Nope, I think there is neither a fix nor an open issue for this right now.
On Mon, 13 Jun 2016 at 11:31 Maximilian Michels wrote:
> Is there an issue or a fix for proper use of the ClojureCleaner in
> CoGroup.where()?
>
> On Fri, Jun 10, 2016 at 8:07 AM, Aljoscha Krettek
>
Serializer and read the "last_updated"
> property
> > - delete the key from RocksDB if the state's "last_updated" property is
> over
> > a month ago
> >
> > Is there any reason this approach wouldn't work, or anything to be
> careful
&g
Hi,
would these super-structure events occur per key? If yes, then I think you
can process this using the currently available windowing mechanism by
writing a custom WindowAssigner and Trigger. This, of course, assumes that
the events arrive in-order, i.e. if A-End arrives before A-Start or if
elem
order
> On 14 Jun 2016 14:04, "Paul Wilson" wrote:
>
>> Hi,
>>
>> No these super-structure events only serve the purpose of defining the
>> boundaries of a join, and do not relate to the keys of the sub-events.
>>
>> Thanks,
>> Paul
>&g
Hi,
right now migrating from one state backend to another is not possible. I
have it in the back of my head, however, that we should introduce a common
serialized representation of state to make this possible in the future.
(Both for checkpoints and savepoints, which use the same mechanism
undernea
Hi,
I'm afraid there is no documentation besides the link that you posted and
this one:
https://ci.apache.org/projects/flink/flink-docs-release-1.0/concepts/concepts.html
.
With what parts are you having trouble? Maybe I can help.
Cheers,
Aljoscha
On Thu, 16 Jun 2016 at 19:31 Bajaj, Abhinav wro
Hi,
I think the problem with the missing Class
com.amazon.ws.emr.hadoop.fs.EmrFileSystem is not specific to RocksDB. The
exception is thrown in the FsStateBackend, which is internally used by the
RocksDB backend to do snapshotting of non-partitioned state. The problem is
that the FsStateBackend tri
Great to hear that it works now! :-)
On Sun, 19 Jun 2016 at 16:33 Paschek, Robert
wrote:
> Hi Mailing List,
>
> after "upgrading" the flink version in my pom.xml to 1.0.3, i get two
> error messages for these output variants, which don't work:
>
> org.apache.flink.api.common.functions.InvalidTyp
Hi,
you can implement a RichSourceFunction. With this you can also implement
open() and close() methods that get called when your source is started on
the worker node and closed respectively. In there, you could determine the
hostname and send it to some centralized service so that it knows the
hos
Hi,
Chiwan is correct. The reason why we're (not yet) using managed memory in
the streaming API (DataStream) is that it was easier to get things up and
running by just using JVM heap. We're hoping to change this at some point
in the future, though.
Cheers,
Aljoscha
On Wed, 22 Jun 2016 at 14:05 Ch
structor, since it seems to get
> initialised later anyway)
>
> Josh
>
>
>
> On Fri, Jun 17, 2016 at 2:53 PM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> I think the problem with the missing Class
>> com.amazon.ws.emr.hadoop.fs.EmrFileSystem is not specific to
Hi,
the two map functions are called by the same thread, so waiting in one
function would block all processing. What you could do is buffer elements
from one input and only process them when an element arrives on the other
input.
Cheers,
Aljoscha
On Sun, 26 Jun 2016 at 13:36 Biplob Biswas wrote:
Maybe. But how do you mean, exactly?
On Mon, 27 Jun 2016 at 11:14 Janardhan Reddy
wrote:
> Hi,
>
> Instead of buffering can we use event creation time and watermarks ?
>
> On Mon, Jun 27, 2016 at 2:32 PM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> the two map functi
Hi,
could you maybe write TypeInformation/TypeSerializer wrappers that lazily
instantiate a type class-based serializer. It might even work using a "lazy
val". Something like this:
class ScalaTypeSerializer[T] extends TypeSerializer[T] {
lazy val serializer = "create the scala serializer"
...
Hi,
I might lead to flooding, yes. But I'm afraid it's the only way to go right
now.
