// setAndStartWorker will set the current worker.
// We always log to both memory logger (for display on the web) and
// console logger (for records / display of command line worker).
func (wi *Instance) setAndStartWorker(wrk Worker, wr *wrangler.Wrangler) (chan struct{}, error) {
wi.currentWorkerMutex.Lock()
defer wi.currentWorkerMutex.Unlock()
if wi.currentWorker != nil {
return nil, fmt.Errorf("A worker is already in progress: %v", wi.currentWorker)
}
wi.currentWorker = wrk
wi.currentMemoryLogger = logutil.NewMemoryLogger()
wi.currentContext, wi.currentCancelFunc = context.WithCancel(wi.backgroundContext)
wi.lastRunError = nil
done := make(chan struct{})
wranglerLogger := wr.Logger()
if wr == wi.wr {
// If it's the default wrangler, do not reuse its logger because it may have been set before.
// Resuing it would result into an endless recursion.
wranglerLogger = logutil.NewConsoleLogger()
}
wr.SetLogger(logutil.NewTeeLogger(wi.currentMemoryLogger, wranglerLogger))
// one go function runs the worker, changes state when done
go func() {
log.Infof("Starting worker...")
var err error
// Catch all panics and always save the execution state at the end.
defer func() {
// The recovery code is a copy of servenv.HandlePanic().
if x := recover(); x != nil {
err = fmt.Errorf("uncaught %v panic: %v", "vtworker", x)
}
wi.currentWorkerMutex.Lock()
wi.currentContext = nil
wi.currentCancelFunc = nil
wi.lastRunError = err
wi.currentWorkerMutex.Unlock()
close(done)
}()
// run will take a long time
err = wrk.Run(wi.currentContext)
}()
return done, nil
}
// setAndStartWorker will set the current worker.
// We always log to both memory logger (for display on the web) and
// console logger (for records / display of command line worker).
func (wi *Instance) setAndStartWorker(wrk Worker, wr *wrangler.Wrangler) (chan struct{}, error) {
wi.currentWorkerMutex.Lock()
defer wi.currentWorkerMutex.Unlock()
if wi.currentContext != nil {
return nil, vterrors.FromError(vtrpcpb.ErrorCode_TRANSIENT_ERROR,
fmt.Errorf("A worker job is already in progress: %v", wi.currentWorker))
}
if wi.currentWorker != nil {
// During the grace period, we answer with a retryable error.
const gracePeriod = 1 * time.Minute
gracePeriodEnd := time.Now().Add(gracePeriod)
if wi.lastRunStopTime.Before(gracePeriodEnd) {
return nil, vterrors.FromError(vtrpcpb.ErrorCode_TRANSIENT_ERROR,
fmt.Errorf("A worker job was recently stopped (%f seconds ago): %v",
time.Now().Sub(wi.lastRunStopTime).Seconds(),
wi.currentWorker))
}
// QUERY_NOT_SERVED = FailedPrecondition => manual resolution required.
return nil, vterrors.FromError(vtrpcpb.ErrorCode_QUERY_NOT_SERVED,
fmt.Errorf("The worker job was stopped %.1f minutes ago, but not reset. You have to reset it manually. Job: %v",
time.Now().Sub(wi.lastRunStopTime).Minutes(),
wi.currentWorker))
}
wi.currentWorker = wrk
wi.currentMemoryLogger = logutil.NewMemoryLogger()
wi.currentContext, wi.currentCancelFunc = context.WithCancel(wi.backgroundContext)
wi.lastRunError = nil
wi.lastRunStopTime = time.Unix(0, 0)
done := make(chan struct{})
wranglerLogger := wr.Logger()
if wr == wi.wr {
// If it's the default wrangler, do not reuse its logger because it may have been set before.
// Resuing it would result into an endless recursion.
wranglerLogger = logutil.NewConsoleLogger()
}
wr.SetLogger(logutil.NewTeeLogger(wi.currentMemoryLogger, wranglerLogger))
// one go function runs the worker, changes state when done
go func() {
log.Infof("Starting worker...")
var err error
// Catch all panics and always save the execution state at the end.
defer func() {
// The recovery code is a copy of servenv.HandlePanic().
if x := recover(); x != nil {
log.Errorf("uncaught vtworker panic: %v\n%s", x, tb.Stack(4))
err = fmt.Errorf("uncaught vtworker panic: %v", x)
}
wi.currentWorkerMutex.Lock()
wi.currentContext = nil
wi.currentCancelFunc = nil
wi.lastRunError = err
wi.lastRunStopTime = time.Now()
wi.currentWorkerMutex.Unlock()
close(done)
}()
// run will take a long time
err = wrk.Run(wi.currentContext)
}()
return done, nil
}
