Re: RE: Spark or Storm
I agree with Cody. Its pretty hard for any framework to provide in built support for that since the semantics completely depends on what data store you want to use it with. Providing interfaces does help a little, but even with those interface, the user still has to do most of the heavy lifting; the user has to understand what is actually going on AND implement all the needed code to ensure unique ID, and the data are atomically updated, according to the capability and APIs provided by the data store. On Fri, Jun 19, 2015 at 7:45 AM, Cody Koeninger wrote: > > http://spark.apache.org/docs/latest/streaming-programming-guide.html#fault-tolerance-semantics > > "semantics of output operations" section > > Is this really not clear? > > As for the general tone of "why doesn't the framework do it for you"... in > my opinion, this is essential complexity for delivery semantics in a > distributed system, not incidental complexity. You need to actually > understand and be responsible for what's going on, unless you're talking > about very narrow use cases (i.e. outputting to a known datastore with > known semantics and schema) > > On Fri, Jun 19, 2015 at 7:26 AM, Ashish Soni > wrote: > >> My understanding for exactly once semantics is it is handled into the >> framework itself but it is not very clear from the documentation , I >> believe documentation needs to be updated with a simple example so that it >> is clear to the end user , This is very critical to decide when some one is >> evaluating the framework and does not have enough time to validate all the >> use cases but to relay on the documentation. >> >> Ashish >> >> On Fri, Jun 19, 2015 at 7:10 AM, bit1...@163.com wrote: >> >>> >>> I think your observation is correct, you have to take care of these >>> replayed data at your end,eg,each message has a unique id or something else. >>> >>> I am using "I think" in the above sentense, because I am not sure and I >>> also have a related question: >>> I am wonderring how direct stream + kakfa is implemented when the Driver >>> is down and restarted, will it always first replay the checkpointed failed >>> batch or will it honor Kafka's offset reset policy(auto.offset.reset). If >>> it honors the reset policy and it is set as "smallest", then it is the at >>> least once semantics; if it set "largest", then it will be at most once >>> semantics? >>> >>> >>> -- >>> bit1...@163.com >>> >>> >>> *From:* Haopu Wang >>> *Date:* 2015-06-19 18:47 >>> *To:* Enno Shioji ; Tathagata Das >>> >>> *CC:* prajod.vettiyat...@wipro.com; Cody Koeninger ; >>> bit1...@163.com; Jordan Pilat ; Will Briggs >>> ; Ashish Soni ; ayan guha >>> ; user@spark.apache.org; Sateesh Kavuri >>> ; Spark Enthusiast ; >>> Sabarish Sasidharan >>> *Subject:* RE: RE: Spark or Storm >>> >>> My question is not directly related: about the "exactly-once semantic", >>> the document (copied below) said spark streaming gives exactly-once >>> semantic, but actually from my test result, with check-point enabled, the >>> application always re-process the files in last batch after gracefully >>> restart. >>> >>> >>> >>> == >>> *Semantics of Received Data* >>> >>> Different input sources provide different guarantees, ranging from *at-least >>> once* to *exactly once*. Read for more details. >>> *With Files* >>> >>> If all of the input data is already present in a fault-tolerant files >>> system like HDFS, Spark Streaming can always recover from any failure and >>> process all the data. This gives *exactly-once* semantics, that all the >>> data will be processed exactly once no matter what fails. >>> >>> >>> >>> >>> -- >>> >>> *From:* Enno Shioji [mailto:eshi...@gmail.com] >>> *Sent:* Friday, June 19, 2015 5:29 PM >>> *To:* Tathagata Das >>> *Cc:* prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; >>> Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; >>> Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan >>> *Subject:* Re: RE: Spark or Storm >>> >>> >>> >>> Fair enough, on second thought, just saying that it should be idempotent >>> is indeed more confusing. >>> &
Re: RE: Spark or Storm
http://spark.apache.org/docs/latest/streaming-programming-guide.html#fault-tolerance-semantics "semantics of output operations" section Is this really not clear? As for the general tone of "why doesn't the framework do it for you"... in my opinion, this is essential complexity for delivery semantics in a distributed system, not incidental complexity. You need to actually understand and be responsible for what's going on, unless you're talking about very narrow use cases (i.e. outputting to a known datastore with known semantics and schema) On Fri, Jun 19, 2015 at 7:26 AM, Ashish Soni wrote: > My understanding for exactly once semantics is it is handled into the > framework itself but it is not very clear from the documentation , I > believe documentation needs to be updated with a simple example so that it > is clear to the end user , This is very critical to decide when some one is > evaluating the framework and does not have enough time to validate all the > use cases but to relay on the documentation. > > Ashish > > On Fri, Jun 19, 2015 at 7:10 AM, bit1...@163.com wrote: > >> >> I think your observation is correct, you have to take care of these >> replayed data at your end,eg,each message has a unique id or something else. >> >> I am using "I think" in the above sentense, because I am not sure and I >> also have a related question: >> I am wonderring how direct stream + kakfa is implemented when the Driver >> is down and restarted, will it always first replay the checkpointed failed >> batch or will it honor Kafka's offset reset policy(auto.offset.reset). If >> it honors the reset policy and it is set as "smallest", then it is the at >> least once semantics; if it set "largest", then it will be at most once >> semantics? >> >> >> -- >> bit1...