stm

stm

Package stm provides Software Transactional Memory operations for Go. This is an alternative to the standard way of writing concurrent code (channels and mutexes). STM makes it easy to perform arbitrarily complex operations in an atomic fashion. One of its primary advantages over traditional locking is that STM transactions are composable, whereas locking functions are not -- the composition will either deadlock or release the lock between functions (making it non-atomic).

The stm API tries to mimic that of Haskell's Control.Concurrent.STM, but Haskell can enforce at compile time that STM variables are not modified outside the STM monad. This is not possible in Go, so be especially careful when using pointers in your STM code. Remember: modifying a pointer is a side effect!

Unlike Haskell, data in Go is not immutable by default, which means you have to be careful when using STM to manage pointers. If two goroutines have access to the same pointer, it doesn't matter whether they retrieved the pointer atomically; modifying the pointer can still cause a data race. To resolve this, either use immutable data structures, or replace pointers with STM variables. A more concrete example is given below.

It remains to be seen whether this style of concurrency has practical applications in Go. If you find this package useful, please tell us about it!

Examples

See the package examples in the Go package docs for examples of common operations.

See cmd/santa-example/main.go for a more complex example.

Pointers

Note that Operation now returns a value of type any, which isn't included in the examples throughout the documentation yet. See the type signatures for Atomically and Operation.

Be very careful when managing pointers inside transactions! (This includes slices, maps, channels, and captured variables.) Here's why:

p := stm.NewVar[[]byte]([]byte{1,2,3})
stm.Atomically(func(tx *stm.Tx) {
	b := p.Get(tx)
	b[0] = 7
	stm.p.Set(tx, b)
})

This transaction looks innocent enough, but it has a hidden side effect: the modification of b is visible outside the transaction. Instead of modifying pointers directly, prefer to operate on immutable values as much as possible. Following this advice, we can rewrite the transaction to perform a copy:

stm.Atomically(func(tx *stm.Tx) {
	b := p.Get(tx)
	c := make([]byte, len(b))
	copy(c, b)
	c[0] = 7
	p.Set(tx, c)
})

This is less efficient, but it preserves atomicity.

In the same vein, it would be a mistake to do this:

type foo struct {
	i int
}
p := stm.NewVar[*foo](&foo{i: 2})
stm.Atomically(func(tx *stm.Tx) {
	f := p.Get(tx)
	f.i = 7
	stm.p.Set(tx, f)
})

...because setting f.i is a side-effect that escapes the transaction. Here, the correct approach is to move the Var inside the struct:

type foo struct {
	i *stm.Var
}
f := foo{i: stm.NewVar[int](2)}
stm.Atomically(func(tx *stm.Tx) {
	i := f.i.Get(tx)
	i = 7
	f.i.Set(tx, i)
})

Benchmarks

In synthetic benchmarks, STM seems to have a 1-5x performance penalty compared to traditional mutex- or channel-based concurrency. However, note that these benchmarks exhibit a lot of data contention, which is where STM is weakest. For example, in BenchmarkIncrementSTM, each increment transaction retries an average of 2.5 times. Less contentious benchmarks are forthcoming.

BenchmarkAtomicGet-4       	50000000	      26.7 ns/op
BenchmarkAtomicSet-4       	20000000	      65.7 ns/op

BenchmarkIncrementSTM-4    	     500	   2852492 ns/op
BenchmarkIncrementMutex-4  	    2000	    645122 ns/op
BenchmarkIncrementChannel-4	    2000	    986317 ns/op

BenchmarkReadVarSTM-4      	    5000	    268726 ns/op
BenchmarkReadVarMutex-4    	   10000	    248479 ns/op
BenchmarkReadVarChannel-4  	   10000	    240086 ns/op

Credits

Package stm was originally created by lukechampine.