Cheers,
Aljoscha
On Mon, 27 Jun 2016 at 17:57 Biplob Biswas wrote:
> Hi,
>
> I was afraid of buffering because I am not sure when the second map
> function
> would get data, wouldn't the first map be flooded wi
Hi,
one thing to add: if you use a ReduceFunction or a FoldFunction for your
window the state will not grow with bigger window sizes or larger numbers
of elements because the result is eagerly computed. In that case, state
size is only dependent on the number of individual keys.
Cheers,
Aljoscha
Hi,
I think this document is still up-to-date since not much was done in these
parts of the code for the 1.0 release and after that.
Maybe Timo can give some insights into what optimizations are done in the
Table API/SQL that will be be released in an updated version in 1.1.
Cheers,
Aljoscha
+Ti
> windows) ?
>
> Thanks and Regards,
> Vishnu Viswanath,
>
> On Tue, Jun 28, 2016 at 5:04 AM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> one thing to add: if you use a ReduceFunction or a FoldFunction for your
>> window the state will not grow with bigger window
Hi,
could you load the properties file when starting the application and add it
to the user functions so that it would be serialized along with them? This
way, you wouldn't have to ship the file to each node.
Cheers,
Aljoscha
On Wed, 29 Jun 2016 at 12:09 Janardhan Reddy
wrote:
> We are running
Hi,
you can explicitly specify that you want processing-time windows like this:
stream.keyBy(...).window(TumblingProcessingTimeWindows.of(Time.seconds(1))).sum(...)
Also note that the timestamp you append in
"writeAsCsv("records-per-second-" + System.currentTimeMillis())" will only
take the times
Hi,
I think there is no way to get the output from these log statements into
the Yarn logs. The reason is that this code is only executed on the client
and not in any Yarn context/container. This code is setting up everything
for Yarn and then control is handed over so it is executed before the Job
Hi Josh,
I think in your case the problem is that Scala might choose different names
for synthetic/generated classes. This will trip up the code that is trying
to restore from a snapshot that was done with an earlier version of the
code where classes where named differently.
I'm afraid I don't kno
Hi,
are you taking about *enableFullyAsyncSnapshots()* in the RocksDB backend.
If not, there is this switch that is described in the JavaDoc:
/**
* Enables fully asynchronous snapshotting of the partitioned state held in
RocksDB.
*
* By default, this is disabled. This means that RocksDB state is c
hen the class loader should be able to
> deserialize your serialized data.
>
> Cheers,
> Till
>
> On Thu, Jun 30, 2016 at 1:55 PM, Aljoscha Krettek
> wrote:
>
>> Hi Josh,
>> I think in your case the problem is that Scala might choose different
>> names for s
Hi,
I'm afraid the only way to do it right now is using the wrapper that can
contain both, as you suggested.
Cheers,
Aljoscha
On Thu, 30 Jun 2016 at 16:50 Martin Neumann wrote:
> Hej,
>
> I'm currently playing around with some machine learning algorithms in
> Flink streaming.
>
> I have an inpu
ger.Task.run(Task.java:546)
>
>
> On Fri, Jul 1, 2016 at 10:21 AM, Josh wrote:
>
>> Thanks guys, that's very helpful info!
>>
>> @Aljoscha I thought I saw this exception on a job that was using the
>> RocksDB state backend, but I'm not sure. I will
Hi,
do you have any data in the coGroup/groupBy operators that you use, besides
the input data?
Cheers,
Aljoscha
On Fri, 1 Jul 2016 at 14:17 Flavio Pompermaier wrote:
> Hi to all,
> I have a Flink job that computes data correctly when launched locally from
> my IDE while it doesn't when launche
Hi,
I think it should be as simple as setting event time as the stream time
characteristic:
env.setStreamTimeCharacteristic(TimeCharacteristic.EventTime)
The problem is that .timeWindow(Time.seconds(10)) will use processing time
if you don't specify a time characteristic. You can enforce using an
0:50 CST
> 100 events in this window
> Mon, 04 Jul 2016 19:10:50 CST
> Mon, 04 Jul 2016 19:10:50 CST
> Mon, 04 Jul 2016 19:10:50 CST
> Mon, 04 Jul 2016 19:10:50 CST
> Mon, 04 Jul 2016 19:10:51 CST
> Mon, 04 Jul 2016 19:10:51 CST
>
>
> On 4 July 2016 at 16:15, Aljoscha Kre
indow" indicating the end of the window.
