flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aljoscha Krettek <aljos...@apache.org>
Subject Re: Stateful streaming question
Date Thu, 15 Jun 2017 10:16:18 GMT
Hi,

Trying to revive this somewhat older thread: have you made any progress? I think going with
a ProcessFunction that keeps all your state internally and periodically outputs to, say, Elasticsearch
using a sink seems like the way to go? You can do the periodic emission using timers in the
ProcessFunction. 

In your use case, does the data you would store in the Flink managed state have links between
data of different keys? This sounds like it could be a problem when it comes to consistency
when outputting to an external system.

Best,
Aljoscha
> On 17. May 2017, at 14:12, Flavio Pompermaier <pompermaier@okkam.it> wrote:
> 
> Hi to all,
> there are a lot of useful discussion points :)
> 
> I'll try to answer to everybody.
> 
> @Ankit: 
> right now we're using Parquet on HDFS to store thrift objects. Those objects are essentially
structured like
> key
> alternative_key
> list of tuples (representing the state of my Object)
> This model could be potentially modeled as a Monoid and it's very well suited for a stateful
streaming computation where updates to a single key state are not as expansive as a call to
any db to get the current list of tuples and update back that list with for an update (IMHO).
Maybe here I'm overestimating Flink streaming capabilities...
> serialization should be ok using thrift, but Flink advice to use tuples to have better
performance so just after reading the data from disk (as a ThriftObject) we convert them to
its equivalent representation as Tuple3<String, String, List<Tuple4>> version
> Since I currently use Flink to ingest data that (in the end) means adding tuples to my
objects, it would be perfect to have an "online" state of the grouped tuples in order to:
> add/remove tuples to my object very quickly
> from time to time, scan the whole online data (or a part of it) and "translate" it into
one ore more JSON indices (and put them into Elasticsearch)
> @Fabian:
> You're right that batch processes are bot very well suited to work with services that
can fail...if in a map function the remote call fails all the batch job fails...this should
be less problematic with streaming because there's checkpointing and with async IO  is should
be the possibile to add some retry/backoff policies in order to not overload remote services
like db or solr/es indices (maybe it's not already there but it should be possible to add).
Am I wrong?
> 
> @Kostas:
> 
> From what I understood Queryable state is usefult for gets...what if I need to scan the
entire db? For us it could be better do periodically dump the state to RocksDb or HDFS but,
as I already said, I'm not sure if it is safe to start a batch job that reads the dumped data
while, in the meantime, a possible update of this dump could happen...is there any potential
problem to data consistency (indeed tuples within grouped objects have references to other
objects keys)?
> 
> Best,
> Flavio
> 
> On Wed, May 17, 2017 at 10:18 AM, Kostas Kloudas <k.kloudas@data-artisans.com <mailto:k.kloudas@data-artisans.com>>
wrote:
> Hi Flavio,
> 
> For setting the retries, unfortunately there is no such setting yet and, if I am not
wrong, in case of a failure of a request, 
> an exception will be thrown and the job will restart. I am also including Till in the
thread as he may know better.
> 
> For consistency guarantees and concurrency control, this depends on your underlying backend.
But if you want to 
> have end-to-end control, then you could do as Ankit suggested at his point 3), i.e have
a single job for the whole pipeline
>  (if this fits your needs of course). This will allow you to set your own “precedence”
rules for your operations.
> 
> Now finally, there is no way currently to expose the state of a job to another job. The
way to do so is either Queryable
> State, or writing to a Sink. If the problem for having one job is that you emit one element
at a time, you can always group
> elements together and emit downstream less often, in batches.
>  
> Finally, if  you need 2 jobs, you can always use a hybrid solution where you keep your
current state in Flink, and you dump it 
> to a Sink that is queryable once per week for example. The Sink then can be queried at
any time, and data will be at most one 
> week old.
> 
> Thanks,
> Kostas
> 
>> On May 17, 2017, at 9:35 AM, Fabian Hueske <fhueske@gmail.com <mailto:fhueske@gmail.com>>
wrote:
>> 
>> Hi Ankit, just a brief comment on the batch job is easier than streaming job argument.
I'm not sure about that. 
>> I can see that just the batch job might seem easier to implement, but this is only
one part of the whole story. The operational side of using batch is more complex IMO. 
>> You need a tool to ingest your stream, you need storage for the ingested data, you
need a periodic scheduler to kick of your batch job, and you need to take care of failures
if something goes wrong. 
>> The streaming case, this is not needed or the framework does it for you.
>> 
>> Just my 2 cents, Fabian
>> 
>> 2017-05-16 20:58 GMT+02:00 Jain, Ankit <ankit.jain@here.com <mailto:ankit.jain@here.com>>:
>> Hi Flavio,
>> 
>> While you wait on an update from Kostas, wanted to understand the use-case better
and share my thoughts-
>> 
>>  
>> 
>> 1)       Why is current batch mode expensive? Where are you persisting the data after
updates? Way I see it by moving to Flink, you get to use RocksDB(a key-value store) that makes
your lookups faster – probably right now you are using a non-indexed store like S3 maybe?
>> 
>> So, gain is coming from moving to a better persistence store suited to your use-case
than from batch->streaming. Myabe consider just going with a different data store.
>> 
>> IMHO, stream should only be used if you really want to act on the new events in real-time.
