I am working on a kafka stream app, and see huge latency variance,
wondering what can cause this?
the processing is very simple and don't have state, linger.ms already
change to 5ms. the message size is around 10K byes and published as 2000
messages/s, network is 10G. using a regular consumer wat
btw, I am using version 0.10.2.0
On Fri, Aug 17, 2018 at 2:04 PM Nan Xu wrote:
> I am working on a kafka stream app, and see huge latency variance,
> wondering what can cause this?
>
> the processing is very simple and don't have state, linger.ms already
> change to 5ms.
latency spikes original on write and not on read: you might
> want to have a look into Kafka Streams JMX metric to see if processing
> latency spikes or throughput drops.
>
> Also watch for GC pauses in the JVM.
>
> Hope this helps.
>
>
> -Matthias
>
> On 8/1
RUNNING since only after
> that the streams client will start to process the first record.
>
>
> Guozhang
>
>
> On Sat, Aug 18, 2018 at 8:52 PM, Nan Xu wrote:
>
> > thanks, which JMX properties indicate "processing latency spikes" /
> > "through
; only the latter has rebalance process, while the Kafak brokers do not
> really have "rebalances" except balancing load by migrating partitions.
>
> Guozhang
>
>
>
> On Sun, Aug 19, 2018 at 7:47 PM, Nan Xu wrote:
>
> > right, so my kafka cluster is alr
I did several test. one is with 10 brokers (remote server),
one with 3 brokers. (local docker)
both exhibit the same behavior, I was thinking the same but from at least
the kafka log, I don't see a rebalance happening. and I am sure my cpu is
only used about half. and all broker still running.
more
expecting 100,000 m/s and less than 10ms latency for a single powerful
broker.
Nan
On Mon, Aug 20, 2018 at 12:45 AM Nan Xu wrote:
> I did several test. one is with 10 brokers (remote server),
> one with 3 brokers. (local docker)
>
> both exhibit the same behavior, I was thinking
ds). Is that aligned
> with the frequency you observe latency spikes?
>
>
> Guozhang
>
>
> On Sun, Aug 19, 2018 at 10:41 PM, Nan Xu wrote:
>
> > did more test and and make the test case simple.
> > all the setup now is a single physical machine. running 3 docker
&g
I setup a local single node test. producer and broker are sitting at the
same VM. broker only has a single node(localhost) and a single partition.
producer produce message as fast as it could in a single thread. all update
to a SINGLE key(String). the kafka broker data directory is memory based
dir
r hard drive bandwidth ? So you can take a look at your iostats
>
> --
> Sent from my iPhone
>
> On Aug 22, 2018, at 8:20 PM, Nan Xu wrote:
>
> I setup a local single node test. producer and broker are sitting at the
> same VM. broker only has a single node(localhost) and
>
>
>
> which is pure stateless, committing will not touch on an state directory at
> all. Hence committing only involves committing offsets to Kafka.
>
>
> Guozhang
>
>
> On Wed, Aug 22, 2018 at 8:11 PM, Nan Xu wrote:
>
> > I
ker machine? it’s
> a
> > Linux kernel parameter.
> >
> > -Sudhir
> >
> > > On Aug 23, 2018, at 4:46 PM, Nan Xu wrote:
> > >
> > > I think I found where the problem is, how to solve and why, still not
> > sure.
> > >
> > >
Hi Guozhang,
Here is the very simple kafka producer/consumer/stream app, using the
latest version and just create 2 topics
input and output
all component are just running on localhost.