>
> The 10s windows should be [19:10:40, 19:10:49], [19:10:50, 19:10:59], ...,
> but in the example above, the events at 19:10:50, which belong to
> [19:10:50, 19:10:59] window were mistakenly put in the [19:10:40, 19:10:49]
> one.
>
> On
s really matter since it is async
> anyways?
>
> Thanks and Regards,
> Vishnu Viswanath,
>
> On Thu, Jun 30, 2016 at 8:07 AM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> are you taking about *enableFullyAsyncSnapshots()* in the RocksDB
>> backend. If not, there is
don't give
enough space for expressing more complex opinions.
Cheers,
Aljoscha
On Mon, 30 May 2016 at 11:23 Aljoscha Krettek wrote:
> Thanks for the feedback! :-) I already read the comments on the file.
>
> On Mon, 30 May 2016 at 11:10 Gyula Fóra wrote:
>
>> Tha
This Blog post goes into the direction of what Jamie suggested:
https://techblog.king.com/rbea-scalable-real-time-analytics-king/ The folks
at King developed a system where users can dynamically inject scripts
written in Groovy into a running general-purpose Flink job.
On Thu, 7 Jul 2016 at 20:34
Hi,
I'm afraid there is no documentation about schedulers, especially at this
low level. Maybe this new design proposal would of interest for you,
though:
https://cwiki.apache.org/confluence/display/FLINK/FLIP-1+%3A+Fine+Grained+Recovery+from+Task+Failures
In
there is a link to the mailing list dis
Hi,
this does not work right now because FileInputFormat does not allow setting
the "enumerateNestedFiles" field directly and the Configuration is
completely ignored in Flink streaming jobs.
Cheers,
Aljoscha
On Wed, 13 Jul 2016 at 11:06 Robert Metzger wrote:
> Hi Dominique,
>
> In Flink 1.1 we'
I left that in on purpose to protect against cases where the combination of
key and namespace can be ambiguous. For example, these two combinations of
key and namespace have the same written representation:
key [0 1 2] namespace [3 4 5] (values in brackets are byte arrays)
key [0 1] namespace [2 3
Hi,
you have to ensure to filter the data that you send back on the feedback
edge, i.e. the loop.closeWith(newCentroids.broadcast()); statement needs to
take a stream that only has the centroids that you want to send back. And
you need to make sure to emit centroids with a good timestamp if you wan
Which is of course only available in 1.1-SNAPSHOT or the upcoming 1.1
release. :-)
On Tue, 19 Jul 2016 at 22:32 Till Rohrmann wrote:
> Hi Dominique,
>
> your problem sounds like a good use case for session windows [1, 2]. If
> you know that there is only a maximum gap between your request and re
Hi,
in the JobManager log there should be a line like this:
2016-07-20 17:19:00,552 INFO
org.apache.flink.runtime.webmonitor.WebRuntimeMonitor - Using directory
/some/dir for web frontend JAR file uploads
if you manually delete the offending jar file from that directory it could
solve your problem
Hi,
the configuration has to be passed using
env.readFile(...).withParameters(ifConf). The InputFormat will then be
properly configured at runtime.
However, Kostas just enhanced the FileInputFormats to allow setting the
parameters directly on the input format. In 1.1-SNAPSHOT and the upcoming
1.1
Hi David,
windows are being processed in order of their end timestamp. So if you
specify an allowed lateness of zero (which will only be possible on Flink
1.1 or by using a custom trigger) you should be able to sort the elements.
The ordering is only valid within one key, though, since windows for
ag
> the watermarks by 20 seconds, then only one instance of the Window (1-5)
> fires with elements A,B,C,D.
>
> Sameer
>
> On Thu, Jul 21, 2016 at 5:17 AM, Aljoscha Krettek
> wrote:
>
>> Hi David,
>> windows are being processed in order of their end timestamp.
t; other sources (except one) would keep sending data and their watermarks.
> Isn't this a risk for a possible Out of Memory Error. Should one always use
> a RocksDB alternative to mitigate such risks.
>
> Sameer
>
>
>
> On Thu, Jul 21, 2016 at 7:52 AM, Aljoscha Krettek
Looping in Max directly because he probably knows the Yarn stuff best.
