Software Transactional Memory

April 2, 2016

Steve Severance has used Software Transactional Memory (STM) pretty intensively in trading and web crawling applications. He shares things he’s learned, some gotchas, and a comparison of the way STM works in Haskell vs other languages.

STM in Haskell


  • Motivation for STM
  • CPUs are getting more cores yet our tools remain fairly primitive
    • Locks and lock convoys
    • Dealing with deadlocks
  • Immutability has helped Haskell some with this
  • What is STM? It is in-memory transactional concurrency that gives us two properties
    • Atomicity: all results of your transaction will be visible at once
    • Isolation of underlying primitives
  • Haskell is the best language in the world in which to implement an STM
    • It was originally 100 lines of code, and even now it’s only about 600 lines of C!
    • It exploits three properties of the language: purity, immutability and laziness
    • Learning about STM was when Steve first started to appreciate laziness; up until that point laziness had only hurt, never helped
  • Haskell retries an STM transaction as needed. Immutability allows it to do this fearlessly, and laziness avoids fully evaluating the transaction
  • STM is harder in other languages. For instance in Clojure you have to manually implement rollback functionality.
  • There are some helpful auxiliary packages such as stm-chans
  • How about performing unsafe IO in the STM monad? Never! If you’re wanting to do this you’re “doing it wrong.”
  • Discussion of the primitives TVar, TMVar, TChan, TQueue, TBQueue, TSem
  • How the retry function works by putting our thread to sleep
  • Using orElse and the Alternative typeclass
  • How often does a transaction block get re-run? Does it get inefficient? Use the stm-stats package to see how well your transactions are performing.
  • The asyc package internally uses STM and has great integration with it
  • Other languages have STM of varying convenience. Clojure, Scala, C++, C#
    • The C# team spent about three years trying to write an STM. They gave up and wrote a retrospective on why it was so hard.
    • Side effects make STM hellish in languages other than Haskell.
  • Q&A