absmach / export

Mainflux Export service that sends messages from one Mainflux cloud to another via MQTT
10 stars 11 forks source link

MF-29 - goroutine in some cases never process channel messages #30

Closed pricelessrabbit closed 4 years ago

pricelessrabbit commented 4 years ago

Fixed #29 : reference type cause the route struct channel to be overwritten by the goroutine that subscribe the routes. The for := range channel reference is changed if not already used and consumer got stuck

Changed the Consume func receiver to value receiver. In this way each worker has owns a copy of route data

Signed-off-by: PricelessRabbit PricelessRabbit@gmail.com

mteodor commented 4 years ago

@PricelessRabbit can you point me where exactly you put the sleep() and it worked you say that route.Messages gets overwritten?

pricelessrabbit commented 4 years ago

@PricelessRabbit can you point me where exactly you put the sleep() and it worked you say that route.Messages gets overwritten?

@mteodor In this point there is a for loop that use local var r https://github.com/mainflux/export/blob/cdbc6f22b1890ae32da86eeccbe962b17ac741de/pkg/export/service.go#L116-L126

but the r.Consume create a pointer reference to r. https://github.com/mainflux/export/blob/cdbc6f22b1890ae32da86eeccbe962b17ac741de/pkg/export/route.go#L86-L88

so then when the loop in the main goroutine continues, it updates r fields with the next route data. But the worker has a pointer to r and not a copy, so all the r fields are updated also in the worker goroutine (also the channel r.Messages).

However, seems (but i'm not shure of that) that if the channel is already filled when the for msg := range line is executed (so there is no "wait" in the first iteration, that the range "keeps" an internal reference of the channel, also if then the channel reference changes, so in some cases (tried with a sleep) the channel is filled before the initialization of the worker and things work as expected also in the bugged implementation

mteodor commented 4 years ago

i think that change should be like https://github.com/mainflux/export/blob/master/pkg/export/service.go#L42 here map[string]*Route and here

func NewRoute(rc config.Route, log logger.Logger, pub messages.Publisher) *Route {
    w := rc.Workers
    if w == 0 {
        w = workers
    }
    r := Route{
        NatsTopic: rc.NatsTopic + "." + NatsAll,
        MqttTopic: rc.MqttTopic,
        Subtopic:  rc.SubTopic,
        Type:      rc.Type,
        Workers:   w,
        Messages:  make(chan *nats.Msg, w),
        logger:    log,
        pub:       pub,
    }
    return &r
}
pricelessrabbit commented 4 years ago

i think that change should be like https://github.com/mainflux/export/blob/master/pkg/export/service.go#L42 here map[string]*Route and here

func NewRoute(rc config.Route, log logger.Logger, pub messages.Publisher) *Route {
  w := rc.Workers
  if w == 0 {
      w = workers
  }
  r := Route{
      NatsTopic: rc.NatsTopic + "." + NatsAll,
      MqttTopic: rc.MqttTopic,
      Subtopic:  rc.SubTopic,
      Type:      rc.Type,
      Workers:   w,
      Messages:  make(chan *nats.Msg, w),
      logger:    log,
      pub:       pub,
  }
  return &r
}

this morning i evaluate and try also in that way. It works, but multiple workers goroutines gets the same route reference, and share the state. This is not a good thing imho because if the route struct change in some manner, all the workers will be affected and this can lead to unexpected behaviours.

mteodor commented 4 years ago

that is the idea, there should be only one instance of each route, and it should not change during the runtime, workers should only process messages and read the route info to know where to send

pricelessrabbit commented 4 years ago

yep it is for that very reason that i think that provide every worker with an immutable value-copy of the route is the most solid solution to avoid possible accidentally modification of data, but if you are ok with the shared state i refactor the fix in that way and update the PR

mteodor commented 4 years ago

@PricelessRabbit yes, please do so, and I'll rethink this what you pointed out and I want to thank you so much