@Max: Do you have any idea how to do this?
On Fri, 22 Jul 2016 at 05:46 김동일 wrote:
> I’saw the source code
> of
> flink-yarn/src/main/java/org/apache/flink/yarn/AbstractYarnClusterDescriptor.java
> Flink ships the FLINK_LIB
Hi,
yes, I'm afraid you would have to use a custom version of the
TumblingProcessingTimeWindows right now.
I've opened a Jira issue for adding an offset setting to the built-in
window assigners: https://issues.apache.org/jira/browse/FLINK-4282
Cheers,
Aljoscha
On Tue, 26 Jul 2016 at 12:51 Hirono
Hi,
you can perform a savepoint without changing your jar, yes.
Automatically taking the latest checkpoint as a savepoint is not possible
right now. We are working on adding support for that, however.
Cheers,
Aljoscha
On Tue, 26 Jul 2016 at 20:28 Shaosu Liu wrote:
> I want to load previous sta
Hi,
when running in local mode the default parallelism is always the number of
(possibly virtual) CPU cores. The parallelism of the sink is set before it
gets a chance to find out how many Kafka partitions there are. I think the
reason for the behavior you're observing is that only one of your two
Hi,
I think the exact thing you're trying do do is not possible right now but I
know of a workaround that some people have used.
For "warming up" the state from the historical data, you would run your
regular Flink job but replace the normal Kafka source by a source that
reads from the historical
When running "mvn clean verify" with Hadoop version 2.6.1 the
Zookeeper/Leader Election tests fail with this:
java.lang.NoSuchMethodError:
org.apache.curator.utils.PathUtils.validatePath(Ljava/lang/String;)Ljava/lang/String;
at
org.apache.curator.framework.imps.NamespaceImpl.(NamespaceImpl.java:37
Hi,
I have to try this to verify but I think the approach works if you give the
two sources different UIDs. The reason is that Flink will ignore state for
which it doesn't have an operator to assign it to. Therefore, the state of
the "historical Kafka source" should be silently discarded.
Cheers,
Hi,
the single-element-windows to me indicate that these originate from
elements that arrived at the window operator after the watermark. In the
current version of Flink these elements will be emitted as a single-element
window. You can avoid this by writing a custom EventTimeTrigger that does
not
Hi,
I'm afraid that's not possible right now. The preferred way of running
would be to have a Yarn cluster per job, that way you can isolate the logs.
Cheers,
Aljoscha
On Thu, 14 Jul 2016 at 09:49 Chawla,Sumit wrote:
> Hi Robert
>
> I actually mean both. Scenarios where multiple jobs are runni
Hi,
yes, if you set the delay to high you will have to wait a long time until
your windows are emitted.
Cheers,
Aljoscha
On Mon, 1 Aug 2016 at 04:52 Sendoh wrote:
> Probably `processAt` is not used adequately because after increasing
> maxDelay
> in watermark to 10 minutes it works as expected.
recover from removed
> tasks/operators without needing to add dummy operators like this.
>
> Josh
>
> On Fri, Jul 29, 2016 at 5:46 PM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> I have to try this to verify but I think the approach works if you give
>> the two so
+Till, looping him in directly, he probably missed this because he was away
for a while.
On Tue, 26 Jul 2016 at 18:21 Sameer W wrote:
> Hi,
>
> It looks like the WithIn clause of CEP uses Tumbling Windows. I could get
> it to use Sliding windows by using an upstream pipeline which uses Sliding
Hi,
if you manually force a parallelism different from 1 after a *windowAll()
then you will get parallel execution of your window. For example, if you do
this:
input.countWindowAll(100).setParallelism(5)
then you will get five parallel instances of the window operator that each
wait for 100 eleme
indow?
>
> Thanks
>
> Andrew
> From mobile
>
> From: Aljoscha Krettek
> Sent: Wednesday, August 3, 17:11
> Subject: Re: Parallel execution on AllWindows
> To: user@flink.apache.org
>
> Hi,
>
> if you manually force a parallelism different from 1 after a *windowAl
Hi,
a watermark cannot be sent before the element that makes you send that
watermark. A watermark of time T tells the system that no element will
arrive in the future with timestamp T or less, thus you cannot send it
before. It seems that what you are trying to achieve can be solved by using
sessio
Hi Davood,
right now, you can only inspect the timestamps by writing a custom operator
that you would use with DataStream.transform(). Measuring latency this way
has some pitfalls, though. The timestamp might be assigned on a different
machine than the machine that will process the tuple at the sin
Hi,
what does the reduce function do exactly? Something like this?