Thanks,
Nan
-Original Message-
From: Nan Xu [mailto:nanxu1...@gmail.com]
Sent: Friday
void spaceMessageWithInterval(){
int i =0 ;
long baseTime = System.nanoTime();
long doneTime = baseTime + duration;
while(true) {
task.run();
pubTime.add(System.nanoTime());
long targetTime = System.nanoTime() + interval;
i
maybe easier to use github.
https://github.com/angelfox123/kperf
On Sat, Aug 25, 2018 at 8:43 PM Nan Xu wrote:
> so I did upgrade to 2.0.0 and still seeing the same result. below is the
> program I am using. I am running everything on a single server. (centos 7,
> 24 core, 32
tly what is the root cause: the
> community can share with your some past experience and a few quick hinters,
> but most likely the issue varies case by case and hence can only be fully
> understandable by yourself.
>
>
> Guozhang
>
> On Sat, Aug 25, 2018 at 6:58 PM, Nan Xu wro
hi,
I was writing a simple stream app, all it does is producer send a sequence
of path and value, for example
path /0 , value 1
path /0/1, value 2
path /0/1/2, value 3
and kafka stream take those input and produce a ktable store.
There is a rule. if parent path is not exist, then child can not i
when I do the transform, for a single input record, I need to output 3
different records, those 3 records are in different classes. I want to
send the each type of records to a separate topic, my understanding is I
should use
context.forward inside the transformer like
Transformer{..
context.fo
the parent node of all 3 sink nodes.
> Then in your Transformer, you can forward the key-value pairs by using one
> of two approaches.
>
> Sending to all child nodes with this call:
>
> context().forward(key, value, To.all()).
>
> Or by listing each child node individually like so
&g
I, sorry if I
> > didn't make this clear from before.
> >
> > Thanks,
> > Bill
> >
> > On Thu, Feb 7, 2019 at 10:41 AM Nan Xu wrote:
> >
> >> thanks, just to make sure I understand this correctly,.
> >>
> >> I have some proce
0).to("topic1"..);
> val allStreams(1).to("topic2"..);
> val allStreams(2).to("topic3"..);
>
> HTH,
> Bill
>
>
>
> On Thu, Feb 7, 2019 at 11:51 AM Nan Xu wrote:
>
> > hmm, but my DSL logic at beginning involve some join between di
[2]
>
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-313%3A+Add+KStream.flatTransform+and+KStream.flatTransformValues
>
>
> Guozhang
>
> On Thu, Feb 7, 2019 at 12:59 PM Nan Xu wrote:
>
> > awesome, this solution is great, thanks a lot.
> >
> > Nan
&
Hi,
Just wondering if there is a way to do a sql like "select key,
value.fieild1 from ktable where key like abc%"
The purpose of this to select some value from a ktable without a fully
defined key. Store.all then filter on them would be very inefficient if
store is big.
Thanks,
Nan
ore
> false positives.
>
> Also, this only work for prefix queries, ie, if you query with a know
> prefix of the key.
>
> Hope this helps.
>
> -Matthias
>
> On 2/12/19 8:25 AM, Nan Xu wrote:
> > Hi,
> >
> > Just wondering if there is a way to do a s
just a general question about the rocksdb in the Kafka stream, I see there
is a folder at /tmp/kafka-stream/, which is used by the rocksdb in the
kafka stream. so when a stream app get restarted, can the store data
directly loaded from this folder? because I see there is very heavy traffic
on the n
hi,
trying the following program and want to see the metadata for
test_store, and nothing get back, the val metaIter =
streams.allMetadata().iterator() size is 0. I can see data in the store
though, but I need metadata so when I have multiple instance running. I can
find the right store.
is th
anyone can give some suggestion? or an explanation why kafka give a big
latency for large payload.
Thanks,
Nan
On Thu, Mar 14, 2019 at 3:53 PM Xu, Nan wrote:
> Hi,
>
> We are using kafka to send messages and there is less than 1% of
> message is very big, close to 30M. understanding kafka i
that's very good information from the slides, thanks. Our design to use
kafka has 2 purpose. one is use it as a cache, we use ktable for that
purpose, second purpose is use as message delivery mechanism to send it to
other system. Because we very much care the latency, the ktable with a
compact top
28 matches
Mail list logo