github.com/sourcegraph/conc

conc: better structured concurrency for go

conc is your toolbelt for structured concurrency in go, making common tasks easier and safer.

go get github.com/sourcegraph/conc

At a glance

• Use conc.WaitGroup if you just want a safer version of sync.WaitGroup
• Use pool.Pool if you want a concurrency-limited task runner
• Use pool.ResultPool if you want a concurrent task runner that collects task results
• Use pool.(Result)?ErrorPool if your tasks are fallible
• Use pool.(Result)?ContextPool if your tasks should be canceled on failure
• Use stream.Stream if you want to process an ordered stream of tasks in parallel with serial callbacks
• Use iter.Map if you want to concurrently map a slice
• Use iter.ForEach if you want to concurrently iterate over a slice
• Use panics.Catcher if you want to catch panics in your own goroutines

All pools are created with pool.New() or pool.NewWithResults[T](), then configured with methods:

• p.WithMaxGoroutines() configures the maximum number of goroutines in the pool
• p.WithErrors() configures the pool to run tasks that return errors
• p.WithContext(ctx) configures the pool to run tasks that should be canceled on first error
• p.WithFirstError() configures error pools to only keep the first returned error rather than an aggregated error
• p.WithCollectErrored() configures result pools to collect results even when the task errored

Goals

The main goals of the package are:

1. Make it harder to leak goroutines
2. Handle panics gracefully
3. Make concurrent code easier to read

Goal #1: Make it harder to leak goroutines

A common pain point when working with goroutines is cleaning them up. It's really easy to fire off a go statement and fail to properly wait for it to complete.

conc takes the opinionated stance that all concurrency should be scoped. That is, goroutines should have an owner and that owner should always ensure that its owned goroutines exit properly.

In conc, the owner of a goroutine is always a conc.WaitGroup. Goroutines are spawned in a WaitGroup with (*WaitGroup).Go(), and (*WaitGroup).Wait() should always be called before the WaitGroup goes out of scope.

In some cases, you might want a spawned goroutine to outlast the scope of the caller. In that case, you could pass a WaitGroup into the spawning function.

func main() {
    var wg conc.WaitGroup
    defer wg.Wait()

    startTheThing(&wg)
}

func startTheThing(wg *conc.WaitGroup) {
    wg.Go(func() { ... })
}

For some more discussion on why scoped concurrency is nice, check out this blog post.

Goal #2: Handle panics gracefully

A frequent problem with goroutines in long-running applications is handling panics. A goroutine spawned without a panic handler will crash the whole process on panic. This is usually undesirable.

However, if you do add a panic handler to a goroutine, what do you do with the panic once you catch it? Some options:

1. Ignore it
2. Log it
3. Turn it into an error and return that to the goroutine spawner
4. Propagate the panic to the goroutine spawner

Ignoring panics is a bad idea since panics usually mean there is actually something wrong and someone should fix it.

Just logging panics isn't great either because then there is no indication to the spawner that something bad happened, and it might just continue on as normal even though your program is in a really bad state.

Both (3) and (4) are reasonable options, but both require the goroutine to have an owner that can actually receive the message that something went wrong. This is generally not true with a goroutine spawned with go, but in the conc package, all goroutines have an owner that must collect the spawned goroutine. In the conc package, any call to Wait() will panic if any of the spawned goroutines panicked. Additionally, it decorates the panic value with a stacktrace from the child goroutine so that you don't lose information about what caused the panic.

Doing this all correctly every time you spawn something with go is not trivial and it requires a lot of boilerplate that makes the important parts of the code more difficult to read, so conc does this for you.

stdlib

conc

type caughtPanicError struct { val any stack []byte } func (e *caughtPanicError) Error() string { return fmt.Sprintf( "panic: %q\n%s", e.val, string(e.stack) ) } func main() { done := make(chan error) go func() { defer func() { if v := recover(); v != nil { done <- &caughtPanicError{ val: v, stack: debug.Stack() } } else { done <- nil } }() doSomethingThatMightPanic() }() err := <-done if err != nil { panic(err) } }

func main() { var wg conc.WaitGroup wg.Go(doSomethingThatMightPanic) // panics with a nice stacktrace wg.Wait() }

Goal #3: Make concurrent code easier to read

Doing concurrency correctly is difficult. Doing it in a way that doesn't obfuscate what the code is actually doing is more difficult. The conc package attempts to make common operations easier by abstracting as much boilerplate complexity as possible.

