Hi Jim, Thanks for sharing. Good luck with hunting down the issue. I've added some comments below.
B/ On 8 December 2014 at 22:23:29, Jim Washburn ([email protected]) wrote: Hi again, I ran more experiments, and now have seen a scenario where we execute the XXX code path, yet don't have a problem with node to node remoting. So my hypothesis regarding this code path being correlated with our issue, must have been incorrect. Sorry for the false alarm. Back to square one... Jim On Mon, Dec 8, 2014 at 9:18 AM, Jim Washburn <[email protected]> wrote: Hi Björn, Thanks for getting back to us! On Mon, Dec 8, 2014 at 6:11 AM, Björn Antonsson <[email protected]> wrote: Hi Jim, The UID in question is the UID of the remote actor system that is trying to connect in to this actor system. It will only be changed on a complete restart of the actor system. Just to make sure we are using the same terminology, referring to the "Actor LifeCycle" diagram in http://doc.akka.io/docs/akka/2.3.7/java/untyped-actors.html - a restart of an actor, creating a new "instance", will *not* change the UID - a Stop or PoisonPill of an actor, followed by an ActorOf(), resulting in a new actor "incarnation", *will* change the UID The UID in the code in the remoting is the UID of the ActorSystem, not the UID of a single ActorRef. The ActorSystem UID is chaged on restart of the ActorSystem. Those two scenarios are separate in my mind from a restart of the entire actor system. Is this is accord with your view? If so, in order to avoid the XXX code path, by making the UID equality false, we could consider creating a new incarnation of the actor on the remote system. Since this happens to be an important actor in our app (Paxos leader), we would have to consider all the repercussions of doing that. The other alternative in terms of making the UID equality false would be to avoid having the local actor system having the remote system's UID stored in its endpoint database. At this point I do not know enough about the remoting to understand why it has this UID stored in the first place, given that the whole JVM was restarted. Which branch did you build exactly? The latest stable bugfix branch is release-2.3. >From github, for akka, when I do a 'git log', the version I have is commit 1312ca61396d2f4e4bb38318bca333e6ec6b62d8 Merge: a570872 6e6f92f Author: Martynas Mickevičius <[email protected]> Date: Fri Nov 14 10:22:14 2014 +0200 from AkkaBuild.scala, version := "2.3-SNAPSHOT", Hope that works as a means of letting you know what I am running. When I did a git clone from github, that is what I got. If you do "git checkout release-2.3" you will get the latest stable bugfix branch. Also which of the machines in you failure scenario is running the XXX code path? OK for our app, we run currently run three nodes on three JVMS. For this discussion, it comes down to an interaction between two nodes. One node we can call the Proposer, which stays running running in our case. The other node, which we can call the Acceptor , is stopped and restarted at the JVM level. That is the machine which running the XXX code path. Regards, Jim B/ On 6 December 2014 at 08:54:29, Jim Washburn ([email protected]) wrote: Hi Martynas, I discussed this with Robert and Dragisa and took a look into it. I cloned the Akka repo (2.3-SNAPSHOT) and put some debug logs into the akka remoting project. Not to imply that there is a bug in akka remoting, but to get a more precise idea of what is happening in terms of the interaction between our application and akka. I found a difference in a certain code path taken in Remoting.scala, between the cases of when we have a problem in restarting our app, and when we don't have a problem. (The restart problem occurs intermittently for us.) The function is def handleInboundAssociation(ia: InboundAssociation): Unit = ia match { case ia @ InboundAssociation(handle: AkkaProtocolHandle) ⇒ endpoints.readOnlyEndpointFor(handle.remoteAddress) match { case Some(endpoint) ⇒ pendingReadHandoffs.get(endpoint) foreach (_.disassociate()) pendingReadHandoffs += endpoint -> handle endpoint ! EndpointWriter.TakeOver(handle, self) case None ⇒ if (endpoints.isQuarantined(handle.remoteAddress, handle.handshakeInfo.uid)) { handle.disassociate(AssociationHandle.Quarantined) } else endpoints.writableEndpointWithPolicyFor(handle.remoteAddress) match { case Some(Pass(ep, None, _)) ⇒ stashedInbound += ep -> (stashedInbound.getOrElse(ep, Vector.empty) :+ ia) case Some(Pass(ep, Some(uid), _)) ⇒ if (handle.handshakeInfo.uid == uid) { pendingReadHandoffs.get(ep) foreach (_.disassociate()) // XXX we have a restart problem pendingReadHandoffs += ep -> handle ep ! EndpointWriter.StopReading(ep, self) } else { context.stop(ep) // YYY we don't have a restart problem endpoints.unregisterEndpoint(ep) pendingReadHandoffs -= ep createAndRegisterEndpoint(handle, refuseUid = Some(uid)) } case state ⇒ createAndRegisterEndpoint(handle, refuseUid = endpoints.refuseUid(handle.remoteAddress)) } } } As shown above, akka executes the code path labelled XXX above , we have a problem with our nodes communicating after a restart. That is the case when the UIDs being compared are equal. In the other case, labelled YYY, when then UIDs are unequal, a new Endpoint is created. It appears that when the UIDs are equal, akka thinks that the endpoint is already set up. I understand that the UIDs get recreated when actors are restarted. And therefore we should use the actorSelection API instead of actorFor. We have gone ahead and replaced actorFor with actorSelection in our app. Still , unfortunately the problem persists. My hypothesis is that there is still some so-to-speak "stale" UID information somewhere in the system. I can understand why an actor on one JVM node (the one which was not restarted) would have the same UID as before. I don't know however how the other node (the one which is restarted) would have that UID in its EndpointRegistry map, and therefore lead to a uid equality. Any other ideas about how to get around this, would be much appreciated. Regards, Jim Washburn On Wednesday, November 12, 2014 4:29:59 PM UTC-8, Robert Preissl wrote: Hi Martynas, Yes, I also tried manipulating the "transport-failure-detector" values (basically multiplied the default interval & pause settings by a factor of 2 and also by 4 and it did not change anything). I am attaching the original server logs for all three nodes. this is the scenario where node1 and node2 are restarting. node3 stays up. and once node1 and node2 come back up, only node2 successfully syncs up with node3. node1 sends messages to node3 but never gets responses back. I appreciate you taking a look and curious if you see something suspicious. Thanks, Robert On Tuesday, November 11, 2014 5:36:40 AM UTC-8, Martynas Mickevičius wrote: Hi all, Robert, you mentioned that you have already tried to change "heartbeat-interval" setting. Did you change that on "watch-failure-detector" or on "transport-failure-detector"? Could you try changing that on "transport-failure-detector"? If you can still reproduce it can you provide either a reproducible code sample or logs from both of the systems when messages can only propagate to one direction? I have tried various situations while restarting one akka node with some decent load and I can't reproduce it. On Fri, Nov 7, 2014 at 10:00 PM, Dragisa Krsmanovic <[email protected]> wrote: Martynas & Robert, To me, the most suspicious thing is that, in this case, connection only works one way. A can talk to B but B can’t reply back to A. There is not that much custom code in that class. It’s not subscribed to Association/Disassociation or any other Akka system event. Actor one node is sending message to actor (ActorRef/ActorSelection) on another node. Receiver clearly receives the messages and replies with "sender ! msg”. We can see that from our logs. But the sender does not receive the message because the link in that direction is disassociated. This is just plain akka-remote. Not clustering. It seems like connection health checks that you added in 2.2 are causing us trouble with false negatives. What are configuration options that we can try ? Can we assign a different dispatcher to heartbeat failure detector to rule out thread starvation ? Dragisa Krsmanovic Ticketfly — Sent from Mailbox On Friday, Nov 7, 2014 at 9:47 AM, Martynas Mickevičius <[email protected]>, wrote: Hello Martynas, This test is already