(a: String, b: String) -> b.toUppercase
If yes, then I would expect a) to be the output you get.
if it is this:
(a: String, b: String) -> a + b.toUppercase
then I would expect this: a,b,cC,d,eE,f,gG,h
Cheers,
Aljoscha
On Sun,
Hi Adam,
sorry for the inconvenience. This is caused by a new file read operator,
specifically how it treats watermarks/timestamps. I opened an issue here
that describes the situation:
https://issues.apache.org/jira/browse/FLINK-4329.
I think this should be fixed for an upcoming 1.1.1 bug fixing r
Hi,
could you maybe post how exactly you specify the window? Also, did you set
a "stream time characteristic", for example EventTime?
That could help us pinpoint the problem.
Cheers,
Aljoscha
On Tue, 9 Aug 2016 at 12:42 Andrew Ge Wu wrote:
> I rolled back to 1.0.3
> If I understand this correc
g something out of ordinary here.
>
>
>
> Thanks!
>
>
> Andrew
>
> On 09 Aug 2016, at 14:18, Aljoscha Krettek wrote:
>
> Hi,
> could you maybe post how exactly you specify the window? Also, did you set
> a "stream time characteristic", for example Even
Hi,
Kostas is right in that the elements are never explicitly sorted by
timestamp. In some cases they might not even be iterated in the order that
they were added so I would normally assume the order of the elements to be
completely arbitrary.
Cheers,
Aljoscha
On Wed, 10 Aug 2016 at 09:44 Kostas
that part, no, Im not explicitly set that.
>
>
> On 09 Aug 2016, at 15:29, Aljoscha Krettek wrote:
>
> Hi,
> are you setting a StreamTimeCharacteristic, i.e.
> env.setStreamTimeCharacteristic?
>
> Cheers,
> Aljoscha
>
> On Tue, 9 Aug 2016 at 14:52 Andrew Ge W
Hi,
I'm afraid you guessed correctly that it is not possible to ensure that
rules arrive before events. I think the way you solved it (with buffering)
is the correct way to go about this.
Cheers,
Aljoscha
On Wed, 10 Aug 2016 at 01:31 Sameer W wrote:
> Hi,
>
> I am using connected streams to sen
Oh, are you by any chance specifying a custom state backend for your job?
For example, RocksDBStateBackend.
Cheers,
Aljoscha
On Wed, 10 Aug 2016 at 11:17 Aljoscha Krettek wrote:
> Hi,
> could you maybe send us the output of "env.getExecutionPlan()". This would
> help us bett
Hi,
from your mail I'm gathering that you are in fact using an Evictor, is that
correct? If not, then the window operator should not keep all the elements
ever received for a window but only the aggregated result.
Side note, there seems to be a bug in EvictingWindowOperator that causes
evicted ele
Hi,
Sameet is right about the snapshotting. The CEP operator behaves more or
less like a FlatMap operator that keeps some more complex state internally.
Snapshotting works the same as with any other operator.
Cheers,
Aljoscha
On Thu, 11 Aug 2016 at 00:54 Sameer W wrote:
> Mans,
>
> I think at t
ermark, trigger's event timer is reached,
>fires and purges and emits current state as event z(time=1, count=2)
>9. Window B receives event, trigger waits for processing time delay,
>then executes fold() and emits event(time=1 => count=2), but internal
>Window stat
Hi,
Sameer is right that Flink currently does not combine for any combination
of assigner, trigger and window function.
Technically, it would be possible to use a combiner for Triggers that don't
observe individual elements but only fire on time. With triggers that
observe elements, such as CountT
Hi,
there is already this FLIP:
https://cwiki.apache.org/confluence/display/FLINK/FLIP-4+%3A+Enhance+Window+Evictor
which
also links to a mailing list discussion. And this FLIP:
https://cwiki.apache.org/confluence/display/FLINK/FLIP-2+Extending+Window+Function+Metadata.