Want to run a set of concurrent tasks with a bounded set of goroutines? Use pool.New(). Want to process an ordered stream of results concurrently, but still maintain order? Try stream.New(). What about a concurrent map over a slice? Take a peek at iter.Map().

Browse some examples below for some comparisons with doing these by hand.

Examples

Each of these examples forgoes propagating panics for simplicity. To see what kind of complexity that would add, check out the "Goal #2" header above.

Spawn a set of goroutines and waiting for them to finish:

stdlib	conc
func main() { var wg sync.WaitGroup for i := 0; i < 10; i++ { wg.Add(1) go func() { defer wg.Done() // crashes on panic! doSomething() }() } wg.Wait() }	func main() { var wg conc.WaitGroup for i := 0; i < 10; i++ { wg.Go(doSomething) } wg.Wait() }

Process each element of a stream in a static pool of goroutines:

stdlib	conc
func process(stream chan int) { var wg sync.WaitGroup for i := 0; i < 10; i++ { wg.Add(1) go func() { defer wg.Done() for elem := range stream { handle(elem) } }() } wg.Wait() }	func process(stream chan int) { p := pool.New().WithMaxGoroutines(10) for elem := range stream { elem := elem p.Go(func() { handle(elem) }) } p.Wait() }

Process each element of a slice in a static pool of goroutines:

stdlib	conc
func process(values []int) { feeder := make(chan int, 8) var wg sync.WaitGroup for i := 0; i < 10; i++ { wg.Add(1) go func() { defer wg.Done() for elem := range feeder { handle(elem) } }() } for _, value := range values { feeder <- value } close(feeder) wg.Wait() }	func process(values []int) { iter.ForEach(values, handle) }

Concurrently map a slice:

stdlib	conc
func concMap( input []int, f func(int) int, ) []int { res := make([]int, len(input)) var idx atomic.Int64 var wg sync.WaitGroup for i := 0; i < 10; i++ { wg.Add(1) go func() { defer wg.Done() for { i := int(idx.Add(1) - 1) if i >= len(input) { return } res[i] = f(input[i]) } }() } wg.Wait() return res }	func concMap( input []int, f func(*int) int, ) []int { return iter.Map(input, f) }

Process an ordered stream concurrently:

stdlib

conc

func mapStream( in chan int, out chan int, f func(int) int, ) { tasks := make(chan func()) taskResults := make(chan chan int) // Worker goroutines var workerWg sync.WaitGroup for i := 0; i < 10; i++ { workerWg.Add(1) go func() { defer workerWg.Done() for task := range tasks { task() } }() } // Ordered reader goroutines var readerWg sync.WaitGroup readerWg.Add(1) go func() { defer readerWg.Done() for result := range taskResults { item := <-result out <- item } }() // Feed the workers with tasks for elem := range in { resultCh := make(chan int, 1) taskResults <- resultCh tasks <- func() { resultCh <- f(elem) } } // We've exhausted input. // Wait for everything to finish close(tasks) workerWg.Wait() close(taskResults) readerWg.Wait() }

func mapStream( in chan int, out chan int, f func(int) int, ) { s := stream.New().WithMaxGoroutines(10) for elem := range in { elem := elem s.Go(func() stream.Callback { res := f(elem) return func() { out <- res } }) } s.Wait() }

Status

This package is currently pre-1.0. There are likely to be minor breaking changes before a 1.0 release as we stabilize the APIs and tweak defaults. Please open an issue if you have questions, concerns, or requests that you'd like addressed before the 1.0 release. Currently, a 1.0 is targeted for March 2023.