@163.com >> >> >> *From:* Haopu Wang >> *Date:* 2015-06-19 18:47 >> *To:* Enno Shioji ; Tathagata Das >> >> *CC:* prajod.vettiyat...@wipro.com; Cody Koeninger ; >> bit1...@163.com; Jordan Pilat ; Will Briggs >> ; Ashish Soni ; ayan guha >> ; user@spark.apache.org; Sateesh Kavuri >> ; Spark Enthusiast ; >> Sabarish >> Sasidharan >> *Subject:* RE: RE: Spark or Storm >> >> My question is not directly related: about the "exactly-once semantic", >> the document (copied below) said spark streaming gives exactly-once >> semantic, but actually from my test result, with check-point enabled, the >> application always re-process the files in last batch after gracefully >> restart. >> >> >> >> == >> *Semantics of Received Data* >> >> Different input sources provide different guarantees, ranging from *at-least >> once* to *exactly once*. Read for more details. >> *With Files* >> >> If all of the input data is already present in a fault-tolerant files >> system like HDFS, Spark Streaming can always recover from any failure and >> process all the data. This gives *exactly-once* semantics, that all the >> data will be processed exactly once no matter what fails. >> >> >> >> >> -- >> >> *From:* Enno Shioji [mailto:eshi...@gmail.com] >> *Sent:* Friday, June 19, 2015 5:29 PM >> *To:* Tathagata Das >> *Cc:* prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; >> Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; >> Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan >> *Subject:* Re: RE: Spark or Storm >> >> >> >> Fair enough, on second thought, just saying that it should be idempotent >> is indeed more confusing. >> >> >> >> I guess the crux of the confusion comes from the fact that people tend to >> assume the work you described (store batch id and skip etc.) is handled by >> the framework, perhaps partly because Storm Trident does handle it (you >> just need to let Storm know if the output operation has succeeded or not, >> and it handles the batch id storing & skipping business). Whenever I >> explain people that one needs to do this additional work you described to >> get end-to-end exactly-once semantics, it usually takes a while to convince >> them. In my limited experience, they tend to interpret "transactional" in >> that sentence to mean that you just have to write to a transactional >> storage like ACID RDB. Pointing them to "Semantics of output operations" is >> usually sufficient though. >> >> >> >> Maybe others like @As
Re: RE: Spark or Storm
auto.offset.reset only applies when there are no starting offsets (either from a checkpoint, or from you providing them explicitly) On Fri, Jun 19, 2015 at 6:10 AM, bit1...@163.com wrote: > > I think your observation is correct, you have to take care of these > replayed data at your end,eg,each message has a unique id or something else. > > I am using "I think" in the above sentense, because I am not sure and I > also have a related question: > I am wonderring how direct stream + kakfa is implemented when the Driver > is down and restarted, will it always first replay the checkpointed failed > batch or will it honor Kafka's offset reset policy(auto.offset.reset). If > it honors the reset policy and it is set as "smallest", then it is the at > least once semantics; if it set "largest", then it will be at most once > semantics? > > > -- > bit1...@163.com > > > *From:* Haopu Wang > *Date:* 2015-06-19 18:47 > *To:* Enno Shioji ; Tathagata Das > *CC:* prajod.vettiyat...@wipro.com; Cody Koeninger ; > bit1...@163.com; Jordan Pilat ; Will Briggs > ; Ashish Soni ; ayan guha > ; user@spark.apache.org; Sateesh Kavuri > ; Spark Enthusiast ; > Sabarish > Sasidharan > *Subject:* RE: RE: Spark or Storm > > My question is not directly related: about the "exactly-once semantic", > the document (copied below) said spark streaming gives exactly-once > semantic, but actually from my test result, with check-point enabled, the > application always re-process the files in last batch after gracefully > restart. > > > > == > *Semantics of Received Data* > > Different input sources provide different guarantees, ranging from *at-least > once* to *exactly once*. Read for more details. > *With Files* > > If all of the input data is already present in a fault-tolerant files > system like HDFS, Spark Streaming can always recover from any failure and > process all the data. This gives *exactly-once* semantics, that all the > data will be processed exactly once no matter what fails. > > > > > -- > > *From:* Enno Shioji [mailto:eshi...