kyo-scheduler-cats

A small bridge that swaps the Cats Effect default IORuntime for one backed by kyo-scheduler. Cats Effect normally executes IO programs on its built-in work-stealing thread pool. This module hands both the compute and the blocking execution contexts to kyo.scheduler.Scheduler, while keeping Cats Effect's own scheduled-task Scheduler (used by IO.sleep and timer-driven combinators) on a dedicated two-thread ScheduledExecutorService.

There is no new API to learn. The reader's first call is either import KyoSchedulerIORuntime.global to bring an implicit runtime into scope for unsafeRun*, or extends KyoSchedulerIOApp to make an IOApp use the kyo-scheduler runtime automatically. The change is observable only as a thread-name shift (threads named kyo-...) and a different scheduling profile underneath unchanged IO code. JVM-only; published for Scala 3 LTS and 2.13.

import cats.effect.IO
import kyo.KyoSchedulerIORuntime.global

val name: String =
    IO.cede.map(_ => Thread.currentThread().getName).unsafeRunSync()
// name starts with "kyo-"

Using kyo-scheduler with Cats Effect

The module exposes one object and one trait. They solve the same problem at two layers: ad-hoc unsafeRun* call sites and full IOApp-shaped applications. Both share a single process-wide kyo-scheduler executor; mixing them in the same process is safe and does not start a second runtime.

KyoSchedulerIORuntime.global

When you already have code that calls unsafeRunSync, unsafeRunAsync, or otherwise picks up an implicit IORuntime, import KyoSchedulerIORuntime.global to redirect that execution onto kyo-scheduler.

import cats.effect.IO
import kyo.KyoSchedulerIORuntime.global

val program: IO[String] =
    IO.cede.map(_ => Thread.currentThread().getName)

val threadName: String = program.unsafeRunSync()
// threadName: "kyo-..."

Caution: global is a lazy val that shadows cats.effect.unsafe.IORuntime.global only when explicitly imported. Without the import, unsafeRun* resolves the stock Cats Effect runtime and your code still runs on io-compute-... threads.

Note: constructing this runtime reads IORuntime.global.config, which forces the stock global runtime to initialize once (and allocate its own pools) as a side effect. The kyo-scheduler runtime then ignores those pools and uses its own executor.

KyoSchedulerIOApp

For applications written against IOApp, replace extends IOApp with extends KyoSchedulerIOApp and nothing else changes. The trait overrides runtime to KyoSchedulerIORuntime.global; run is defined exactly as with stock IOApp.

import cats.effect.ExitCode
import cats.effect.IO
import kyo.KyoSchedulerIOApp

object Main extends KyoSchedulerIOApp:
    def run(args: List[String]): IO[ExitCode] =
        IO.println(Thread.currentThread().getName).as(ExitCode.Success)

Note: KyoSchedulerIOApp overrides only runtime. It does not override runtimeConfig or computeWorkerThreadCount, so any configuration normally exposed on IOApp is inherited from IORuntime.global.config rather than from user overrides on the app subclass.

Choosing between them

Both entry points produce the same runtime; the choice is shape-only. Use KyoSchedulerIOApp when your program has a single main defined as an IOApp. Use import KyoSchedulerIORuntime.global when execution is scattered across libraries, tests, or scripts that summon an IORuntime implicitly at multiple call sites. Mixing both in one process is fine: they share the same lazy val, so the executor is allocated once.

What changes underneath

Three things differ from the stock Cats Effect runtime:

Compute and blocking are the same executor. The IORuntime is built with kyo.scheduler.Scheduler.get.asExecutionContext passed in both the compute and the blocking slot. Cats Effect's normal split (a cached blocking pool distinct from the compute pool) is collapsed, so IO.blocking runs on the same scheduler as IO.cede.

import cats.effect.IO
import kyo.KyoSchedulerIORuntime.global

val blockingThread: String =
    IO.blocking(Thread.currentThread().getName).unsafeRunSync()
// blockingThread: "kyo-..."

Timers run on a fixed 2-thread ScheduledExecutorService. IO.sleep and other timer-driven combinators are dispatched by a dedicated cats.effect.unsafe.Scheduler whose thread pool size is hard-coded to 2 regardless of CPU count. High-frequency IO.sleep workloads share these two threads.

The runtime shutdown hook is a no-op. IORuntime(...) is constructed with () => () as the shutdown action, so calling IORuntime#shutdown() does not stop the kyo-scheduler executor. The scheduler lives for the JVM's lifetime.