flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Maximilian Bode <maximilian.b...@tngtech.com>
Subject Re: Jobmanager HA with Rolling Sink in HDFS
Date Wed, 09 Mar 2016 07:31:10 GMT
Hi Aljoscha,

yeah I should have been clearer. I did mean those accumulators but am not trusting them in
the sense of total number (as you said, they are reset on failure). On the other hand, if
they do not change for a while it is pretty obvious that the job has ingested everything in
the queue. But you are right, this is kind of heuristic. In combination with the fact that
the DateTimeBucketer does not create new folders I believe this should be sufficient to decide
when the job has basically finished, though.

So the setup is the following: The Flink job consists of a FlinkKafkaConsumer08, a map containing
just an IntCounter accumulator and finally a rolling sink writing to HDFS. I start it in a
per-job yarn session with n=3, s=4. Then I pour 2 million records in the Kafka queue the application
is reading from. If no job/task managers are killed, the behavior is exactly as expected:
the output files in HDFS grow with time and I can exactly monitor via the accumulator when
every record has been ingested from Kafka. After that time, I give the job a few seconds and
then cancel it via the web interface. Then still some time later (to give the job the chance
to output the few records still hanging around) a wc -l on the output files yields exactly
the expected 2 million.

On the other hand, if I kill a task manager while the job is in progress, one of the 12 output
files seems to be missing as described before. A wc -l on only the relevant bytes as I mentioned
in an earlier mail then leads to a number smaller than 2 million.

We are using an FsStateBackend in HDFS with a checkpoint interval of 10s.

Cheers,
 Max
—
Maximilian Bode * Junior Consultant * maximilian.bode@tngtech.com
TNG Technology Consulting GmbH, Betastr. 13a, 85774 Unterföhring
Geschäftsführer: Henrik Klagges, Christoph Stock, Dr. Robert Dahlke
Sitz: Unterföhring * Amtsgericht München * HRB 135082