@gmail.com] > *Sent:* Friday, June 19, 2015 5:29 PM > *To:* Tathagata Das > *Cc:* prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; > Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; > Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan > *Subject:* Re: RE: Spark or Storm > > > > Fair enough, on second thought, just saying that it should be idempotent > is indeed more confusing. > > > > I guess the crux of the confusion comes from the fact that people tend to > assume the work you described (store batch id and skip etc.) is handled by > the framework, perhaps partly because Storm Trident does handle it (you > just need to let Storm know if the output operation has succeeded or not, > and it handles the batch id storing & skipping business). Whenever I > explain people that one needs to do this additional work you described to > get end-to-end exactly-once semantics, it usually takes a while to convince > them. In my limited experience, they tend to interpret "transactional" in > that sentence to mean that you just have to write to a transactional > storage like ACID RDB. Pointing them to "Semantics of output operations" is > usually sufficient though. > > > > Maybe others like @Ashish can weigh on this; did you interpret it in this > way? > > > > What if we change the statement into: > > "end-to-end exactly-once semantics (if your updates to downstream systems > are idempotent or transactional). To learn how to make your updates > idempotent or transactional, see the "Semantics of output operations" > section in this chapter > <https://spark.apache.org/docs/latest/streaming-programming-guide.html#fault-tolerance-semantics> > " > > > > That way, it's clear that it's not sufficient to merely write to a > "transactional storage" like ACID store. > > > > > > > > > > > > > > > > On Fri, Jun 19, 2015 at 9:08 AM, Tathagata Das > wrote: > > If the current documentation is confusing, we can definitely improve the > documentation. However, I dont not understand why is the term > "transactional" confusing. If your output operation has to add 5, then the > user has to implement the following mechanism > > > > 1. If the unique id of the batch of data is already present in the store, > then skip the update > > 2. Otherwise atomically do both, the update operation as well as store the > unique id of t
Re: RE: Spark or Storm
My understanding for exactly once semantics is it is handled into the framework itself but it is not very clear from the documentation , I believe documentation needs to be updated with a simple example so that it is clear to the end user , This is very critical to decide when some one is evaluating the framework and does not have enough time to validate all the use cases but to relay on the documentation. Ashish On Fri, Jun 19, 2015 at 7:10 AM, bit1...@163.com wrote: > > I think your observation is correct, you have to take care of these > replayed data at your end,eg,each message has a unique id or something else. > > I am using "I think" in the above sentense, because I am not sure and I > also have a related question: > I am wonderring how direct stream + kakfa is implemented when the Driver > is down and restarted, will it always first replay the checkpointed failed > batch or will it honor Kafka's offset reset policy(auto.offset.reset). If > it honors the reset policy and it is set as "smallest", then it is the at > least once semantics; if it set "largest", then it will be at most once > semantics? > > > -- > bit1...@163.com > > > *From:* Haopu Wang > *Date:* 2015-06-19 18:47 > *To:* Enno Shioji ; Tathagata Das > *CC:* prajod.vettiyat...@wipro.com; Cody Koeninger ; > bit1...@163.com; Jordan Pilat ; Will Briggs > ; Ashish Soni ; ayan guha > ; user@spark.apache.org; Sateesh Kavuri > ; Spark Enthusiast ; > Sabarish > Sasidharan > *Subject:* RE: RE: Spark or Storm > > My question is not directly related: about the "exactly-once semantic", > the document (copied below) said spark streaming gives exactly-once > semantic, but actually from my test result, with check-point enabled, the > application always re-process the files in last batch after gracefully > restart. > > > > == > *Semantics of Received Data* > > Different input sources provide different guarantees, ranging from *at-least > once* to *exactly once*. Read for more details. > *With Files* > > If all of the input data is already present in a fault-tolerant files > system like HDFS, Spark Streaming can always recover from any failure and > process all the data. This gives *exactly-once* semantics, that all the > data will be processed exactly once no matter what fails. > > > > > -- > > *From:* Enno Shioji [mailto:eshi...@gmail.com] > *Sent:* Friday, June 19, 2015 5:29 PM > *To:* Tathagata Das > *Cc:* prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; > Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; > Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan > *Subject:* Re: RE: Spark or Storm > > > > Fair enough, on second thought, just saying that it should be idempotent > is indeed more confusing. > > > > I guess the crux of the confusion comes from the fact that people tend to > assume the work you described (store batch id and skip etc.) is handled by > the framework, perhaps partly because Storm Trident does handle it (you > just need to let Storm know if the output operation has succeeded or not, > and it handles the batch id storing & skipping business). Whenever I > explain people that one needs to do this additional work you described to > get end-to-end exactly-once semantics, it usually takes a while to convince > them. In my limited experience, they tend to interpret "transactional" in > that sentence to mean that you just have to write to a transactional > storage like ACID RDB. Pointing them to "Semantics of output operations" is > usually sufficient though. > > > > Maybe others like @Ashish can weigh on this; did you interpret it in this > way? > > > > What if we change the statement into: > > "end-to-end exactly-once semantics (if your updates to downstream systems > are idempotent or transactional). To learn how to make your updates > idempotent or transactional, see the "Semantics of output operations" > section in this chapter > <https://spark.apache.org/docs/latest/streaming-programming-guide.html#fault-tolerance-semantics> > " > > > > That way, it's clear that it's not sufficient to merely write to a > "transactional storage" like ACID store. > > > > > > > > > > > > > > > > On Fri, Jun 19, 2015 at 9:08 AM, Tathagata Das > wrote: > > If the current documentation is confusing, we can definitely improve the > documentation. However, I dont not understand why is the term > "transactional" confusing. If your output operation has to add 5,
