Skip to main content

Transactional memory (STM)

Software transactional memory, or STM, is an abstraction for concurrent state modification. With STM, one can write code that concurrently accesses state, and that can easily be composed without exposing details of how it ensures safety guarantees. Programs running within an STM transaction will neither deadlock nor have race-conditions. The base building blocks of STM are TVars and the primitives retry, orElse, and catch.

info

The API of arrow-fx-stm is based on Haskell's stm package, and the implementation is based on the GHC implementation for fine-grained locks. For further information, you can read Composable memory transactions by Tim Harris, Simon Marlow, Simon Peyton Jones, and Maurice Herlihy.

Reading and writing concurrent state

Those values that live under the umbrella of STM must be defined as TVars (short for transactional variable). You can think of a TVar<A> as a variable holding values of type A, but where concurrent modification is protected. TVars are not the only transactional data structure (more on that later), but in any case, to modify one, you need to be inside the STM context. This is achieved either by defining our functions with STM as the receiver, or using stm to create lambda functions with such a receiver.

By itself, a function using STM as a receiver does not perform any computations. We say it's just a description of a transaction. Running a transaction is then done using atomically.

The example below shows a banking service moving money from one account to another with STM. Should the first account not have enough money, we throw an exception. This code is guaranteed never to deadlock and to never produce an invalid state by committing after the read state has changed concurrently.

import arrow.fx.stm.atomically
import arrow.fx.stm.TVar
import arrow.fx.stm.STM

fun STM.transfer(from: TVar<Int>, to: TVar<Int>, amount: Int): Unit {
withdraw(from, amount)
deposit(to, amount)
}

fun STM.deposit(acc: TVar<Int>, amount: Int): Unit {
val current = acc.read()
acc.write(current + amount)
// or the shorthand acc.modify { it + amount }
}

fun STM.withdraw(acc: TVar<Int>, amount: Int): Unit {
val current = acc.read()
require(current - amount >= 0) { "Not enough money in the account!" }
acc.write(current - amount)
}

suspend fun example() {
val acc1 = TVar.new(500)
val acc2 = TVar.new(300)
// check initial balances
acc1.unsafeRead() shouldBe 500
acc2.unsafeRead() shouldBe 300
// perform transaction
atomically { transfer(acc1, acc2, 50) }
// check final balances
acc1.unsafeRead() shouldBe 450
acc2.unsafeRead() shouldBe 350
}

One additional guarantee of STM is that the whole transaction is executed atomically. That means we can modify several TVars in one transaction, and we'll never observe an intermediate state.

Other STM data structures

The following types are built upon TVars and provided out of the box with Arrow:

  • TQueue: transactional mutable queue,
  • TMVar: mutable transactional variable that may be empty,
  • TSet, TMap: transactional Set and Map,
  • TArray: array of TVars,
  • TSemaphore: transactional semaphore.
tip

Note that, in most cases, using these data structures is much more efficient than wrapping their "regular" version in a TVar. For example, a TSet<A> performs better than a TVar<Set<A>> because the latter needs to "lock" the entire set on modification, whereas the former knows that only the affected entries need to be taken into account.

Retries

It is sometimes beneficial to manually abort the current transaction if an invalid state has been reached. For example, a TQueue had no elements to read. The aborted transaction will automatically restart once any previously accessed variable has changed.

Here in this example, we've changed withdraw to use retry and thus wait until enough money is in the account, which after a few seconds happens to be the case.

fun STM.transfer(from: TVar<Int>, to: TVar<Int>, amount: Int): Unit {
withdraw(from, amount)
deposit(to, amount)
}

fun STM.deposit(acc: TVar<Int>, amount: Int): Unit {
val current = acc.read()
acc.write(current + amount)
// or the shorthand acc.modify { it + amount }
}

fun STM.withdraw(acc: TVar<Int>, amount: Int): Unit {
val current = acc.read()
if (current - amount >= 0) acc.write(current - amount)
else retry()
// the two lines above could also be written
// check(current - amount >= 0)
// acc.write(it - amount)`
}

suspend fun example() = coroutineScope {
val acc1 = TVar.new(0)
val acc2 = TVar.new(300)
// check initial balances
acc1.unsafeRead() shouldBe 0
acc2.unsafeRead() shouldBe 300
// simulate some time until the money is found
async {
delay(500)
atomically { acc1.write(100_000_000) }
}
// concurrently attempt the transaction
atomically {
transfer(acc1, acc2, 50)
}
// check final balances
acc1.unsafeRead() shouldBe (100_000_000 - 50)
acc2.unsafeRead() shouldBe 350
}

retry can be used to implement a lot of complex transactions, and lies at the heart of the implementation of more complex transactional data structured like TMVar and TQueue.

caution

A transaction that sees an invalid state (which includes reading a TVar that has been changed concurrently) will restart and try again. This usually means we rerun the function entirely. Therefore it is recommended to keep transactions small and never to use code that has side effects inside.

  • Transactions may be aborted at any time, so accessing resources may never trigger finalizers,
  • Transactions may rerun an arbitrary amount of times before finishing, and thus all effects will rerun.

Branching

The counterpart to retry is orElse, which allows detecting if a branch has called retry and then use a fallback instead. If the fallback retries as well, then the whole transaction retries.

In the example below, we use orElse to return null whenever the check fails. By default, check retries the transaction, hence the need for orElse.

fun STM.transaction(v: TVar<Int>): Int? =
stm {
val result = v.read()
check(result in 0 .. 10)
result
} orElse { null }

suspend fun example() {
val v = TVar.new(100)
// check initial balance
v.unsafeRead() shouldBe 100
// value is outside the range, should fail
atomically { transaction(v) } shouldBe null
// value is outside the range, should succeed
atomically { v.write(5) }
atomically { transaction(v) } shouldBe 5
}

Exceptions

Throwing inside STM will let the exception bubble up to either a catch handler or to atomically, which will rethrow it. Note that this catch does not refer to the built-in exception one, but rather to the function of the same name in the STM module.

danger

Using try {...} catch (e: Exception) {...} is not encouraged because any state change inside try will not be undone when an exception occurs. The recommended way of catching exceptions is to use the catch function, which properly rolls back the transaction.