Kotlin/Native for AWS Lambda

Amazon announced Lambda Runtime API on AWS re:Invent 2018. It allows developers, among other things, to build Lambda functions using any technology they want via so-called Custom Runtimes. Yes, it’s now possible to author a function on PHP, Perl, Pascal (anybody?) or even Bash (they use it in the docs)!

Nice, isn’t it? Let’s build an AWS Lambda function with Kotlin/Native!

But first, let’s figure out what should be done. How do runtimes work?

A runtime’s job is to:

  • Execute the function’s initialization logic. In the case of Java, it means starting the JVM, loading the classes and running static initializers.

  • Locate the handler passed through the "Handler" configuration parameter. For Java-based Lambdas, it’s either an FCDN of a class, like some.package.Handler, or a reference to a method, like some.package.Handler::method.

  • Execute the handler for each incoming event.

Here is a picture to help you grasp a function’s lifecycle:

Basically, when you author a Lambda function using one of the supported runtimes, like Java, Nodejs or Go, you are concentrated on the event processing loop in the center. The runtime handles the initialization and passes the events directly to your handler in the form of objects or structs (the naming depends on the programming language).

In the case of a custom runtime, it’s all your job.

A custom runtime is just an executable file named bootstrap in your function’s deployment package that is used as an entry point. The file can be included in your the deployment package directly, or in a layer. AWS Lambda executes it with the configuration passed via environment variables. The bootstrap should initialize the required resources and enter the event processing loop. AWS Lambda provides an HTTP API for custom runtimes to receive the events and send the responses back. Your custom runtime should call this API in a loop and fetch the events. For each event, it could either invoke a handler or processes it on its own.

Let’s take a look at this Bash-based Lambda function that responds with an Amazon API Gateway Proxy Integration response that sends HTTP redirects to this blog for all the requests.

bootstrap
#!/bin/sh

set -euo pipefail

while true
do
    HEADERS="$(mktemp)" # (1)

    EVENT_DATA=$(curl -v -sS -LD "$HEADERS" -X GET "http://${AWS_LAMBDA_RUNTIME_API}/2018-06-01/runtime/invocation/next") # (2)

    INVOCATION_ID=$(grep -Fi Lambda-Runtime-Aws-Request-Id "$HEADERS" | tr -d '[:space:]' | cut -d: -f2) # (3)

    echo $EVENT_DATA # (4)

    curl -v -sS -X POST "http://${AWS_LAMBDA_RUNTIME_API}/2018-06-01/runtime/invocation/$INVOCATION_ID/response" -H "Content-Type: application/json" -d '{"statusCode": 307, "headers": {"Location": "https://madhead.me"}}' # (5)
done
  1. Create a file to capture the headers.

  2. Receive next event and dump it’s headers to a temporary file created on the previous line.

  3. Parse the headers to find the request id.

  4. Log the event to STDOUT.

  5. Respond to the event using the request id.

There is no initialization or clean-up here and there is no external handler: it’s a very basic function and the events are processed right in bootstrap script. Note the while true loop that polls for events: once this function is started by AWS Lambda it will stay alive and process the requests until Lambda decides to recycle it. Also note the way how the function interacts with the AWS Lambda: by calling HTTP API with curl.

Now, when you know the basics of custom runtimes, let’s implement echo Lambda function in Kotlin/Native.

First, set up the project. Latest IntelliJ IDEA supports Kotlin/Native… natively, so it can be done with "File" → "New" → "Project…" → "Kotlin" → "Kotlin/Native":

IDEA will create a Gradle project with Kotlin Multiplatform plugin and native target.

Let’s add two dependencies: Ktor and kotlinx.serialization. We’ll use first as an HTTP client and second for JSON encoding.

In order for Kotlin/Native dependencies to work properly in Gradle, you need to enable GRADLE_METADATA feature. Furthermore, kotlinx.serialization’s plugin is not published on Gradle Plugin Portal (yet), so let’s use some black magic to tell it how to find the plugin. Both tweaks should go to the settings.gradle.kts, so here is it:

settings.gradle.kts
import org.gradle.api.internal.FeaturePreviews

rootProject.name = "kotlin-native-lambda"

enableFeaturePreview(FeaturePreviews.Feature.GRADLE_METADATA.name)

pluginManagement {
    repositories {
        gradlePluginPortal()
        add(jcenter())
    }

    resolutionStrategy {
        eachPlugin {
            if (requested.id.id == "kotlinx-serialization") {
                useModule("org.jetbrains.kotlin:kotlin-serialization:${requested.version}")
            }
        }
    }
}

Now Gradle "knows" kotlinx-serialization plugin and we can go and configure the build:

build.gradle.kts
plugins {
    kotlin("multiplatform").version("1.3.21")
    id("kotlinx-serialization").version("1.3.21") // (1)
}

repositories {
    jcenter()
    maven("https://kotlin.bintray.com/kotlinx") // (2)
}

kotlin { // (3)
    linuxX64("bootstrap") { // (4)
        binaries {
            executable("bootstrap") { // (5)
                entryPoint = "by.dev.madhead.kotlin_native_lambda.main" // (6)
            }
        }
    }
}

dependencies {
    val bootstrapMainImplementation by configurations // (7)

    bootstrapMainImplementation("io.ktor:ktor-client-curl-linuxx64:1.1.3") // (8)
    bootstrapMainImplementation("io.ktor:ktor-client-json-linuxx64:1.1.3") // (8)
}

tasks {
    wrapper {
        gradleVersion = "5.3"
        distributionType = Wrapper.DistributionType.ALL
    }
}
  1. Use org.jetbrains.kotlin:kotlin-serialization:1.3.21 as Gradle plugin. The resolution rule comes from pluginManagement.resolutionStrategy in settings.gradle.kts configured previously. One day, the plugin will be published officially and these lines could be removed. I recommend applying plugins via plugins block versus apply so that Kotlin extension functions are made available to configure them.