Re: RE: Spark or Storm
I think your observation is correct, you have to take care of these replayed data at your end,eg,each message has a unique id or something else. I am using "I think" in the above sentense, because I am not sure and I also have a related question: I am wonderring how direct stream + kakfa is implemented when the Driver is down and restarted, will it always first replay the checkpointed failed batch or will it honor Kafka's offset reset policy(auto.offset.reset). If it honors the reset policy and it is set as "smallest", then it is the at least once semantics; if it set "largest", then it will be at most once semantics? bit1...@163.com From: Haopu Wang Date: 2015-06-19 18:47 To: Enno Shioji; Tathagata Das CC: prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan Subject: RE: RE: Spark or Storm My question is not directly related: about the "exactly-once semantic", the document (copied below) said spark streaming gives exactly-once semantic, but actually from my test result, with check-point enabled, the application always re-process the files in last batch after gracefully restart. == Semantics of Received Data Different input sources provide different guarantees, ranging from at-least once to exactly once. Read for more details. With Files If all of the input data is already present in a fault-tolerant files system like HDFS, Spark Streaming can always recover from any failure and process all the data. This gives exactly-once semantics, that all the data will be processed exactly once no matter what fails. From: Enno Shioji [mailto:eshi...@gmail.com] Sent: Friday, June 19, 2015 5:29 PM To: Tathagata Das Cc: prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan Subject: Re: RE: Spark or Storm Fair enough, on second thought, just saying that it should be idempotent is indeed more confusing. I guess the crux of the confusion comes from the fact that people tend to assume the work you described (store batch id and skip etc.) is handled by the framework, perhaps partly because Storm Trident does handle it (you just need to let Storm know if the output operation has succeeded or not, and it handles the batch id storing & skipping business). Whenever I explain people that one needs to do this additional work you described to get end-to-end exactly-once semantics, it usually takes a while to convince them. In my limited experience, they tend to interpret "transactional" in that sentence to mean that you just have to write to a transactional storage like ACID RDB. Pointing them to "Semantics of output operations" is usually sufficient though. Maybe others like @Ashish can weigh on this; did you interpret it in this way? What if we change the statement into: "end-to-end exactly-once semantics (if your updates to downstream systems are idempotent or transactional). To learn how to make your updates idempotent or transactional, see the "Semantics of output operations" section in this chapter" That way, it's clear that it's not sufficient to merely write to a "transactional storage" like ACID store. On Fri, Jun 19, 2015 at 9:08 AM, Tathagata Das wrote: If the current documentation is confusing, we can definitely improve the documentation. However, I dont not understand why is the term "transactional" confusing. If your output operation has to add 5, then the user has to implement the following mechanism 1. If the unique id of the batch of data is already present in the store, then skip the update 2. Otherwise atomically do both, the update operation as well as store the unique id of the batch. This is pretty much the definition of a transaction. The user has to be aware of the transactional semantics of the data store while implementing this functionality. You CAN argue that this effective makes the whole updating sort-a idempotent, as even if you try doing it multiple times, it will update only once. But that is not what is generally considered as idempotent. Writing a fixed count, not an increment, is usually what is called idempotent. And so just mentioning that the output operation must be idempotent is, in my opinion, more confusing. To take a page out of the Storm / Trident guide, even they call this exact conditional updating of Trident State as "transactional" operation. See "transactional spout" in the Trident State guide - https://storm.apache.org/documentation/Trident-state In the end, I am totally open the suggestions and PRs on how to make the programming guide easier to understand. :) TD On Thu, J
RE: RE: Spark or Storm