The former proposes to enhan
Thanks for letting us know!
On Thu, 11 Aug 2016 at 07:33 Janardhan Reddy
wrote:
> sorry my bad, i was using some other version.
>
>
>
> On Thu, Aug 11, 2016 at 4:47 AM, Janardhan Reddy <
> janardhan.re...@olacabs.com> wrote:
>
>> Hi,
>>
>> the flink-dist_2.11-1.0.0.jar jar present in lib folder
Hi,
the input elements to a window function should not be modified. Could you
maybe achieve something using a Fold? Maybe if you went into a bit more
details we could figure something out together.
Cheers,
Aljoscha
On Tue, 16 Aug 2016 at 10:38 Ufuk Celebi wrote:
> Hey Paul! I think the window c
Hi,
this might be due to a bug in the Flink 1.1.0 maven dependencies. Can you
try updating to Flink 1.1.1?
Cheers,
Aljoscha
On Mon, 22 Aug 2016 at 07:48 wrote:
> Hi,
> every one , when i use scala version 2.10,and set the sbt project(add
> those:flink-core,flink-scala,flink-streaming-scala,
Hi,
you mean the directory is completely empty? Can you check in the JobManager
dashboard whether it reports any successful checkpoints for the job? One
possible explanation is an optimization that the FsStateBackend performs:
when the state is very small it will not actually be written to files bu
ark to the window
> function metadata in FLIP-2?
>
> From: Shannon Carey
> Date: Friday, August 12, 2016 at 6:24 PM
> To: Aljoscha Krettek , "user@flink.apache.org" <
> user@flink.apache.org>
>
> Subject: Re: Firing windows multiple times
>
> Thanks Al
Hi Aris,
I think you're on the right track with using a CoFlatMap for this. Could
you maybe post the code of your CoFlatMapFunction (or you could send it to
me privately if you have concerns with publicly posting it) then I could
have a look.
Cheers,
Aljoscha
On Mon, 29 Aug 2016 at 15:48 aris kol
rent watermark) on an ongoing basis.
> The windowing function would be responsible for evicting old data based on
> the current watermark.
>
> Does that make sense? Does it seem logical, or am I misunderstanding
> something about how Flink works?
>
> -Shannon
>
>
> Fr
>}
>
>flatMap(in, out)
>
> }
> }
> }
>
>
> applyWithState throws the exception and my intuition says I am doing
> seriously wrong in the instantiation. I tried to make something work using
> the mapWithState implementation as a guide and I ended up here.
>
&g
Hi Konstantin,
I think this is not possible with the current API but I've been thinking
about similar stuff this week. Let me quickly outline what I was thinking
and then you can tell me whether that would also be helpful for you.
The basic problem is this: I want to be able to write ITCases that
Just checking, all the elements that would fall into a window of length X
also have X as a property? In that case you should be able to do something
like this:
public Collection assignWindows(PojoType element, long
timestamp, WindowAssignerContext context) {
long size = element.windowSize;
Just one clarification: even with a specified allowed lateness the window
will still be evaluated once the watermark passes the end of the window.
It's just that with allowed lateness the window contents and state will be
kept around a bit longer to allow eventual late elements to update the
result
ll
> be on average 0.5 months stale. A year-long window is even worse.
>
> -Shannon
>
> From: Aljoscha Krettek
> Date: Tuesday, August 30, 2016 at 9:08 AM
> To: Shannon Carey , "user@flink.apache.org" <
> user@flink.apache.org>
>
> Subject: Re: Firing
Hi,
from this I would expect to get as many HashMaps as you have keys. The
winFunction is also executed per-key so it cannot combine the HashMaps of
all keys.
Does this describe the behavior that you're seeing?
Cheers,
Aljoscha
On Fri, 2 Sep 2016 at 17:37 Luis Mariano Guerra
wrote:
> hi!
>
> I
Hi,
right now it is not possible to mix the DataSet and the DataStream API. The
reason for the "task not serializable" error is that putting the DataSet
into the map function tries to serialize the DataSet, which is not possible.
Cheers,
Aljoscha
On Tue, 30 Aug 2016 at 16:31 wrote:
> Hi,
>
That's true. The reason why it works in Flink is that a slow downstream
operator will back pressure an upstream operator which will then slow down.