the simplest scenario we can come up with. It comes from a cluster simulation test framework we have developed to simulate our business needs. If we find the time we can write a simple ping-pong test. but not sure if this is possible. is there any more logging we can try? or changing parameters? (however, i already tried changing the "heartbeat-interval", etc.) Are there any plans to make this more robust in Akka 2.3.7? i fear we need to revert back to older Akka versions. and, can it be that we experience similar issues as reported in this ticket? https://github.com/akka/akka/issues/13860 Thanks, Robert On Friday, November 7, 2014 9:47:42 AM UTC-8, Martynas Mickevičius wrote: I think these messages are fine. After node{1,2} comes back on node3 should associate with new nodes. >From the logs I see that there is quite a lot of custom code running (such as >diva.core.engine.PaxosDistributedKeyManager) which is listening for >Association/Disassociation Events. Have you tried the restart scenario with >some load with simplest actors possible and see if you can reproduce the issue? On Fri, Nov 7, 2014 at 7:37 PM, Robert Preissl <[email protected]> wrote: Hello Martynas, Well, I think I can rule this option out because: - without any load on the system (my scenario 1 in my orig. post) a restart works fine. - also, most of the time node3 can send back messages to node2. but node3 does not send to node1. (however, sometimes both nodes, node1 and node2 do not hear back from node3) - and we also tried with ActorSelection and it did not work. is it suspicious to see Disassociated messages? or is this just a symptom? Thanks, Robert On Friday, November 7, 2014 9:15:42 AM UTC-8, Martynas Mickevičius wrote: Hi Robert, as you mentioned and from the logs you provided its seems that messages are flowing from node{1,2} to node3 after restart, but not to the other direction. Would it be possible that your application tries to send messages from node3 to node{1,2} using ActorRefs which were resolved before the restart of node{1,2}? ActorRef includes actor UID which changes after Actor is stopped and started again, which happens upon node restart. Here is a quick example code that illustrates that situation. If so, you should send messages using ActorSelection or re-resolve ActorRefs after node restart or periodically. On Fri, Nov 7, 2014 at 3:58 AM, Robert Preissl <[email protected]> wrote: Hello Endre, First of all, thanks for replying so quickly! Second, I need to mention that we use Akka remoting. and not Akka clustering (yet). not sure if this makes a difference. What I mean that in our restart scenario (where first node1 and node2 are simultaneously restarted. and then node3) when node1 and node2 are coming back up, it seems that the connection node1 -> node3 works fine. but the connection node3 -> node1 does not. So, to answer your question, yes, it seems we loose messages from node3. I attached more detail logs below. (and please excuse the many log lines; i tried to clean it up as much as possible) what is interesting to see is this line: processing Event(Disassociated [akka.tcp://[email protected]:8900] -> [akka.tcp://[email protected]:8900] 10.57.0.43 is node3. and 10.57.0.41 is node1, by the way. so, the connection between node3 and node1 is Disassociated; which explains maybe why node1 never hears back from node3 when it tries to sync up. We looked a bit in the akka source code and found that stopping an EndpointWriter (I think) triggers a "Disassociated" to be fired, right? and we can see this stop in a few log lines above: [akka://DivaPCluster/system/endpointManager/reliableEndpointWriter-akka.tcp%3A%2F%2FDivaPCluster%4010.57.0.41%3A8900-0/endpointWriter] akka.remote.EndpointWriter - stopping so, why is it stopping? is this our problem here? the logs from node3 are attached as a file. Thank you! Robert On Wednesday, November 5, 2014 12:55:04 PM UTC-8, Robert Preissl wrote: hello! I am having a problem in my remote Akka production system, which consists of 3 nodes running with the latest version of Akka (2.3.6.): In more details, I am experiencing errors with "rolling restarts" of the cluster (for deployment, etc. we cannot afford any downtime), where a restart happens in the following sequence 1.) restart node1 and node2. 