I download Flink 1.5 and I'm trying to run it in standalone mode.
1 job manager, 2 task managers.
I can run flink job when I run in local mode: 1 machine as both job manager
and task manager.
But when I add 2 remote machines as slaves and try to run, I am seeing this
error in the log and the
Hi guys,
I have another question related to the KPL problem. I wonder what the
consequences of overwhelming KPL internal queue (kinesis) can be.
From my observation in experimenting with 1.4.2 (which does not have
backpressure support yet in the open pr stated below), when the flink cluster
Hi,all:
I want to use some other broadcast resources such as list or map in the
flatmap function or customer triggers, but I don’t find some api to satisfy.
Anyone can help?
thanks
Alas, that error appears to be a red herring. Admin mistyped the cancel
command leading to the error. But immediately corrected it, resulting in
the job being canceled next. So seems unrelated to the job coming back to
life later on.
On Wed, Jun 20, 2018 at 10:04 AM Elias Levy
wrote:
> The
hmm, this indeed looks odd. Looping in Till (cc) who might know more
about this.
On 20.06.2018 16:43, Dominik Wosiński wrote:
Hello,
I'm not sure whether the problem is connected with bad configuration
or it's some inconsistency in the documentation but according to this
The source of the issue may be this error that occurred when the job was
being canceled on June 5:
June 5th 2018, 14:59:59.430 Failure during cancellation of job
c59dd3133b1182ce2c05a5e2603a0646 with savepoint.
java.io.IOException: Failed to create savepoint directory at
--checkpoint-dir
at
We had an unusual situation last night. One of our Flink clusters
experienced some connectivity issues, with lead to the the single job
running on the cluster failing and then being restored.
And then something odd happened. The cluster decided to also restore an
old version of the job. One we
Thanks, Gordon. You are quick and It is very helpful to me.
I tried some other alternatives to resolve this, finally thought about
rewriting the FlinkKinesisProducer class for our need. Glad that I asked before
I started.
Really appreciate the quick response.
From: "Tzu-Li (Gordon) Tai"
Date:
Hi Gavin,
The problem is that the Kinesis producer currently does not propagate
backpressure properly.
Records are added to the internally used KPL client’s queue, without any queue
size limit.
This is considered a bug, and already has a pull request for it [1], which we
should probably push
Hi guys,
I am new to flink framework. And we are building an application that takes
kinesis stream for both flink source and sink.
The flink version we are using is 1.4.2, which is also the version for the
flink-connector-kinesis. We built the flink-connector-kinesis jar explicitly
with KPL
Hi Fabian and Chesnay,
As Chesnay pointed out, it seems that I need to write the current counter
(which is defined inside Trigger) into state which I think should be the
operator state of the window operator.
However, as I previously said, TriggerContext allows for users to access
only the
So, finally, I have got this working. The issue was because of a poor
library which was using xerces 2.6 :).
In this process, I found few things missing from the doc would like to
contribute the same.
I really appreciate the support provided.
Thanks,
On Tue, 19 Jun 2018 at 4:05 PM, Ted Yu
Hello,
I'm not sure whether the problem is connected with bad configuration or
it's some inconsistency in the documentation but according to this document:
https://cwiki.apache.org/confluence/display/FLINK/FLIP-19%3A+Improved+BLOB+storage+architecture
. *I*f a job fails, all non-HA files'
Hi,
You can “skip” the corrupted message by returning `null` from the deserialize
method on the user-provided DeserializationSchema.
This lets the Kafka connector consider the record as processed, advances the
offset, but doesn’t emit anything downstream for it.
Hope this helps!
Cheers,
Hi,
You can use FlatMap instead of Map, and only collect valid elements.
Regards,
Kien
On 6/20/2018 7:57 AM, chrisr123 wrote:
First time I'm trying to get this to work so bear with me. I'm trying to
learn checkpointing with Kafka and handling "bad" messages, restarting
without losing
Hi Amol,
> In above code also it will sort the records in specific time window only.
