...this feels like a bug to me. I think Erik-Jan's expectation that nothing would have begun for ListenUDP given primary node only config is fair. I also think our current position of 'just not calling onTrigger' is fair too but less intuitive for users.
What do ya'll think? On Wed, Jun 5, 2019 at 12:10 PM Erik-Jan <erik...@gmail.com> wrote: > This is what I do basically. I want a highly available setup with a > minimum of components since each component can (and will) fail. Traffic > reaches all nodes but only a single node should read it. > > Op wo 5 jun. 2019 18:07 schreef James Srinivasan < > james.sriniva...@gmail.com>: > >> In our case the stream is UDP broadcast, so available to all nodes >> anyway. I've been meaning to test UDP multicast but not got round to it yet. >> >> >> On Wed, 5 Jun 2019, 17:03 Bryan Bende, <bbe...@gmail.com> wrote: >> >>> That is probably a valid point, but how about putting a load balancer >>> in front to handle that? >>> >>> On Wed, Jun 5, 2019 at 11:30 AM James Srinivasan >>> <james.sriniva...@gmail.com> wrote: >>> > >>> > Presumably you'd want to mirror the stream to all nodes for when the >>> primary node changes? >>> > >>> > On Wed, 5 Jun 2019, 13:46 Bryan Bende, <bbe...@gmail.com> wrote: >>> >> >>> >> The processor is started on all nodes, but onTrigger method is only >>> >> executed on the primary node. >>> >> >>> >> This is something we've discussed trying to improve before, but the >>> >> real question is why are you sending data to the other nodes if you >>> >> don't expect the processor to execute there? >>> >> >>> >> On Wed, Jun 5, 2019 at 7:04 AM Erik-Jan <erik...@gmail.com> wrote: >>> >> > >>> >> > I figured it out after further testing. The processor runs on all >>> nodes, despite the explicit "run on primary node only" option that I >>> selected. But only on the primary node the queue is processed. On the other >>> nodes the queue gets filled until the max is reached after which the error >>> message starts appearing. What I missed before is that the message is >>> coming from the other, non-primary nodes. >>> >> > I'm not sure if this is intended behavior or if it is a bug though! >>> For me it's a bug since I really want this processor to run on the primary >>> only. >>> >> > >>> >> > Op di 4 jun. 2019 16:34 schreef Erik-Jan <erik...@gmail.com>: >>> >> >> >>> >> >> Hi Bryan, >>> >> >> >>> >> >> Yes I have considerably increased the numbers in the controller >>> settings. >>> >> >> I don't mind getting my hands dirty, increasing the timeout is >>> worth a try. >>> >> >> >>> >> >> The errors seems to appear after quite a while. Usually I see >>> these messages the next morning so testing and experimenting with this >>> error takes a lot of time. >>> >> >> >>> >> >> Today I've been trying to reproduce this on a virtual machine with >>> the same OS, Nifi and Java versions but to no avail. The difference is that >>> this VM is not a cluster, has limited memory and cpu and still is able to >>> handle much more UDP data with the error appearing only a few times so far >>> after hours of running. It leads me to thinking there must be something in >>> the configuration of the cluster thats causing this. I will also try a >>> vanilla Nifi install on one of the nodes without clustering to see if my >>> configuration and cluster setup is somehow the cause. >>> >> >> >>> >> >> Op di 4 jun. 2019 om 16:14 schreef Bryan Bende <bbe...