Michael Walker

Announce: dejafu-0.1.0.0

Overloadable primitives for testable, potentially non-deterministic, concurrency.
Published on

I will be giving a presentation about Déjà Fu at the Haskell Symposium next week.

Déjà Fu is a library for developing and testing concurrent Haskell programs, it provides a typeclass-abstraction over GHC’s regular concurrency API, allowing the concrete implementation to be swapped out.

GitHackageJenkins

Why do we need this? Well, concurrency is really hard to get right. Empirical studies have found that many real-world concurrency bugs can be exhibited with small test cases using as few as two threads: so it’s not just big concurrent programs that are hard, small ones are too. We as programmers just don’t seem to have a very good intuition for traditional threads-and-shared-memory-style concurrency. The typical approach to testing concurrent programs is to just run them lots of times, but that doesn’t provide any hard coverage guarantees, and then we need to wonder: how many runs do we need?

Fortunately, there has been a lot of research into testing concurrency in the past several years. Systematic concurrency testing is an approach where the source of nondeterminism, the actual scheduler, is swapped out for one under the control of the testing framework. This allows possible schedules to be systematically explored, giving real coverage guarantees for our tests.

This is a library implementing systematic concurrency testing. It provides two typeclasses, MonadConc to abstract over much of Control.Concurrent and related modules, and MonadSTM, to similarly abstract over Control.Monad.STM.

How to use it

If you’re not making use of any IO in your code other than for concurrency, the transition to using MonadConc and MonadSTM will probably just be a textual substitution:

  • IO is replaced with MonadConc m => m
  • STM with MonadSTM m => m
  • *IORef with *CRef
  • *MVar with *CVar
  • *TVar with *CTVar
  • Most other things have the same name, and so can be replaced by just swapping imports around.

If you are using other IO, you will need a gentle sprinkling of MonadIO and liftIO in your code as well.

Is this really just a drop-in replacement for IO/STM?

That’s the idea, yes.

More specifically, the IO instance of MonadConc and the STM instance of MonadSTM just use the regular IO and STM functions, and so should have no noticeable change in behaviour, except for CRef, the IORef equivalent, where modifyCRef behaves like atomicModifyIORef, not modifyIORef1.

  • Haskell Systematic Concurrency Testing: My first approach was to use a Par-like model based on two typeclasses capturing what can be done with shared state. This is alright if all you care about are MVars, but starts to feel a bit forced once you start to add in things like STM and exceptions.

  • Pre-emption Bounding: There are a lot of possible schedules, even for very simple cases. Some method of reducing them must be used in order to complete testing in a sensible amount of time. One simple approach is to just limit the number of pre-emptive context switches in a schedule.

  • Reducing Combinatorial Explosion: Pre-emption bounding works, but still results in a lot of redundant work being done. We can characterise the execution of a concurrent program by the ordering of dependent events, as that is all which can affect the result. This lets us massively reduce the number of schedules to test, and is core of the technique used in this release.


  1. There are some other differences which can lead to incorrect results when testing, but which should not affect code when used in an IO or STM context. Specifically: Control.Monad.Conc.Class.getNumCapabilities can lie to encourage more concurrency when testing; and Control.Exception.catch can catch exceptions from pure code, but Control.Monad.Conc.Class.catch can’t (except for the IO instance).