2.) once 1. completed, restart node3. but we only observe failures once there is a load (even small load) on the system. So, I want to describe two scenarios: Scenario 1 - no load on the system: Restart works. if there is no load on the system at all, the restarting seems to work fine. I.e., with detailed logging I can observe that node3 logs the following events: (in chronological order) 13:09:48.769 WARN [akka.tcp://DivaPCluster@NODE_3:8900/system/endpointManager/reliableEndpointWriter-akka.tcp0-1] akka.remote.ReliableDeliverySupervisor - Association with remote system [akka.tcp://DivaPCluster@NODE_2:8900] has failed, address is now gated for [5000] ms. Reason is: [Disassociated]. 13:09:48.823 WARN [akka.tcp://DivaPCluster@NODE_3:8900/system/endpointManager/reliableEndpointWriter-akka.tcp0-0] akka.remote.ReliableDeliverySupervisor - Association with remote system [akka.tcp://DivaPCluster@NODE_1:8900] has failed, address is now gated for [5000] ms. Reason is: [Disassociated]. 13:10:10.661 DEBUG [Remoting] Remoting - Associated [akka.tcp://DivaPCluster@NODE_3:8900] <- [akka.tcp://DivaPCluster@NODE_2:8900] 13:10:10.987 DEBUG [Remoting] Remoting - Associated [akka.tcp://DivaPCluster@NODE_3:8900] <- [akka.tcp://DivaPCluster@NODE_1:8900] Since node1 and node2 restart, it is fine that the association is gated between node3 -> node1 (and between node3 -> node2) for a while. And I assume it becomes active again since "a successful inbound connection is accepted from a remote system during Gate it automatically transitions to Active" (as you describe in http://doc.akka.io/docs/akka/snapshot/java/remoting.html) this can be verified since I can see the logs on node1 that it tries to connect at this point in time after the restart: 13:10:10.861 (and the connection becomes active on node3; managing node3 -> node1; at time 13:10:10.987 as you can see above) so, everything cool here and the system restarts fine! Scenario 2 - easy load on the system: Restart fails due to Unrecoverable "gated" state Similar to Scenario 1 above, I can observe the "gated" messages for links node3 -> node1 and node3 -> node2. However, I never see that the links become active again! and the restart never recovers and I need to manually stop my nodes and start up again. This is surprising since I clearly see that node1 and node2 (after they restarted) send message to node3. and node3 successfully logs the reception of these messages. So, why does in this scenario the connection not become active again?? It is a successful inbound connection that should make the link active again as you describe on your site? Any help on this is greatly appreciated. otherwise we need to roll back to Scala 2.10 (or 2.9) and an older version of Akka. Thanks, Robert -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- Martynas Mickevičius Typesafe – Reactive Apps on the JVM -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- Martynas Mickevičius Typesafe – Reactive Apps on the JVM -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- Martynas Mickevičius Typesafe – Reactive Apps on the JVM -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- Björn Antonsson Typesafe – Reactive Apps on the JVM twitter: @bantonsson -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to a topic in the Google Groups "Akka User List" group. To unsubscribe from this topic, visit https://groups.google.com/d/topic/akka-user/i_fWDe8GN0I/unsubscribe. To unsubscribe from this group and all its topics, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout. -- Björn Antonsson Typesafe – Reactive Apps on the JVM twitter: @bantonsson -- >>>>>>>>>> Read the docs: http://akka.io/docs/ >>>>>>>>>> Check the FAQ: >>>>>>>>>> http://doc.akka.io/docs/akka/current/additional/faq.html >>>>>>>>>> Search the archives: https://groups.google.com/group/akka-user --- You received this message because you are subscribed to the Google Groups "Akka User List" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To post to this group, send email to [email protected]. Visit this group at http://groups.google.com/group/akka-user. For more options, visit https://groups.google.com/d/optout.