All windows will be emitted as watermark passes the end of the window. The
watermark only increases. So the non-overlapping windows should be also sorted
by time and as a consequence the records across windows
to that last one: i'm accessing S3 from one EC2 instance which has a IAM
Role attached ...
I'll get back to you when i have those stacktraces printed ... will have to
build the project and package the custom version first, might take some
time, and also some vacation is up next ...
Cheers
--
Hi,
I was looking in this more, and I have couple of suspicions, but it’s still
hard to tell which is correct. Could you for example place a breakpoint (or add
a code there to print a stack trace) in
org.apache.log4j.helpers.AppenderAttachableImpl#addAppender
And check who is calling it? Since
Checkpointing of metrics is a manual process.
The operator must write the current value into state, retrieve it on
restore and restore the counter's count.
On 20.06.2018 12:10, Fabian Hueske wrote:
Hi Dongwon,
You are of course right! We need to decrement the counter when the
window is
Hi Avihai,
Rafi pointed out the two common approaches to deal with this situation. Let
me expand a bit on those.
1) Transactional producing in to queues: There are two approaches to
accomplish exactly-once producing into queues, 1) using a system with
transactional support such as Kafka or 2)
Hi Dongwon,
You are of course right! We need to decrement the counter when the window
is closed.
The idea of using Trigger.clear() (the clean up method is called clear()
instead of onClose()) method is great!
It will be called when the window is closed but also when it is merged.
So, I think you
Hi,
Good point, sorry for confusion, BoundedOutOfOrdernessTimestampExtractor of
course does not buffer records, you need to apply windowing (e.g.
TumblingEventTimeWindows) for that and then sort the window output by time and
emit records in sorted order.
You can also use windowAll which
Hi,
it is possible that the number of processing time timers can grow, because
internal timers are scoped by time, key, and namespace (typically this means
„window“, because each key can be part of multiple windows). So if the number
of keys in your application is steadily growing this can
Hi,
I think a global ordering is a bit impractical on production, but in theroy,
you still can do that. You need to
- Firstly fix the operate's parallelism to 1(except the source node).
- If you want to sort the records within a bouned time, then you can keyBy() a
constant and window it,
It will, but it defaults to jobmanager.rpc.address if no rest.address has
been specified.
On Wed, Jun 20, 2018 at 9:49 AM Chesnay Schepler wrote:
> Shouldn't the non-HA case be covered by rest.address?
>
> On 20.06.2018 09:40, Till Rohrmann wrote:
>
> Hi Sampath,
>
> it is no longer possible to
Shouldn't the non-HA case be covered by rest.address?
On 20.06.2018 09:40, Till Rohrmann wrote:
Hi Sampath,
it is no longer possible to not start the rest server endpoint by
setting rest.port to -1. If you do this, then the cluster won't start.
The comment in the flink-conf.yaml holds only
Hi Sampath,
it is no longer possible to not start the rest server endpoint by setting
rest.port to -1. If you do this, then the cluster won't start. The comment
in the flink-conf.yaml holds only true for the legacy mode.
In non-HA setups we need the jobmanager.rpc.address to derive the hostname
Hi Avihai,
The problem is that every message queuing sink only provides at-least-once
> guarantee
>
>From what I see, possible messaging queue which guarantees exactly-once is
Kafka 0.11, while using the Kafka transactional messaging
I was worried this might be the case.
The rest.port handling was simply copied from the legacy web-server,
which explicitly allowed shutting it down.
It may (I'm not entirely sure) also not be necessary for all deployment
modes; for example if the job is baked into the job/taskmanager images.
I think you can set the target-directory to null. But I'm not sure why
this particular request requires this, other request allow optional
fields to simply be ommitted...
On 20.06.2018 06:12, Siew Wai Yow wrote:
Hi all,
Seems pass in target-directory is a must now for checkpoints REST API,
Hi Chesnay
Adding on to this point you made - " the rpc address is still *required *due
to some technical implementations; it may be that you can set this to some
arbitrary value however."
For job submission to happen successfully we should give specific rpc
address and not any arbitrary value.
31 matches
Mail list logo