github typelevel/cats-effect v3.6.0-RC1
on GitHub

pre-release10 days ago

This is the fifty-third release in the Cats Effect 3.x lineage. It is fully binary compatible with every 3.x release and expected to be fully source-compatible with the eventual 3.6.x lineage.

This release candidate is expected to be very stable, and our main rationale for releasing it as such rather than as a full 3.6.0 is to allow the ecosystem (and adventurous end users!) to try it out and find unforeseen problems prior to full release. If fixing those problems require breaking compatibility prior to 3.6.0 (while maintaining bincompat with 3.x) we will do so, but otherwise you should expect this release to roughly represent the exact API surface area we will ultimately release.

Warning

Please note that Cats Effect 3.6.x is targeting Scala Native 0.4.x, which means it does not at present support native multithreading. This was a very intentional choice meant to give us an opportunity to break binary compatibility only on Scala Native if necessary when upgrading to Scala Native 0.5 and as we discover the impacts of the integrated runtime on the downstream ecosystem. It also reduces risk since Native multithreading in the runtime is a significant lift and we want to make sure we isolate that change from the already significant changes in 3.6.0. This functionality is already under development and we plan to release it in 3.7.0 as soon as possible.

What's Changed

This release contains the largest change to the Cats Effect internal work scheduler since the introduction of work stealing back in the run up to 3.0.0. In particular, we have finally completed implementation of the integrated runtime we first conceived of back in early 2022. Well, I say "we", but really @armanbilge deserves the absolute lion's share of the credit here. He did almost all of the work not only on the direct implementation in Cats Effect, but also the downstream functionality built on top of it in Fs2 and in Http4s… not only on the JVM but also in Scala Native and across multiple operating systems! It is hard to overstate how gargantuan of an effort this was, both in design and implementation.

As immense as the effort has been, the impact is even more dramatic. We've already observed scenarios in which the performance of microservices implemented on top of Http4s Ember on the JVM leveraging io_uring is improved by over 3.5x! That is to say, the performance is 3.5x higher than baseline using classical NIO2! (as measured by TechEmpower, suggesting performance more than 2x higher than pekko-http) Now, TechEmpower does have its limitations as a framework, but it does represent a useful standard candle for certain access patterns, and improving an already very fast trio of frameworks by this magnitude is not something that happens every day. Even more excitingly, this is just the tip of the iceberg! We ran these tests with very early, almost PoC code with a lot of performance-impacting caveats. The best is yet to come.

And if that wasn't enough, the biggest impacts are not on the JVM but actually on Scala Native. Without diving too far into the details, this type of syscall scheduling mechanism is exactly what is necessary on Scala Native to achieve any measure of performance from downstream services. It is actually remarkably similar to what the Go runtime does, though implemented in a more extensible way (e.g. Go does not, at present, have native support for io_uring, whereas Cats Effect can and has implemented this via a third-party library).

Going forward, Cats Effect will be maintaining support for the following polling systems:

  • JVM
    • Selector
  • Native
    • kqueue
    • epoll

Support for io_uring is already under heavy development, and we expect to incorporate this into Cats Effect at some future date, supporting both Native and JVM. The integrated runtime was specifically designed with io_uring's unique data patterns in mind and we have already been able to demonstrate that downstream layers of the ecosystem will be able to fully leverage its benefits. Additionally, we believe we can meaningfully improve JVM performance by implementing epoll and kqueue support for the JVM (and not just Native), the former for situations where io_uring is unavailable. This is mostly because Selector is actually very very slow, and while the performance of the Selector-based runtime is roughly on par with what we've been able to achieve on top of NIO2 (which has its own syscall management and is incompatible with the integrated runtime), we know we can do quite a bit better.

On top of the above, this release contains a large number of other enhancements covering a number of major scenarios. Most noticeably, thanks to @kamilkloch, explicitly importing cats.syntax.all._ is no longer required for most usage involving explicit IO! You will still need the syntax import for any use of parametric effects (commonly referred to as "tagless final" or F[_]), and unfortunately due to the way Scala's type inference works, it is also required for the very common traverse function, but nearly any and all other cases have had their implicit search tweaked to avoid this inconvenience. This in turn means that . completion now mostly Just Works™ when using IO (and when it doesn't, hunting for methods on the IO companion object definitely works).

