feat: improve model rate limit tracking
This commit is contained in:
@@ -64,7 +64,7 @@ func (s *Service) execute(ctx context.Context, task store.GatewayTask, user *aut
|
||||
}
|
||||
candidates, err := s.store.ListModelCandidates(ctx, task.Model, modelType, user)
|
||||
if err != nil {
|
||||
failed, finishErr := s.failTask(ctx, task.ID, "no_model_candidate", err.Error(), task.RunMode == "simulation", err)
|
||||
failed, finishErr := s.failTask(ctx, task.ID, store.ModelCandidateErrorCode(err), err.Error(), task.RunMode == "simulation", err)
|
||||
if finishErr != nil {
|
||||
return Result{}, finishErr
|
||||
}
|
||||
@@ -279,7 +279,7 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
|
||||
return clients.Response{}, err
|
||||
}
|
||||
reservations := s.rateLimitReservations(ctx, user, candidate, body)
|
||||
limitResult, err := s.store.ReserveRateLimits(ctx, task.ID, reservations)
|
||||
limitResult, err := s.store.ReserveRateLimits(ctx, task.ID, attemptID, reservations)
|
||||
if err != nil {
|
||||
clientErr := &clients.ClientError{Code: "rate_limit", Message: err.Error(), Retryable: false}
|
||||
_ = s.store.FinishTaskAttempt(ctx, store.FinishTaskAttemptInput{
|
||||
@@ -292,6 +292,12 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
|
||||
})
|
||||
return clients.Response{}, clientErr
|
||||
}
|
||||
rateReservationsFinalized := false
|
||||
defer func() {
|
||||
if !rateReservationsFinalized {
|
||||
_ = s.store.ReleaseRateLimitReservations(context.WithoutCancel(ctx), limitResult.Reservations, "attempt_failed")
|
||||
}
|
||||
}()
|
||||
defer s.store.ReleaseConcurrencyLeases(context.WithoutCancel(ctx), limitResult.LeaseIDs)
|
||||
|
||||
if err := s.store.RecordClientAssignment(ctx, candidate); err != nil {
|
||||
@@ -397,6 +403,10 @@ func (s *Service) runCandidate(ctx context.Context, task store.GatewayTask, user
|
||||
return clients.Response{}, err
|
||||
}
|
||||
}
|
||||
if err := s.store.CommitRateLimitReservations(ctx, limitResult.Reservations, tokenUsageAmounts(response.Usage)); err != nil {
|
||||
return clients.Response{}, err
|
||||
}
|
||||
rateReservationsFinalized = true
|
||||
if err := s.store.FinishTaskAttempt(ctx, store.FinishTaskAttemptInput{
|
||||
AttemptID: attemptID,
|
||||
Status: "succeeded",
|
||||
|
||||
Reference in New Issue
Block a user