spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Tathagata Das <t...@databricks.com>
Subject Re: Spark or Storm
Date Wed, 17 Jun 2015 21:21:30 GMT
To add more information beyond what Matei said and answer the original
question, here are other things to consider when comparing between Spark
Streaming and Storm.

* Unified programming model and semantics - Most occasions you have to
process the same data again in batch jobs. If you have two separate systems
for batch and streaming, its much much harder to share the code. You will
have to deal with different processing models, with their own semantics.
Compare Storm's "join" vs doing an usual batch join, where as Spark and
Spark Streaming share the same join semantics as they are based on same RDD
model underneath.

* Integration with Spark ecosystem - Many people really want to go beyond
basic streaming ETL and into advanced streaming analytics.
  - Combine stream processing with static datasets
  - Apply dynamically updated machine learning models (i.e. offline
learning and online prediction, or even continuous learning and
prediction),
  - Apply DataFrame and SQL operation with streaming
 These things are pretty easy to do with the spark ecosystem

* Operational management - You have to consider the operational cost of
managing two separate systems for batch and stream processing (with their
own deployment models), vs managing one single engine with one deployment
model.

* Performance - According to Intel's independent study, Spark Streaming in
Kafka direct mode can have 2.5-3x throughput than Trident with 0.5GB batch
size. And at that batch size, the latency of Trident is 30 seconds,
compared to 1.5 seconds for Spark Streaming. This is from a talk that Intel
gave in the Spark Summit (https://spark-summit.org/2015/) two days ago.
Slides will be available soon, but here is a sneak peek - throughput -
http://i.imgur.com/u6pf4rB.png   and latency - http://imgur.com/c46MJ4i
I will post the link to the slides when it comes out, hopefully next week.



On Wed, Jun 17, 2015 at 11:55 AM, Matei Zaharia <matei.zaharia@gmail.com>
wrote:

> The major difference is that in Spark Streaming, there's no *need* for a
> TridentState for state inside your computation. All the stateful operations
> (reduceByWindow, updateStateByKey, etc) automatically handle exactly-once
> processing, keeping updates in order, etc. Also, you don't need to run a
> separate transactional system (e.g. MySQL) to store the state.
>
> After your computation runs, if you want to write the final results (e.g.
> the counts you've been tracking) to a storage system, you use one of the
> output operations (saveAsFiles, foreach, etc). Those actually will run in
> order, but some might run multiple times if nodes fail, thus attempting to
> write the same state again.
>
> You can read about how it works in this research paper:
> http://people.csail.mit.edu/matei/papers/2013/sosp_spark_streaming.pdf.
>
> Matei
>
> On Jun 17, 2015, at 11:49 AM, Enno Shioji <eshioji@gmail.com> wrote:
>
> Hi Matei,
>
>
> Ah, can't get more accurate than from the horse's mouth... If you don't
> mind helping me understand it correctly..
>
> From what I understand, Storm Trident does the following (when used with
> Kafka):
> 1) Sit on Kafka Spout and create batches
> 2) Assign global sequential ID to the batches
> 3) Make sure that all result of processed batches are written once to
> TridentState, *in order* (for example, by skipping batches that were
> already applied once, ultimately by using Zookeeper)
>
> TridentState is an interface that you have to implement, and the
> underlying storage has to be transactional for this to work. The necessary
> skipping etc. is handled by Storm.
>
> In case of Spark Streaming, I understand that
> 1) There is no global ordering; e.g. an output operation for batch
> consisting of offset [4,5,6] can be invoked before the operation for offset
> [1,2,3]
> 2) If you wanted to achieve something similar to what TridentState does,
> you'll have to do it yourself (for example using Zookeeper)
>
> Is this a correct understanding?
>
>
>
>
> On Wed, Jun 17, 2015 at 7:14 PM, Matei Zaharia <matei.zaharia@gmail.com>
> wrote:
>
>> This documentation is only for writes to an external system, but all the
>> counting you do within your streaming app (e.g. if you use
>> reduceByKeyAndWindow to keep track of a running count) is exactly-once.