> Am 08.03.2016 um 17:46 schrieb Aljoscha Krettek <aljoscha@apache.org>:
> 
> Hi,
> with accumulator you mean the ones you get from RuntimeContext.addAccumulator/getAccumulator?
I’m afraid these are not fault-tolerant which means that the count in these probably doesn’t
reflect the actual number of elements that were processed. When a job fails and restarts the
accumulators should start from scratch. This makes me wonder how yours ever reach the required
2 mio, for it to be considered “done”.
> 
> This keeps getting more mysterious…
> 
> By the way, what are you using as StateBackend and checkpoint interval?
> 
> Cheers,
> Aljoscha
>> On 08 Mar 2016, at 13:38, Maximilian Bode <maximilian.bode@tngtech.com> wrote:
>> 
>> Hi,
>> thanks for the fast answer. Answers inline.
>> 
>>> Am 08.03.2016 um 13:31 schrieb Aljoscha Krettek <aljoscha@apache.org>:
>>> 
>>> Hi,
>>> a missing part file for one of the parallel sinks is not necessarily a problem.
This can happen if that parallel instance of the sink never received data after the job successfully
restarted.
>>> 
>>> Missing data, however, is a problem. Maybe I need some more information about
your setup:
>>> 
>>> - When are you inspecting the part files?
>> Some time after the cluster is shut down
>>> - Do you shutdown the Flink Job before checking? If so, how do you shut it down.
>> Via 'cancel' in the Jobmanager Web Interface. Some records seem to be written only
after cancelling the job, right?
>>> - When do you know whether all the data from Kafka was consumed by Flink and
has passed through the pipeline into HDFS?
>> I have an accumulator in a map right before writing into HDFS. Also, the RollingSink
has a DataTimeBucketer which makes it transparent when no new data is arriving anymore as
the last bucket is from some minutes ago.
>>> 
>>> Cheers,
>>> Aljoscha
>>>> On 08 Mar 2016, at 13:19, Maximilian Bode <maximilian.bode@tngtech.com>
wrote:
>>>> 
>>>> Hi Aljoscha,
>>>> 
>>>> oh I see. I was under the impression this file was used internally and the
output being completed at the end. Ok, so I extracted the relevant lines using
>>>> 	for i in part-*; do head -c $(cat "_$i.valid-length" | strings) "$i" >
"$i.final"; done
>>>> which seems to do the trick.
>>>> 
>>>> Unfortunately, now some records are missing again. In particular, there are
the files
>>>> 	part-0-0, part-1-0, ..., part-10-0, part-11-0, each with corresponding .valid-length
files
>>>> 	part-0-1, part-1-1, ..., part-10-0
>>>> in the bucket, where job parallelism=12. So it looks to us as if one of the
files was not even created in the second attempt. This behavior seems to be what somewhat
reproducible, cf. my earlier email where the part-11 file disappeared as well.
>>>> 
>>>> Thanks again for your help.
>>>> 
>>>> Cheers,
>>>> Max
>>>> —
>>>> Maximilian Bode * Junior Consultant * maximilian.bode@tngtech.com
>>>> TNG Technology Consulting GmbH, Betastr. 13a, 85774 Unterföhring
>>>> Geschäftsführer: Henrik Klagges, Christoph Stock, Dr. Robert Dahlke
>>>> Sitz: Unterföhring * Amtsgericht München * HRB 135082
>>>> 
>>>>> Am 08.03.2016 um 11:05 schrieb Aljoscha Krettek <aljoscha@apache.org>:
>>>>> 
>>>>> Hi,
>>>>> are you taking the “.valid-length” files into account. The problem
with doing “exactly-once” with HDFS is that before Hadoop 2.7 it was not possible to truncate
files. So the trick we’re using is to write the length up to which a file is valid if we
would normally need to truncate it. (If the job fails in the middle of writing the output
files have to be truncated to a valid position.) For example, say you have an output file
part-8-0. Now, if there exists a file part-8-0.valid-length this file tells you up to which
position the file part-8-0 is valid. So you should only read up to this point.
>>>>> 
>>>>> The name of the “.valid-length” suffix can also be configured, by
the way, as can all the other stuff.
>>>>> 
>>>>> If this is not the problem then I definitely have to investigate further.
I’ll also look into the Hadoop 2.4.1 build problem.
>>>>> 
>>>>> Cheers,
>>>>> Aljoscha
>>>>>> On 08 Mar 2016, at 10:26, Maximilian Bode <maximilian.bode@tngtech.com>
wrote:
>>>>>> 
>>>>>> Hi Aljoscha,
>>>>>> thanks again for getting back to me. I built from your branch and
the exception is not occurring anymore. The RollingSink state can be restored.
>>>>>> 
>>>>>> Still, the exactly-once guarantee seems not to be fulfilled, there
are always some extra records after killing either a task manager or the job manager. Do you
have an idea where this behavior might be coming from? (I guess concrete numbers will not
help greatly as there are so many parameters influencing them. Still, in our test scenario,
we produce 2 million records in a Kafka queue but in the final output files there are on the
order of 2.1 million records, so a 5% error. The job is running in a per-job YARN session
with n=3, s=4 with a checkpointing interval of 10s.)
>>>>>> 
>>>>>> On another (maybe unrelated) note: when I pulled your branch, the