My question is not directly related: about the "exactly-once semantic", the document (copied below) said spark streaming gives exactly-once semantic, but actually from my test result, with check-point enabled, the application always re-process the files in last batch after gracefully restart. == Semantics of Received Data Different input sources provide different guarantees, ranging from at-least once to exactly once. Read for more details. With Files If all of the input data is already present in a fault-tolerant files system like HDFS, Spark Streaming can always recover from any failure and process all the data. This gives exactly-once semantics, that all the data will be processed exactly once no matter what fails. From: Enno Shioji [mailto:eshi...@gmail.com] Sent: Friday, June 19, 2015 5:29 PM To: Tathagata Das Cc: prajod.vettiyat...@wipro.com; Cody Koeninger; bit1...@163.com; Jordan Pilat; Will Briggs; Ashish Soni; ayan guha; user@spark.apache.org; Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan Subject: Re: RE: Spark or Storm Fair enough, on second thought, just saying that it should be idempotent is indeed more confusing. I guess the crux of the confusion comes from the fact that people tend to assume the work you described (store batch id and skip etc.) is handled by the framework, perhaps partly because Storm Trident does handle it (you just need to let Storm know if the output operation has succeeded or not, and it handles the batch id storing & skipping business). Whenever I explain people that one needs to do this additional work you described to get end-to-end exactly-once semantics, it usually takes a while to convince them. In my limited experience, they tend to interpret "transactional" in that sentence to mean that you just have to write to a transactional storage like ACID RDB. Pointing them to "Semantics of output operations" is usually sufficient though. Maybe others like @Ashish can weigh on this; did you interpret it in this way? What if we change the statement into: "end-to-end exactly-once semantics (if your updates to downstream systems are idempotent or transactional). To learn how to make your updates idempotent or transactional, see the "Semantics of output operations" section in this chapter <https://spark.apache.org/docs/latest/streaming-programming-guide.html#f ault-tolerance-semantics> " That way, it's clear that it's not sufficient to merely write to a "transactional storage" like ACID store. On Fri, Jun 19, 2015 at 9:08 AM, Tathagata Das wrote: If the current documentation is confusing, we can definitely improve the documentation. However, I dont not understand why is the term "transactional" confusing. If your output operation has to add 5, then the user has to implement the following mechanism 1. If the unique id of the batch of data is already present in the store, then skip the update 2. Otherwise atomically do both, the update operation as well as store the unique id of the batch. This is pretty much the definition of a transaction. The user has to be aware of the transactional semantics of the data store while implementing this functionality. You CAN argue that this effective makes the whole updating sort-a idempotent, as even if you try doing it multiple times, it will update only once. But that is not what is generally considered as idempotent. Writing a fixed count, not an increment, is usually what is called idempotent. And so just mentioning that the output operation must be idempotent is, in my opinion, more confusing. To take a page out of the Storm / Trident guide, even they call this exact conditional updating of Trident State as "transactional" operation. See "transactional spout" in the Trident State guide - https://storm.apache.org/documentation/Trident-state In the end, I am totally open the suggestions and PRs on how to make the programming guide easier to understand. :) TD On Thu, Jun 18, 2015 at 11:47 PM, Enno Shioji wrote: Tbh I find the doc around this a bit confusing. If it says "end-to-end exactly-once semantics (if your updates to downstream systems are idempotent or transactional)", I think most people will interpret it that as long as you use a storage which has atomicity (like MySQL/Postgres etc.), a successful output operation for a given batch (let's say "+ 5") is going to be issued exactly-once against the storage. However, as I understand it that's not what this statement means. What it is saying is, it will always issue "+5" and never, say "+6", because it makes sure a message is processed exactly-once internally. However, it *may* issue "+5" more than once for a given batch, and it is up to the developer to deal with this by eithe
Re: RE: Spark or Storm
tegration-of-spark-streaming.html >>> >>> >>> >>> Note the use of checkpoints to persist the Kafka offsets in Spark >>> Streaming itself, and not in zookeeper. >>> >>> >>> >>> Also this statement:”.. This allows one to build a Spark Streaming + >>> Kafka pipelines with end-to-end exactly-once semantics (if your updates to >>> downstream systems are idempotent or transactional).” >>> >>> >>> >>> >>> >>> *From:* Cody Koeninger [mailto:c...@koeninger.org] >>> *Sent:* 18 June 2015 19:38 >>> *To:* bit1...@163.com >>> *Cc:* Prajod S Vettiyattil (WT01 - BAS); jrpi...@gmail.com; >>> eshi...@gmail.com; wrbri...@gmail.com; asoni.le...@gmail.com; ayan >>> guha; user; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; >>> sabarish.sasidha...@manthan.com >>> *Subject:* Re: RE: Spark or Storm >>> >>> >>> >>> That general description is accurate, but not really a specific issue of >>> the direct steam. It applies to anything consuming from kafka (or, as >>> Matei already said, any streaming system really). You can't have exactly >>> once semantics, unless you know something more about how you're storing >>> results. >>> >>> >>> >>> For "some unique id", topicpartition and offset is usually the obvious >>> choice, which is why it's important that the direct stream gives you access >>> to the offsets. >>> >>> >>> >>> See https://github.com/koeninger/kafka-exactly-once for more info >>> >>> >>> >>> >>> >>> >>> >>> On Thu, Jun 18, 2015 at 6:47 AM, bit1...