:
License
Copyright [2019] The Parapet Project Developers
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
∏åRÂπ∑†