  2. kotlinx.serialization is not published in JCenter (yet), so we need that additional repository.

  3. Kotlin Multiplatform Project configuration block.

  4. We configure single Linux x64 binary as it is the platform used by AWS Lambda. Other supported platforms include Android ARM, iOS ARM and x64, other Linux variants, Windows, Mac OS and Web Assembly.

  5. As long as we’re building an executable file, we need to say that to Gradle. Other options are dynamic and static libraries and Objective-C frameworks.

  6. As with good old Gradle’s application plugin, we need to specify the entry point.

  7. Just a trick to tell Gradle about the existing configuration and avoid using string literals within the dependencies block.

  8. Finally, the libraries we need: Ktor HTTP client based on cURL and Ktor JSON facilities.

Almost done. We’ll be using two data classes to model input and output of a Lambda function for AWS Lambda Proxy Integration:

ProxyIntegrationRequest
@kotlinx.serialization.Serializable
data class ProxyIntegrationRequest(
        val resource: String,
        val path: String,
        val httpMethod: String,
        val headers: Map<String, String>?,
        val multiValueHeaders: Map<String, List<String>>?,
        val queryStringParameters: Map<String, String>?,
        val multiValueQueryStringParameters: Map<String, List<String>>?,
        val pathParameters: Map<String, String>?,
        val stageVariables: Map<String, String>?,
        val body: String?,
        val isBase64Encoded: Boolean?
)
ProxyIntegrationResponse
@kotlinx.serialization.Serializable
data class ProxyIntegrationResponse(
        val isBase64Encoded: Boolean = false,
        val statusCode: Int,
        val headers: Map<String, String>? = null,
        val multiValueHeaders: Map<String, List<String>>? = null,
        val body: String? = null
)

Finally, the bootstrap’s code:

main.kt
fun main() = runBlocking { // (1)
    val client = HttpClient(Curl) // (2)

    while (true) { // (3)
        val invocation = client.call("http://${getenv("AWS_LAMBDA_RUNTIME_API")!!.toKString()}/2018-06-01/runtime/invocation/next") {
            method = HttpMethod.Get
        } // (4)

        // (5)
        val invocationId = invocation.response.headers["Lambda-Runtime-Aws-Request-Id"]
        val payload = invocation.response.content.readRemaining().readText(Charsets.UTF_8)
        val proxyIntegrationRequest =
                try {
                    Json.nonstrict.parse(ProxyIntegrationRequest.serializer(), payload)
                } catch (e: Exception) {
                    // (6)
                }

        println(proxyIntegrationRequest) // (7)

        // (8)
        client.call("http://${getenv("AWS_LAMBDA_RUNTIME_API")!!.toKString()}/2018-06-01/runtime/invocation/$invocationId/response") {
            method = HttpMethod.Post
            body = TextContent(
                    Json.nonstrict.stringify(
                            ProxyIntegrationResponse.serializer(),
                            ProxyIntegrationResponse(
                                    statusCode = 200,
                                    headers = mapOf(
                                            "Content-Type" to "text/plain"
                                    ),
                                    body = proxyIntegrationRequest.toString()
                            )
                    ),
                    ContentType.Application.Json
            )
        }
    }
}
  1. As long as we’ll be using coroutines thanks to Ktor, we need a coroutine scope. The simplest way to acquire one is runBlocking.

  2. Configure the HTTP client using the Curl engine. This is the initialization phase from the picture at the beginning of this post.

  3. Enter the event loop.

  4. Fetch next event to process.

  5. Parse the event. Feel’s better than grep, isn’t it?

  6. Never swallow exceptions like that.

  7. Log the request. Function’s STDOUT is redirected to AWS CloudWatch where you’ll be able to find the logs.

  8. Echo the request to the response.

That’s it!

Take a look at the full source on my GitLab. It even has a CI/CD configuration where you can see how to build, pack and deploy functions with custom runtimes!

Questions? Ask in comments below!

But I’ll probably answer one right now. The performance. Is it worth it or not?

I have tested multiple different runtimes in a simple task of sending HTTP redirects (just like a Bash function above). Take a look at the results:

I know it’s a mess but the results are:

  1. Kotlin/Native and Bash are not very performant. It’s still better that cold-started JVM functions, though (who literally sucks).

  2. Golang is probably the best choice for Lambda functions as it provides the fastest cold starts and the best performance in general.

  3. Pre-warmed JVM functions perform very well (close to Golang).

  4. Script languages provide good cold start times but pre-warmed scripted functions concede to Golang and JVM.

Thanks for reading to the end! Have fun with Kotlin!

comments powered by Disqus