@gmail.com>: >>> >> >>> >>> >> >>> Hi Erik, >>> >> >>> >>> >> >>> It sounds like you have tried most of the common tuning options >>> that >>> >> >>> can be done. I would have expected batching + increasing >>> concurrent >>> >> >>> tasks from 1 to 3-5 to be the biggest improvement. >>> >> >>> >>> >> >>> Have you increased the number of threads in your overall thread >>> pool >>> >> >>> according to your hardware? (from the top right menu controller >>> >> >>> settings) >>> >> >>> >>> >> >>> I would be curious what happens if you did some tests increasing >>> the >>> >> >>> timeout where it attempts to place the message in the queue from >>> 100ms >>> >> >>> to 200ms and then maybe 500ms if it still happens. >>> >> >>> >>> >> >>> I know this requires a code change since that timeout is >>> hard-coded, >>> >> >>> but it sounds like you already went down that path with trying a >>> >> >>> different queue :) >>> >> >>> >>> >> >>> -Bryan >>> >> >>> >>> >> >>> On Tue, Jun 4, 2019 at 4:28 AM Erik-Jan <erik...@gmail.com> >>> wrote: >>> >> >>> > >>> >> >>> > Hi, >>> >> >>> > >>> >> >>> > I'm experimenting with a locally installed 3 node nifi cluster. >>> This cluster receives UDP packets on the primary node. >>> >> >>> > These nodes are pretty powerful, have a good network >>> connection, have lots of memory and SSD disks. I gave nifi 24G of java heap >>> (xms and xmx). >>> >> >>> > >>> >> >>> > I have configured a ListenUDP processor that listens on a UDP >>> port and it receives somewhere between 20000 to 50000 packets per 5 >>> minutes. It's "Max size of message queue" is large enough (1M), I gave it 5 >>> concurrent tasks, it's running on the primary node only. >>> >> >>> > >>> >> >>> > The problem: after running for a while, I get the following >>> error: "internal queue at maximum capacity, could not queue event." >>> >> >>> > >>> >> >>> > I have reviewed the source code and understand when this >>> happens. It happens when the processor tries to store an event in a java >>> LinkedBlockingQueue and that queue reached its maximum capacity. The >>> offer() method has a 100ms timeout in which it waits for space to free up >>> and then it fails and the event gets dropped. In the logs I see exactly 10 >>> of these error messages per second (10 x 100ms is 1 second). Despite these >>> errors, I still get a very good rate of events that get through to the next >>> processors. Actually, it seems pretty much all of the other events get >>> through since the message rate in ListenUDP and the followup processor are >>> very much alike. The followup processors can easily handle the load and >>> there are no full queues, congestions or anything like that. >>> >> >>> > >>> >> >>> > What I have tried so far: >>> >> >>> > >>> >> >>> > Increasing the "Max Size of Message Queue" setting helps, but >>> only delays the errors. They eventually return. >>> >> >>> > >>> >> >>> > Increasing heap space is a suggestion I read from a past post: >>> I think 24G is more than enough actually? Perhaps even too much? >>> >> >>> > >>> >> >>> > Increasing parallelism: concurrent tasks set to 5 or 10 does >>> not help. >>> >> >>> > >>> >> >>> > I modified the code to use an ArrayBlockingQueue instead of the >>> LinkedBlockingQueue, thinking it was some kind of garbage collection. This >>> didn't help. >>> >> >>> > >>> >> >>> > I increased "Receive Buffer Size", "Max Size of Socket Buffer" >>> but to no avail. >>> >> >>> > >>> >> >>> > I tried batching. This helps a bit, like increasing the "Max >>> Size of Message Queue" it only seems to delay the eventual error messages >>> though. >>> >> >>> > >>> >> >>> > I reproduced this on my local workstation. I installed nifi, >>> did no OS tuning at all, set the heap size to 4GB. I generate 1.3M UDP >>> packets per 5 minutes (the max I can reach with a simple python script). >>> With "Max Size of Message Queue" set to only 100, soon the error appears. >>> In the ListenUDP processor I see 1.34M events out, on the followup >>> processor I see 1.34M events incoming. The error is not as frequent as on >>> the cluster though, only a few every couple of minutes while the data rate >>> is much higher and the queue much smaller. I'm a bit desperate and hope >>> anyone can help me out. Why am I getting this error on a relatively quiet >>> cluster with not that much load? >>> >> >>> > >>> >> >>> > Best regards, >>> >> >>> > Erik-Jan van Baaren >>> >>