>> When you write to a storage system, no matter which streaming framework you
>> use, you'll have to make sure the writes are idempotent, because the
>> storage system can't know whether you meant to write the same data again or
>> not. But the place where Spark Streaming helps over Storm, etc is for
>> tracking state within your computation. Without that facility, you'd not
>> only have to make sure that writes are idempotent, but you'd have to make
>> sure that updates to your own internal state (e.g. reduceByKeyAndWindow)
>> are exactly-once too.
>>
>> Matei
>>
>>
>> On Jun 17, 2015, at 8:26 AM, Enno Shioji <eshioji@gmail.com> wrote:
>>
>> The thing is, even with that improvement, you still have to make updates
>> idempotent or transactional yourself. If you read
>> http://spark.apache.org/docs/latest/streaming-programming-guide.html#fault-tolerance-semantics
>>
>> that refers to the latest version, it says:
>>
>> Semantics of output operations
>>
>> Output operations (like foreachRDD) have *at-least once* semantics, that
>> is, the transformed data may get written to an external entity more than
>> once in the event of a worker failure. While this is acceptable for saving
>> to file systems using the saveAs***Files operations (as the file will
>> simply get overwritten with the same data), additional effort may be
>> necessary to achieve exactly-once semantics. There are two approaches.
>>
>>    -
>>
>>    *Idempotent updates*: Multiple attempts always write the same data.
>>    For example, saveAs***Files always writes the same data to the
>>    generated files.
>>    -
>>
>>    *Transactional updates*: All updates are made transactionally so that
>>    updates are made exactly once atomically. One way to do this would be the
>>    following.
>>    - Use the batch time (available in foreachRDD) and the partition
>>       index of the transformed RDD to create an identifier. This identifier
>>       uniquely identifies a blob data in the streaming application.
>>       - Update external system with this blob transactionally (that is,
>>       exactly once, atomically) using the identifier. That is, if the identifier
>>       is not already committed, commit the partition data and the identifier
>>       atomically. Else if this was already committed, skip the update.
>>
>>
>> So either you make the update idempotent, or you have to make it
>> transactional yourself, and the suggested mechanism is very similar to what
>> Storm does.
>>
>>
>>
>>
>> On Wed, Jun 17, 2015 at 3:51 PM, Ashish Soni <asoni.learn@gmail.com>
>> wrote:
>>
>>> @Enno
>>> As per the latest version and documentation Spark Streaming does offer
>>> exactly once semantics using improved kafka integration , Not i have not
>>> tested yet.
>>>
>>> Any feedback will be helpful if anyone is tried the same.
>>>
>>> http://koeninger.github.io/kafka-exactly-once/#7
>>>
>>>
>>> https://databricks.com/blog/2015/03/30/improvements-to-kafka-integration-of-spark-streaming.html
>>>
>>>
>>>
>>> On Wed, Jun 17, 2015 at 10:33 AM, Enno Shioji <eshioji@gmail.com> wrote:
>>>
>>>> AFAIK KCL is *supposed* to provide fault tolerance and load balancing
>>>> (plus additionally, elastic scaling unlike Storm), Kinesis providing the
>>>> coordination. My understanding is that it's like a naked Storm worker
>>>> process that can consequently only do map.
>>>>
>>>> I haven't really used it tho, so can't really comment how it compares
>>>> to Spark/Storm. Maybe somebody else will be able to comment.
>>>>
>>>>
>>>>
>>>> On Wed, Jun 17, 2015 at 3:13 PM, ayan guha <guha.ayan@gmail.com> wrote:
>>>>
>>>>> Thanks for this. It's kcl based kinesis application. But because its
>>>>> just a Java application we are thinking to use spark on EMR or storm
for
>>>>> fault tolerance and load balancing. Is it a correct approach?