@163.com >>> wrote: >>> >>> I am wondering how direct stream api ensures end-to-end exactly once >>> semantics >>> >>> >>> >>> I think there are two things involved: >>> >>> 1. From the spark streaming end, the driver will replay the Offset range >>> when it's down and restarted,which means that the new tasks will process >>> some already processed data. >>> >>> 2. From the user end, since tasks may process already processed data, >>> user end should detect that some data has already been processed,eg, >>> >>> use some unique ID. >>> >>> >>> >>> Not sure if I have understood correctly. >>> >>> >>> >>> >>> -- >>> >>> bit1...@163.com >>> >>> >>> >>> *From:* prajod.vettiyat...@wipro.com >>> >>> *Date:* 2015-06-18 16:56 >>> >>> *To:* jrpi...@gmail.com; eshi...@gmail.com >>> >>> *CC:* wrbri...@gmail.com; asoni.le...@gmail.com; guha.a...@gmail.com; >>> user@spark.apache.org; sateesh.kav...@gmail.com; >>> sparkenthusi...@yahoo.in; sabarish.sasidha...@manthan.com >>> >>> *Subject:* RE: Spark or Storm >>> >>> >>not being able to read from Kafka using multiple nodes >>> >>> >>> >>> > Kafka is plenty capable of doing this.. >>> >>> >>> >>> I faced the same issue before Spark 1.3 was released. >>> >>> >>> >>> The issue was not with Kafka, but with Spark Streaming’s Kafka >>> connector. Before Spark 1.3.0 release one Spark worker would get all the >>> streamed messages. We had to re-partition to distribute the processing. >>> >>> >>> >>> From Spark 1.3.0 release the Spark Direct API for Kafka supported >>> parallel reads from Kafka streamed to Spark workers. See the “Approach 2: >>> Direct Approach” in this page: >>> http://spark.apache.org/docs/1.3.0/streaming-kafka-integration.html. >>> Note that is also mentions zero data loss and exactly once semantics for >>> kafka integration. >>> >>> >>> >>> >>> >>> Prajod >>> >>> >>> >>> *From:* Jordan Pilat [mailto:jrpi...@gmail.com] >>> *Sent:* 18 June 2015 03:57 >>> *To:* Enno Shioji >>> *Cc:* Will Briggs; asoni.le...@gmail.com; ayan guha; user; Sateesh >>> Kavuri; Spark Enthusiast; Sabarish Sasidharan >>> *Subject:* Re: Spark or Storm >>> >>> >>> >>> >not being able to read from Kafka using multiple nodes >>> >>> Kafka is plenty capable of doin
Re: RE: Spark or Storm
If the current documentation is confusing, we can definitely improve the documentation. However, I dont not understand why is the term "transactional" confusing. If your output operation has to add 5, then the user has to implement the following mechanism 1. If the unique id of the batch of data is already present in the store, then skip the update 2. Otherwise atomically do both, the update operation as well as store the unique id of the batch. This is pretty much the definition of a transaction. The user has to be aware of the transactional semantics of the data store while implementing this functionality. You CAN argue that this effective makes the whole updating sort-a idempotent, as even if you try doing it multiple times, it will update only once. But that is not what is generally considered as idempotent. Writing a fixed count, not an increment, is usually what is called idempotent. And so just mentioning that the output operation must be idempotent is, in my opinion, more confusing. To take a page out of the Storm / Trident guide, even they call this exact conditional updating of Trident State as "transactional" operation. See "transactional spout" in the Trident State guide - https://storm.apache.org/documentation/Trident-state In the end, I am totally open the suggestions and PRs on how to make the programming guide easier to understand. :) TD On Thu, Jun 18, 2015 at 11:47 PM, Enno Shioji wrote: > Tbh I find the doc around this a bit confusing. If it says "end-to-end > exactly-once semantics (if your updates to downstream systems are > idempotent or transactional)", I think most people will interpret it that > as long as you use a storage which has atomicity (like MySQL/Postgres > etc.), a successful output operation for a given batch (let's say "+ 5") is > going to be issued exactly-once against the storage. > > However, as I understand it that's not what this statement means. What it > is saying is, it will always issue "+5" and never, say "+6", because it > makes sure a message is processed exactly-once internally. However, it > *may* issue "+5" more than once for a given batch, and it is up to the > developer to deal with this by either making the output operation > idempotent (e.g. "set 5"), or "transactional" (e.g. keep track of batch IDs > and skip already applied batches etc.). > > I wonder if it makes more sense to drop "or transactional" from the > statement, because if you think about it, ultimately what you are asked to > do is to make the writes idempotent even with the "transactional" approach, > & "transactional" is a bit loaded and would be prone to lead to > misunderstandings (even though in fairness, if you read the fault tolerance > chapter it explicitly explains it). > > > > On Fri, Jun 19, 2015 at 2:56 AM, wrote: > >> More details on the Direct API of Spark 1.3 is at the databricks blog: >> https://databricks.com/blog/2015/03/30/improvements-to-kafka-integration-of-spark-streaming.html >> >> >> >> Note the use of checkpoints to persist the Kafka offsets in Spark >> Streaming itself, and not in zookeeper. >> >> >> >> Also this statement:”.. This allows one to build a Spark Streaming + >> Kafka pipelines with end-to-end exactly-once semantics (if your updates to >> downstream systems are idempotent or transactional).” >> >> >> >> >> >> *From:* Cody Koeninger [mailto:c...@koeninger.org] >> *Sent:* 18 June 2015 19:38 >> *To:* bit1...@163.com >> *Cc:* Prajod S Vettiyattil (WT01 - BAS); jrpi...@gmail.com; >> eshi...@gmail.com; wrbri...@gmail.com; asoni.le...