Skip to content

Commit

Permalink
feat: Add time.Sleep to mitigate race condition. (#1923)
Browse files Browse the repository at this point in the history
The ShuffleQueue scheduler strategy has an infrequent race condition, as
explained by the comment:

```
	// A race condition is possible when the last active table asynchronously
	// queues a relation. The table finishes (calling `.Done()`) a moment
	// before the queue receives the `.Push()`. At this point, the queue is
	// empty and there are no active workers.
	//
	// A moment later, the queue receives the `.Push()` and queues a new task.
	//
	// This is a very infrequent case according to tests, but it happens.
```

After many attempts at a more elegant solution, I finally yielded:

```
time.Sleep(10 * time.Millisecond)
```

Looks ugly, but after running the tests 300 times (so around 3000
syncs), it works 🤷

```
✓ cloudquery/plugin-sdk main* $ go test ./scheduler -count=100 -run TestScheduler             ⏱ 15:04:12
ok  	github.com/cloudquery/plugin-sdk/v4/scheduler	143.523s
✓ cloudquery/plugin-sdk main* $ go test ./scheduler -count=100 -run TestScheduler              ⏱ 15:06:56
ok  	github.com/cloudquery/plugin-sdk/v4/scheduler	142.796s
✓ cloudquery/plugin-sdk main* $ go test ./scheduler -count=100 -run TestScheduler              ⏱ 15:09:22
ok  	github.com/cloudquery/plugin-sdk/v4/scheduler	144.304s
```
  • Loading branch information
marianogappa authored Oct 4, 2024
1 parent d51e172 commit 83dfcad
Showing 1 changed file with 11 additions and 0 deletions.
11 changes: 11 additions & 0 deletions scheduler/queue/active_work_signal.go
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@ package queue
import (
"sync"
"sync/atomic"
"time"
)

// activeWorkSignal is a thread-safe coordinator for awaiting a worker pool
Expand Down Expand Up @@ -58,6 +59,16 @@ func (s *activeWorkSignal) IsIdle() bool {

// Wait blocks until the count of active workers changes.
func (s *activeWorkSignal) Wait() {
// A race condition is possible when the last active table asynchronously
// queues a relation. The table finishes (calling `.Done()`) a moment
// before the queue receives the `.Push()`. At this point, the queue is
// empty and there are no active workers.
//
// A moment later, the queue receives the `.Push()` and queues a new task.
//
// This is a very infrequent case according to tests, but it happens.
time.Sleep(10 * time.Millisecond)

if s.activeWorkUnitCount.Load() <= 0 {
return
}
Expand Down

0 comments on commit 83dfcad

Please sign in to comment.