It is generally harder to get a streaming job correct than a batch one.
>> 
>>  
>> 
>> 2)       If current setup is expensive due to serialization-deserialization then
that should be fixed by moving to a faster format (maybe AVRO? - I don’t have a lot of expertise
in that). I don’t see how that problem will go away with Flink – so still need to handle
serialization.
>> 
>>  
>> 
>> 3)       Even if you do decide to move to Flink – I think you can do this with
one job, two jobs are not needed. At every incoming event, check the previous state and update/output
to kafka or whatever data store you are using.
>> 
>>  
>> 
>>  
>> 
>> Thanks
>> 
>> Ankit
>> 
>>  
>> 
>> From: Flavio Pompermaier <pompermaier@okkam.it <mailto:pompermaier@okkam.it>>
>> Date: Tuesday, May 16, 2017 at 9:31 AM
>> To: Kostas Kloudas <k.kloudas@data-artisans.com <mailto:k.kloudas@data-artisans.com>>
>> Cc: user <user@flink.apache.org <mailto:user@flink.apache.org>>
>> Subject: Re: Stateful streaming question
>> 
>>  
>> 
>> Hi Kostas,
>> 
>> thanks for your quick response. 
>> 
>> I also thought about using Async IO, I just need to figure out how to correctly handle
parallelism and number of async requests. 
>> 
>> However that's probably the way to go..is it possible also to set a number of retry
attempts/backoff when the async request fails (maybe due to a too busy server)?
>> 
>>  
>> 
>> For the second part I think it's ok to persist the state into RocksDB or HDFS, my
question is indeed about that: is it safe to start reading (with another Flink job) from RocksDB
or HDFS having an updatable state "pending" on it? Should I ensure that state updates are
not possible until the other Flink job hasn't finish to read the persisted data?
>> 
>>  
>> 
>> And another question...I've tried to draft such a processand basically I have the
following code:
>> 
>>  
>> 
>> DataStream<MyGroupedObj> groupedObj = tuples.keyBy(0)
>> 
>>         .flatMap(new RichFlatMapFunction<Tuple4, MyGroupedObj>() {
>> 
>>  
>> 
>>           private transient ValueState<MyGroupedObj> state;
>> 
>>  
>> 
>>           @Override
>> 
>>           public void flatMap(Tuple4 t, Collector<MyGroupedObj> out) throws
Exception {
>> 
>>             MyGroupedObj current = state.value();
>> 
>>             if (current == null) {
>> 
>>               current = new MyGroupedObj();
>> 
>>             }
>> 
>>             ....
>> 
>>            current.addTuple(t);
>> 
>>             ... 
>> 
>>             state.update(current);
>> 
>>             out.collect(current);
>> 
>>           }
>> 
>>           
>> 
>>           @Override
>> 
>>           public void open(Configuration config) {
>> 
>>             ValueStateDescriptor<MyGroupedObj> descriptor =
>> 
>>                       new ValueStateDescriptor<>( "test",TypeInformation.of(MyGroupedObj.class));
>> 
>>               state = getRuntimeContext().getState(descriptor);
>> 
>>           }
>> 
>>         });
>> 
>>     groupedObj.print();
>> 
>>  
>> 
>> but obviously this way I emit the updated object on every update while, actually,
I just want to persist the ValueState somehow (and make it available to another job that runs
one/moth for example). Is that possible?
>> 
>>  
>> 
>>  
>> 
>> On Tue, May 16, 2017 at 5:57 PM, Kostas Kloudas <k.kloudas@data-artisans.com <mailto:k.kloudas@data-artisans.com>>
wrote:
>> 
>> Hi Flavio,
>> 
>>  
>> 
>> From what I understand, for the first part you are correct. You can use Flink’s
internal state to keep your enriched data.
>> 
>> In fact, if you are also querying an external system to enrich your data, it is worth
looking at the AsyncIO feature:
>> 
>>  
>> 
>> https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/stream/asyncio.html
<https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/stream/asyncio.html>
>>  
>> 
>> Now for the second part, currently in Flink you cannot iterate over all registered
keys for which you have state. A pointer 
>> 
>> to look at the may be useful is the queryable state:
>> 
>>  
>> 
>> https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/stream/queryable_state.html
<https://ci.apache.org/projects/flink/flink-docs-release-1.2/dev/stream/queryable_state.html>
>>  
>> 
>> This is still an experimental feature, but let us know your opinion if you use it.
>> 
>>  
>> 
>> Finally, an alternative would be to keep state in Flink, and periodically flush it
to an external storage system, which you can
>> 
>> query at will.
>> 
>>  
>> 
>> Thanks,
>> 
>> Kostas
>> 
>>  
>> 
>>  
>> 
>> On May 16, 2017, at 4:38 PM, Flavio Pompermaier <pompermaier@okkam.it <mailto:pompermaier@okkam.it>>
wrote:
>> 
>>  
>> 
>> Hi to all,
>> 
>> we're still playing with Flink streaming part in order to see whether it can improve
our current batch pipeline.
>> 
>> At the moment, we have a job that translate incoming data (as Row) into Tuple4, groups
them together by the first field and persist the result to disk (using a thrift object). When
we need to add tuples to those grouped objects we need to read again the persisted data, flat
it back to Tuple4, union with the new tuples, re-group by key and finally persist.
>> 
>>  
>> 
>> This is very expansive to do with batch computation while is should pretty straightforward
to do with streaming (from what I understood): I just need to use ListState. Right?
>> 
>> Then, let's say I need to scan all the data of the stateful computation (key and
values), in order to do some other computation, I'd like to know:
>> 
>> how to do that? I.e. create a DataSet/DataSource<Key,Value> from the stateful
data in the stream
>> is there any problem to access the stateful data without stopping incoming data (and
thus possible updates to the states)?
>> Thanks in advance for the support,
>> 
>> Flavio
>> 
>>  
>> 
>>  
>> 
>> 
>> 
>> 
>>  
>> 
>> --
>> 
>> Flavio Pompermaier
>> Development Department
>> 
>> OKKAM S.r.l.
>> Tel. +(39) 0461 1823908 <tel:+39%200461%20182%203908>
> 
> 


Mime
View raw message