The technical implementation of this relies on the fact that Flink uses a
bounded pool of network buffers. A sending operator writes data to network
bu
Hi,
which version of Flink are you using? Are the checkpoints being reported as
successful in the Web Frontend, i.e. in the "checkpoints" tab of the
running job?
Cheers,
Aljoscha
On Fri, 2 Sep 2016 at 12:17 Dong-iL, Kim wrote:
> Hi,
>
> I’m using HDFS as state backend.
> The checkpoints folder
+Max Michels
Directly looping in Max. You recently worked on the clients code, do you
have any Idea if and how this is possible?
On Fri, 2 Sep 2016 at 14:38 Paul Wilson wrote:
> Hi,
>
> I'd like to write a client that can execute an already 'uploaded' JAR
> (i.e. the JAR is deployed and availab
,
Aljoscha
On Fri, 2 Sep 2016 at 18:26 Luis Mariano Guerra
wrote:
> On Fri, Sep 2, 2016 at 5:24 PM, Aljoscha Krettek
> wrote:
>
>> Hi,
>> from this I would expect to get as many HashMaps as you have keys. The
>> winFunction is also executed per-key so it cannot combine
h can have an impact on the processing guarantees
> when a failure/recovery occurs if we don't do it carefully. Also, we're not
> particularly sophisticated yet with regard to avoiding unnecessary queries
> to the time series data.
>
> -Shannon
>
>
> From: Aljoscha K
ljoscha
On Mon, 5 Sep 2016 at 17:35 Luis Mariano Guerra
wrote:
> On Mon, Sep 5, 2016 at 12:30 PM, Aljoscha Krettek
> wrote:
>
>> Hi,
>>
> for this you would have to use a non-parallel window, i.e. something like
>> stream.windowAll().apply(...). This does not
Hi,
I'd be very happy to give you pointers for FLIP-2 and FLIP-4. Why don't you
start a separate thread on the dev list so that we don't hijack this thread.
For FLIP-4 we also have to coordinate with Vishnu, he was driving FLIP-4
but lately everyone has been a bit inactive on that. Let's see if he
Hi,
you don't need the BlockedEventState class, you should be able to just do
this:
private transient ValueState blockedRoads;
@Override
public void open(final org.apache.flink.configuration.Configuration
parameters) throws Exception {
final ValueStateDescri
Hi,
yes AJ that observation is correct. Let's see what Shannon has to say about
this but it might be that all "higher-level" aggregates will have to be
based on the first level and can then update at the speed of that aggregate.
Cheers,
Aljoscha
On Mon, 12 Sep 2016 at 05:03 aj.h wrote:
> In the
Hi,
this is a different job from the Kafka Job that you have running, right?
Could you maybe post the code for that as well?
Cheers,
Aljoscha
On Tue, 13 Sep 2016 at 20:14 amir bahmanyari wrote:
> Hi Robert,
> Sure, I am forwarding it to user. Sorry about that. I followed the
> "robot's" instru
Hi,
the problem might be that your timestamp/watermark assigner is run in
parallel and that only one parallel instance of those operators emits the
watermark because only one of those parallel instances sees the element
with _3 == 9000. For the watermark to advance at an operator it needs to
advanc
Hi,
WindowFunction.apply() will be called once for each window so you should be
able to do the setup/teardown in there. open() and close() are called at
the start of processing, end of processing, respectively.
Cheers,
Aljoscha
On Wed, 14 Sep 2016 at 09:04 Swapnil Chougule
wrote:
> Hi Team,
>
>
hile #slots=64 is the same.
>
> Its still slow for a relatively large file though.
> Pls advice if something I can try to improve the cluster performance.
> Thanks+regards
>
> --
> *From:* Aljoscha Krettek
> *To:* user@flink.apache.org; amir bahm
Hi,
when is the "first time". It seems you have tried this repeatedly so what
differentiates a "first time" from the other times? Are you closing your
IDE in-between or do you mean running the job a second time within the same
program?
Cheers,
Aljoscha
On Fri, 9 Sep 2016 at 16:40 Yassine MARZOUGU
501 - 600 of 1265 matches
Mail list logo