>>>>> On 17 Jun 2015 23:07, "Enno Shioji" <eshioji@gmail.com> wrote:
>>>>>
>>>>>> Hi Ayan,
>>>>>>
>>>>>> Admittedly I haven't done much with Kinesis, but if I'm not mistaken
>>>>>> you should be able to use their "processor" interface for that. In
this
>>>>>> example, it's incrementing a counter:
>>>>>> https://github.com/awslabs/amazon-kinesis-data-visualization-sample/blob/master/src/main/java/com/amazonaws/services/kinesis/samples/datavis/kcl/CountingRecordProcessor.java
>>>>>>
>>>>>> Instead of incrementing a counter, you could do your transformation
>>>>>> and send it to HBase.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Wed, Jun 17, 2015 at 1:40 PM, ayan guha <guha.ayan@gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> Great discussion!!
>>>>>>>
>>>>>>> One qs about some comment: Also, you can do some processing with
>>>>>>> Kinesis. If all you need to do is straight forward transformation
and you
>>>>>>> are reading from Kinesis to begin with, it might be an easier
option to
>>>>>>> just do the transformation in Kinesis
>>>>>>>
>>>>>>> - Do you mean KCL application? Or some kind of processing
>>>>>>> withinKineis?
>>>>>>>
>>>>>>> Can you kindly share a link? I would definitely pursue this route
as
>>>>>>> our transformations are really simple.
>>>>>>>
>>>>>>> Best
>>>>>>>
>>>>>>> On Wed, Jun 17, 2015 at 10:26 PM, Ashish Soni <asoni.learn@gmail.com
>>>>>>> > wrote:
>>>>>>>
>>>>>>>> My Use case is below
>>>>>>>>
>>>>>>>> We are going to receive lot of event as stream ( basically
Kafka
>>>>>>>> Stream ) and then we need to process and compute
>>>>>>>>
>>>>>>>> Consider you have a phone contract with ATT and every call
/ sms /
>>>>>>>> data useage you do is an event and then it needs  to calculate
your bill on
>>>>>>>> real time basis so when you login to your account you can
see all those
>>>>>>>> variable as how much you used and how much is left and what
is your bill
>>>>>>>> till date ,Also there are different rules which need to be
considered when
>>>>>>>> you calculate the total bill one simple rule will be 0-500
min it is free
>>>>>>>> but above it is $1 a min.
>>>>>>>>
>>>>>>>> How do i maintain a shared state  ( total amount , total
min ,
>>>>>>>> total data etc ) so that i know how much i accumulated at
any given point
>>>>>>>> as events for same phone can go to any node / executor.
>>>>>>>>
>>>>>>>> Can some one please tell me how can i achieve this is spark
as in
>>>>>>>> storm i can have a bolt which can do this ?
>>>>>>>>
>>>>>>>> Thanks,
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> On Wed, Jun 17, 2015 at 4:52 AM, Enno Shioji <eshioji@gmail.com>
>>>>>>>> wrote:
>>>>>>>>
>>>>>>>>> I guess both. In terms of syntax, I was comparing it
with Trident.
>>>>>>>>>
>>>>>>>>> If you are joining, Spark Streaming actually does offer
windowed
>>>>>>>>> join out of the box. We couldn't use this though as our
event stream can
>>>>>>>>> grow "out-of-sync", so we had to implement something
on top of Storm. If
>>>>>>>>> your event streams don't become out of sync, you may
find the built-in join
>>>>>>>>> in Spark Streaming useful. Storm also has a join keyword
but its semantics
>>>>>>>>> are different.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> > Also, what do you mean by "No Back Pressure" ?
>>>>>>>>>
>>>>>>>>> So when a topology is overloaded, Storm is designed so
that it
>>>>>>>>> will stop reading from the source. Spark on the other
hand, will keep
>>>>>>>>> reading from the source and spilling it internally. This
maybe fine, in
>>>>>>>>> fairness, but it does mean you have to worry about the
persistent store
>>>>>>>>> usage in the processing cluster, whereas with Storm you
don't have to worry
>>>>>>>>> because the messages just remain in the data store.