@gmail.com; ayan guha; >> user; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; >> sabarish.sasidha...@manthan.com >> *Subject:* Re: RE: Spark or Storm >> >> >> >> That general description is accurate, but not really a specific issue of >> the direct steam. It applies to anything consuming from kafka (or, as >> Matei already said, any streaming system really). You can't have exactly >> once semantics, unless you know something more about how you're storing >> results. >> >> >> >> For "some unique id", topicpartition and offset is usually the obvious >> choice, which is why it's important that the direct stream gives you access >> to the offsets. >> >> >> >> See https://github.com/koeninger/kafka-exactly-once for more info >> >> >> >> >> >> >> >> On Thu, Jun 18, 2015 at 6:47 AM, bit1...@163.com wrote: >> >> I am wondering how direct stream api ensures
Re: RE: Spark or Storm
Tbh I find the doc around this a bit confusing. If it says "end-to-end exactly-once semantics (if your updates to downstream systems are idempotent or transactional)", I think most people will interpret it that as long as you use a storage which has atomicity (like MySQL/Postgres etc.), a successful output operation for a given batch (let's say "+ 5") is going to be issued exactly-once against the storage. However, as I understand it that's not what this statement means. What it is saying is, it will always issue "+5" and never, say "+6", because it makes sure a message is processed exactly-once internally. However, it *may* issue "+5" more than once for a given batch, and it is up to the developer to deal with this by either making the output operation idempotent (e.g. "set 5"), or "transactional" (e.g. keep track of batch IDs and skip already applied batches etc.). I wonder if it makes more sense to drop "or transactional" from the statement, because if you think about it, ultimately what you are asked to do is to make the writes idempotent even with the "transactional" approach, & "transactional" is a bit loaded and would be prone to lead to misunderstandings (even though in fairness, if you read the fault tolerance chapter it explicitly explains it). On Fri, Jun 19, 2015 at 2:56 AM, wrote: > More details on the Direct API of Spark 1.3 is at the databricks blog: > https://databricks.com/blog/2015/03/30/improvements-to-kafka-integration-of-spark-streaming.html > > > > Note the use of checkpoints to persist the Kafka offsets in Spark > Streaming itself, and not in zookeeper. > > > > Also this statement:”.. This allows one to build a Spark Streaming + > Kafka pipelines with end-to-end exactly-once semantics (if your updates to > downstream systems are idempotent or transactional).” > > > > > > *From:* Cody Koeninger [mailto:c...@koeninger.org] > *Sent:* 18 June 2015 19:38 > *To:* bit1...@163.com > *Cc:* Prajod S Vettiyattil (WT01 - BAS); jrpi...@gmail.com; > eshi...@gmail.com; wrbri...@gmail.com; asoni.le...@gmail.com; ayan guha; > user; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; > sabarish.sasidha...@manthan.com > *Subject:* Re: RE: Spark or Storm > > > > That general description is accurate, but not really a specific issue of > the direct steam. It applies to anything consuming from kafka (or, as > Matei already said, any streaming system really). You can't have exactly > once semantics, unless you know something more about how you're storing > results. > > > > For "some unique id", topicpartition and offset is usually the obvious > choice, which is why it's important that the direct stream gives you access > to the offsets. > > > > See https://github.com/koeninger/kafka-exactly-once for more info > > > > > > > > On Thu, Jun 18, 2015 at 6:47 AM, bit1...@163.com wrote: > > I am wondering how direct stream api ensures end-to-end exactly once > semantics > > > > I think there are two things involved: > > 1. From the spark streaming end, the driver will replay the Offset range > when it's down and restarted,which means that the new tasks will process > some already processed data. > > 2. From the user end, since tasks may process already processed data, user > end should detect that some data has already been processed,eg, > > use some unique ID. > > > > Not sure if I have understood correctly. > > > > > -- > > bit1...@163.com > > > > *From:* prajod.vettiyat...@wipro.com > > *Date:* 2015-06-18 16:56 > > *To:* jrpi...@gmail.com; eshi...@gmail.com > > *CC:* wrbri...@gmail.com; asoni.le...@gmail.com; guha.a...@gmail.com; > user@spark.apache.org; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; > sabarish.sasidha...@manthan.com > > *Subject:* RE: Spark or Storm > > >>not being able to read from Kafka using multiple nodes > > > > > Kafka is plenty capable of doing this.. > > > > I faced the same issue before Spark 1.3 was released. > > > > The issue was not with Kafka, but with Spark Streaming’s Kafka connector. > Before Spark 1.3.0 release one Spark worker would get all the streamed > messages. We had to re-partition to distribute the processing. > > > > From Spark 1.3.0 release the Spark Direct API for Kafka supported parallel > reads from Kafka streamed to Spark workers. See the “Approach 2: Direct > Approach” in this page: > http://spark.apache.org/docs/1.3.0/streaming-kafka-integration.html. Note > that is also mentions zero data loss and exactly once semant
RE: RE: Spark or Storm