An additional class of work went into substantially improving the internal instrumentation and metrics associated with the runtime. As previously, these are all exposed as JMX MBeans on the JVM, but we have additionally added public APIs which allow applications to self-instrument (rather than forcing external untyped instrumentation via JMX). These metrics have been organized to mirror the structure of the runtime itself and can all be found within the cats.effect.unsafe.metrics package.

If you're an end-user of Cats Effect (in other words, you run applications built on Cats Effect, rather than building libraries or frameworks), one way you can very much aid the future development of the library is to incorporate these metrics (likely via JMX) into your application observability and key metrics. This should allow you to start to build up an intuition for what metrics are significant indicators of good or bad or just plain normal behavior. This is exactly the type of information that is difficult for library authors to realistically determine, but if end-users learn this on our behalf, you can help educate us on the precise details of how the runtime behaves for your application in your production context, shaping the future optimization and tuning of the framework! (note: Cats Effect does not proactively send metrics anywhere from your application, it merely exposes them to your own monitoring infrastructure)

Enhancements

Bug Fixes

Documentation

Behind the Scenes

  • Configure mergify to add 🤖 label by @armanbilge in #3630
  • Temporarily disable ARM CI, update Docker images by @armanbilge in #3696
  • Cherry-pick Cirrus CI updates to series/3.x by @armanbilge in #3805
  • Remove plugins we get transitively by @armanbilge in #3798
  • Drop macOS jobs from Cirrus by @armanbilge in #3832
  • Fix bad merge by @armanbilge in #3852
  • Add JDK 21 to CI matrix by @armanbilge in #3871
  • Swap Oracle JDK 21 for Temurin JDK 21 by @armanbilge in #3874
  • register->access by @armanbilge in #3867
  • Only run mdoc on Temurin 17 by @armanbilge in #3921
  • Skip starvation test (which is breaking CI) by @djspiewak in #3928
  • Update nscplugin, sbt-scala-native, ... to 0.4.17 in series/3.x by @typelevel-steward in #3965
  • Update sbt-jmh to 0.4.7 in series/3.x by @typelevel-steward in #3913
  • Update specs2-core, specs2-scalacheck to 4.20.5 in series/3.x by @typelevel-steward in #3971
  • Updated macos GHA runners to use the new M series by @djspiewak in #3988
  • Update scalacheck to 1.17.1 in series/3.x by @typelevel-steward in #4053
  • Update cats-core, cats-free, ... to 2.11.0 in series/3.x by @typelevel-steward in #4079
  • Use correct typelevel-nix overlay by @steinybot in #4107
  • Update scalafmt-core to 3.8.3 in series/3.x by @typelevel-steward in #4108
  • register -> accessPoller by @armanbilge in #4117
  • Update scala-library to 2.12.20 in series/3.x by @typelevel-steward in #4129
  • Update sbt-scalafix to 0.13.0 in series/3.x by @typelevel-steward in #4146
  • Update to Scala 2.13.15 and fix Any inference by @mzuehlke in #4148
  • Update scala3-library, ... to 3.3.4 in series/3.x by @typelevel-steward in #4145
  • Update sbt-typelevel to 0.7.4 in series/3.x by @typelevel-steward in #4149
  • update GraalVM from 17 to 21 by @mzuehlke in #4152
  • Update sbt-buildinfo to 0.13.1 in series/3.x by @typelevel-steward in #4165
  • Update cats-mtl, cats-mtl-laws to 1.3.1 in series/3.x by @typelevel-steward in #4182
  • Update sbt-scalajs, scalajs-compiler, ... to 1.17.0 in series/3.x by @typelevel-steward in #4192
  • flake.lock: Update by @typelevel-steward in #4157
  • Update sbt, scripted-plugin to 1.10.6 in series/3.x by @typelevel-steward in #4191
  • Update sbt-mdoc to 2.6.2 in series/3.x by @typelevel-steward in #4198
  • Remove "metrics" that are LocalQueue implementation details by @armanbilge in #4203
  • Refactor WSTP metrics by @armanbilge in #4205
  • "completed" -> "succeeded" in PollerMetrics by @armanbilge in #4210
  • Improve currently-failing tests by @djspiewak in #4211
  • Update sbt, scripted-plugin to 1.10.7 in series/3.x by @typelevel-steward in #4212

Uncategorized

  • Use AtomicReferenceArray to keep track of worker threads by @biochimia in #4196

New Contributors

Full Changelog: v3.5.7...v3.6.0-RC1

Check out latest releases or
releases around typelevel/cats-effect v3.6.0-RC1

Don't miss a new cats-effect release

NewReleases is sending notifications on new releases.

Get notifications