>>>>>>>>>
>>>>>>>>> Spark came up with the idea of rate limiting, but I don't
feel
>>>>>>>>> this is as nice as back pressure because it's very difficult
to tune it
>>>>>>>>> such that you don't cap the cluster's processing power
but yet so that it
>>>>>>>>> will prevent the persistent storage to get used up.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> On Wed, Jun 17, 2015 at 9:33 AM, Spark Enthusiast <
>>>>>>>>> sparkenthusiast@yahoo.in> wrote:
>>>>>>>>>
>>>>>>>>>> When you say Storm, did you mean Storm with Trident
or Storm?
>>>>>>>>>>
>>>>>>>>>> My use case does not have simple transformation.
There are
>>>>>>>>>> complex events that need to be generated by joining
the incoming event
>>>>>>>>>> stream.
>>>>>>>>>>
>>>>>>>>>> Also, what do you mean by "No Back PRessure" ?
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>   On Wednesday, 17 June 2015 11:57 AM, Enno Shioji
<
>>>>>>>>>> eshioji@gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> We've evaluated Spark Streaming vs. Storm and ended
up sticking
>>>>>>>>>> with Storm.
>>>>>>>>>>
>>>>>>>>>> Some of the important draw backs are:
>>>>>>>>>> Spark has no back pressure (receiver rate limit can
alleviate
>>>>>>>>>> this to a certain point, but it's far from ideal)
>>>>>>>>>> There is also no exactly-once semantics. (updateStateByKey
can
>>>>>>>>>> achieve this semantics, but is not practical if you
have any significant
>>>>>>>>>> amount of state because it does so by dumping the
entire state on every
>>>>>>>>>> checkpointing)
>>>>>>>>>>
>>>>>>>>>> There are also some minor drawbacks that I'm sure
will be fixed
>>>>>>>>>> quickly, like no task timeout, not being able to
read from Kafka using
>>>>>>>>>> multiple nodes, data loss hazard with Kafka.
>>>>>>>>>>
>>>>>>>>>> It's also not possible to attain very low latency
in Spark, if
>>>>>>>>>> that's what you need.
>>>>>>>>>>
>>>>>>>>>> The pos for Spark is the concise and IMO more intuitive
syntax,
>>>>>>>>>> especially if you compare it with Storm's Java API.
>>>>>>>>>>
>>>>>>>>>> I admit I might be a bit biased towards Storm tho
as I'm more
>>>>>>>>>> familiar with it.
>>>>>>>>>>
>>>>>>>>>> Also, you can do some processing with Kinesis. If
all you need to
>>>>>>>>>> do is straight forward transformation and you are
reading from Kinesis to
>>>>>>>>>> begin with, it might be an easier option to just
do the transformation in
>>>>>>>>>> Kinesis.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Wed, Jun 17, 2015 at 7:15 AM, Sabarish Sasidharan
<
>>>>>>>>>> sabarish.sasidharan@manthan.com> wrote:
>>>>>>>>>>
>>>>>>>>>> Whatever you write in bolts would be the logic you
want to apply
>>>>>>>>>> on your events. In Spark, that logic would be coded
in map() or similar
>>>>>>>>>> such  transformations and/or actions. Spark doesn't
enforce a structure for
>>>>>>>>>> capturing your processing logic like Storm does.
>>>>>>>>>> Regards
>>>>>>>>>> Sab
>>>>>>>>>> Probably overloading the question a bit.
>>>>>>>>>>
>>>>>>>>>> In Storm, Bolts have the functionality of getting
triggered on
>>>>>>>>>> events. Is that kind of functionality possible with
Spark streaming? During
>>>>>>>>>> each phase of the data processing, the transformed
data is stored to the
>>>>>>>>>> database and this transformed data should then be
sent to a new pipeline
>>>>>>>>>> for further processing
>>>>>>>>>>
>>>>>>>>>> How can this be achieved using Spark?