Travis build did not go through for -Dhadoop.version=2.4.1. I have not looked into this further
as of now, is this one of the tests known to fail sometimes?
>>>>>> 
>>>>>> Cheers,
>>>>>> Max
>>>>>> <travis.log>
>>>>>> —
>>>>>> Maximilian Bode * Junior Consultant * maximilian.bode@tngtech.com
>>>>>> TNG Technology Consulting GmbH, Betastr. 13a, 85774 Unterföhring
>>>>>> Geschäftsführer: Henrik Klagges, Christoph Stock, Dr. Robert Dahlke
>>>>>> Sitz: Unterföhring * Amtsgericht München * HRB 135082
>>>>>> 
>>>>>>> Am 07.03.2016 um 17:20 schrieb Aljoscha Krettek <aljoscha@apache.org>:
>>>>>>> 
>>>>>>> Hi Maximilian,
>>>>>>> sorry for the delay, we where very busy with the release last
week. I had a hunch about the problem but I think I found a fix now. The problem is in snapshot
restore. When restoring, the sink tries to clean up any files that where previously in progress.
If Flink restores to the same snapshot twice in a row then it will try to clean up the leftover
files twice but they are not there anymore, this causes the exception.
>>>>>>> 
>>>>>>> I have a fix in my branch: https://github.com/aljoscha/flink/tree/rolling-sink-fix
>>>>>>> 
>>>>>>> Could you maybe try if this solves your problem? Which version
of Flink are you using? You would have to build from source to try it out. Alternatively I
could build it and put it onto a maven snapshot repository for you to try it out.
>>>>>>> 
>>>>>>> Cheers,
>>>>>>> Aljoscha
>>>>>>>> On 03 Mar 2016, at 14:50, Aljoscha Krettek <aljoscha@apache.org>
wrote:
>>>>>>>> 
>>>>>>>> Hi,
>>>>>>>> did you check whether there are any files at your specified
HDFS output location? If yes, which files are there?
>>>>>>>> 
>>>>>>>> Cheers,
>>>>>>>> Aljoscha
>>>>>>>>> On 03 Mar 2016, at 14:29, Maximilian Bode <maximilian.bode@tngtech.com>
wrote:
>>>>>>>>> 
>>>>>>>>> Just for the sake of completeness: this also happens
when killing a task manager and is therefore probably unrelated to job manager HA.
>>>>>>>>> 
>>>>>>>>>> Am 03.03.2016 um 14:17 schrieb Maximilian Bode <maximilian.bode@tngtech.com>:
>>>>>>>>>> 
>>>>>>>>>> Hi everyone,
>>>>>>>>>> 
>>>>>>>>>> unfortunately, I am running into another problem
trying to establish exactly once guarantees (Kafka -> Flink 1.0.0-rc3 -> HDFS).
>>>>>>>>>> 
>>>>>>>>>> When using
>>>>>>>>>> 
>>>>>>>>>> RollingSink<Tuple3<Integer,Integer,String>>
sink = new RollingSink<Tuple3<Integer,Integer,String>>("hdfs://our.machine.com:8020/hdfs/dir/outbound");
>>>>>>>>>> sink.setBucketer(new NonRollingBucketer());
>>>>>>>>>> output.addSink(sink);
>>>>>>>>>> 
>>>>>>>>>> and then killing the job manager, the new job manager
is unable to restore the old state throwing
>>>>>>>>>> ---
>>>>>>>>>> java.lang.Exception: Could not restore checkpointed
state to operators and functions
>>>>>>>>>> 	at org.apache.flink.streaming.runtime.tasks.StreamTask.restoreState(StreamTask.java:454)
>>>>>>>>>> 	at org.apache.flink.streaming.runtime.tasks.StreamTask.invoke(StreamTask.java:209)
>>>>>>>>>> 	at org.apache.flink.runtime.taskmanager.Task.run(Task.java:559)
>>>>>>>>>> 	at java.lang.Thread.run(Thread.java:744)
>>>>>>>>>> Caused by: java.lang.Exception: Failed to restore
state to function: In-Progress file hdfs://our.machine.com:8020/hdfs/dir/outbound/part-5-0
was neither moved to pending nor is still in progress.
>>>>>>>>>> 	at org.apache.flink.streaming.api.operators.AbstractUdfStreamOperator.restoreState(AbstractUdfStreamOperator.java:168)
>>>>>>>>>> 	at org.apache.flink.streaming.runtime.tasks.StreamTask.restoreState(StreamTask.java:446)
>>>>>>>>>> 	... 3 more
>>>>>>>>>> Caused by: java.lang.RuntimeException: In-Progress
file hdfs://our.machine.com:8020/hdfs/dir/outbound/part-5-0 was neither moved to pending nor
is still in progress.
>>>>>>>>>> 	at org.apache.flink.streaming.connectors.fs.RollingSink.restoreState(RollingSink.java:686)
>>>>>>>>>> 	at org.apache.flink.streaming.connectors.fs.RollingSink.restoreState(RollingSink.java:122)
>>>>>>>>>> 	at org.apache.flink.streaming.api.operators.AbstractUdfStreamOperator.restoreState(AbstractUdfStreamOperator.java:165)
>>>>>>>>>> 	... 4 more
>>>>>>>>>> ---
>>>>>>>>>> I found a resolved issue [1] concerning Hadoop 2.7.1.
We are in fact using 2.4.0 – might this be the same issue?
>>>>>>>>>> 
>>>>>>>>>> Another thing I could think of is that the job is
not configured correctly and there is some sort of timing issue. The checkpoint interval is
10 seconds, everything else was left at default value. Then again, as the NonRollingBucketer
is used, there should not be any timing issues, right?
>>>>>>>>>> 
>>>>>>>>>> Cheers,
>>>>>>>>>> Max
>>>>>>>>>> 
>>>>>>>>>> [1] https://issues.apache.org/jira/browse/FLINK-2979
>>>>>>>>>> 
>>>>>>>>>> —
>>>>>>>>>> Maximilian Bode * Junior Consultant * maximilian.bode@tngtech.com
>>>>>>>>>> TNG Technology Consulting GmbH, Betastr. 13a, 85774
Unterföhring
>>>>>>>>>> Geschäftsführer: Henrik Klagges, Christoph Stock,
Dr. Robert Dahlke
>>>>>>>>>> Sitz: Unterföhring * Amtsgericht München * HRB
135082
>>>>>>>>>> 
>>>>>>>>> 
>>>>>>>> 
>>>>>>> 
>>>>>> 
>>>>> 
>>>> 
>>> 
>> 
> 


Mime
View raw message