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> On Wed, Jun 17, 2015 at 10:10 AM, Spark Enthusiast
<
>>>>>>>>>> sparkenthusiast@yahoo.in> wrote:
>>>>>>>>>>
>>>>>>>>>> I have a use-case where a stream of Incoming events
have to be
>>>>>>>>>> aggregated and joined to create Complex events. The
aggregation will have
>>>>>>>>>> to happen at an interval of 1 minute (or less).
>>>>>>>>>>
>>>>>>>>>> The pipeline is :
>>>>>>>>>>                                   send events
>>>>>>>>>>                      enrich event
>>>>>>>>>> Upstream services -------------------> KAFKA --------->
event
>>>>>>>>>> Stream Processor ------------> Complex Event Processor
------------>
>>>>>>>>>> Elastic Search.
>>>>>>>>>>
>>>>>>>>>> From what I understand, Storm will make a very good
ESP and Spark
>>>>>>>>>> Streaming will make a good CEP.
>>>>>>>>>>
>>>>>>>>>> But, we are also evaluating Storm with Trident.
>>>>>>>>>>
>>>>>>>>>> How does Spark Streaming compare with Storm with
Trident?
>>>>>>>>>>
>>>>>>>>>> Sridhar Chellappa
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>   On Wednesday, 17 June 2015 10:02 AM, ayan guha
<
>>>>>>>>>> guha.ayan@gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> I have a similar scenario where we need to bring
data from
>>>>>>>>>> kinesis to hbase. Data volecity is 20k per 10 mins.
Little manipulation of
>>>>>>>>>> data will be required but that's regardless of the
tool so we will be
>>>>>>>>>> writing that piece in Java pojo.
>>>>>>>>>> All env is on aws. Hbase is on a long running EMR
and kinesis on
>>>>>>>>>> a separate cluster.
>>>>>>>>>> TIA.
>>>>>>>>>> Best
>>>>>>>>>> Ayan
>>>>>>>>>> On 17 Jun 2015 12:13, "Will Briggs" <wrbriggs@gmail.com>
wrote:
>>>>>>>>>>
>>>>>>>>>> The programming models for the two frameworks are
conceptually
>>>>>>>>>> rather different; I haven't worked with Storm for
quite some time, but
>>>>>>>>>> based on my old experience with it, I would equate
Spark Streaming more
>>>>>>>>>> with Storm's Trident API, rather than with the raw
Bolt API. Even then,
>>>>>>>>>> there are significant differences, but it's a bit
closer.
>>>>>>>>>>
>>>>>>>>>> If you can share your use case, we might be able
to provide
>>>>>>>>>> better guidance.
>>>>>>>>>>
>>>>>>>>>> Regards,
>>>>>>>>>> Will
>>>>>>>>>>
>>>>>>>>>> On June 16, 2015, at 9:46 PM, asoni.learn@gmail.com
wrote:
>>>>>>>>>>
>>>>>>>>>> Hi All,
>>>>>>>>>>
>>>>>>>>>> I am evaluating spark VS storm ( spark streaming
 ) and i am not
>>>>>>>>>> able to see what is equivalent of Bolt in storm inside
spark.
>>>>>>>>>>
>>>>>>>>>> Any help will be appreciated on this ?
>>>>>>>>>>
>>>>>>>>>> Thanks ,
>>>>>>>>>> Ashish
>>>>>>>>>>
>>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>>>>>>>>> For additional commands, e-mail: user-help@spark.apache.org
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> ---------------------------------------------------------------------
>>>>>>>>>> To unsubscribe, e-mail: user-unsubscribe@spark.apache.org
>>>>>>>>>> For additional commands, e-mail: user-help@spark.apache.org
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Best Regards,
>>>>>>> Ayan Guha
>>>>>>>
>>>>>>
>>>>>>
>>>>
>>>
>>
>>
>
>

Mime
View raw message