More details on the Direct API of Spark 1.3 is at the databricks blog: https://databricks.com/blog/2015/03/30/improvements-to-kafka-integration-of-spark-streaming.html Note the use of checkpoints to persist the Kafka offsets in Spark Streaming itself, and not in zookeeper. Also this statement:”.. This allows one to build a Spark Streaming + Kafka pipelines with end-to-end exactly-once semantics (if your updates to downstream systems are idempotent or transactional).” From: Cody Koeninger [mailto:c...@koeninger.org] Sent: 18 June 2015 19:38 To: bit1...@163.com Cc: Prajod S Vettiyattil (WT01 - BAS); jrpi...@gmail.com; eshi...@gmail.com; wrbri...@gmail.com; asoni.le...@gmail.com; ayan guha; user; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; sabarish.sasidha...@manthan.com Subject: Re: RE: Spark or Storm That general description is accurate, but not really a specific issue of the direct steam. It applies to anything consuming from kafka (or, as Matei already said, any streaming system really). You can't have exactly once semantics, unless you know something more about how you're storing results. For "some unique id", topicpartition and offset is usually the obvious choice, which is why it's important that the direct stream gives you access to the offsets. See https://github.com/koeninger/kafka-exactly-once for more info On Thu, Jun 18, 2015 at 6:47 AM, bit1...@163.com<mailto:bit1...@163.com> mailto:bit1...@163.com>> wrote: I am wondering how direct stream api ensures end-to-end exactly once semantics I think there are two things involved: 1. From the spark streaming end, the driver will replay the Offset range when it's down and restarted,which means that the new tasks will process some already processed data. 2. From the user end, since tasks may process already processed data, user end should detect that some data has already been processed,eg, use some unique ID. Not sure if I have understood correctly. bit1...@163.com<mailto:bit1...@163.com> From: prajod.vettiyat...@wipro.com<mailto:prajod.vettiyat...@wipro.com> Date: 2015-06-18 16:56 To: jrpi...@gmail.com<mailto:jrpi...@gmail.com>; eshi...@gmail.com<mailto:eshi...@gmail.com> CC: wrbri...@gmail.com<mailto:wrbri...@gmail.com>; asoni.le...@gmail.com<mailto:asoni.le...@gmail.com>; guha.a...@gmail.com<mailto:guha.a...@gmail.com>; user@spark.apache.org<mailto:user@spark.apache.org>; sateesh.kav...@gmail.com<mailto:sateesh.kav...@gmail.com>; sparkenthusi...@yahoo.in<mailto:sparkenthusi...@yahoo.in>; sabarish.sasidha...@manthan.com<mailto:sabarish.sasidha...@manthan.com> Subject: RE: Spark or Storm >>not being able to read from Kafka using multiple nodes > Kafka is plenty capable of doing this.. I faced the same issue before Spark 1.3 was released. The issue was not with Kafka, but with Spark Streaming’s Kafka connector. Before Spark 1.3.0 release one Spark worker would get all the streamed messages. We had to re-partition to distribute the processing. From Spark 1.3.0 release the Spark Direct API for Kafka supported parallel reads from Kafka streamed to Spark workers. See the “Approach 2: Direct Approach” in this page: http://spark.apache.org/docs/1.3.0/streaming-kafka-integration.html. Note that is also mentions zero data loss and exactly once semantics for kafka integration. Prajod From: Jordan Pilat [mailto:jrpi...@gmail.com<mailto:jrpi...@gmail.com>] Sent: 18 June 2015 03:57 To: Enno Shioji Cc: Will Briggs; asoni.le...@gmail.com<mailto:asoni.le...@gmail.com>; ayan guha; user; Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan Subject: Re: Spark or Storm >not being able to read from Kafka using multiple nodes Kafka is plenty capable of doing this, by clustering together multiple consumer instances into a consumer group. If your topic is sufficiently partitioned, the consumer group can consume the topic in a parallelized fashion. If it isn't, you still have the fault tolerance associated with clustering the consumers. OK JRP On Jun 17, 2015 1:27 AM, "Enno Shioji" mailto:eshi...@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
Re: RE: Spark or Storm
That general description is accurate, but not really a specific issue of the direct steam. It applies to anything consuming from kafka (or, as Matei already said, any streaming system really). You can't have exactly once semantics, unless you know something more about how you're storing results. For "some unique id", topicpartition and offset is usually the obvious choice, which is why it's important that the direct stream gives you access to the offsets. See https://github.com/koeninger/kafka-exactly-once for more info On Thu, Jun 18, 2015 at 6:47 AM, bit1...@163.com wrote: > I am wondering how direct stream api ensures end-to-end exactly once > semantics > > I think there are two things involved: > 1. From the spark streaming end, the driver will replay the Offset range > when it's down and restarted,which means that the new tasks will process > some already processed data. > 2. From the user end, since tasks may process already processed data, user > end should detect that some data has already been processed,eg, > use some unique ID. > > Not sure if I have understood correctly. > > > -- > bit1...@163.com > > > *From:* prajod.vettiyat...@wipro.com > *Date:* 2015-06-18 16:56 > *To:* jrpi...@gmail.com; eshi...@gmail.com > *CC:* wrbri...@gmail.com; asoni.le...@gmail.com; guha.a...@gmail.com; > user@spark.apache.org; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; > sabarish.sasidha...@manthan.com > *Subject:* RE: Spark or Storm > > >>not being able to read from Kafka using multiple nodes > > > > > Kafka is plenty capable of doing this.. > > > > I faced the same issue before Spark 1.3 was released. > > > > The issue was not with Kafka, but with Spark Streaming’s Kafka connector. > Before Spark 1.3.0 release one Spark worker would get all the streamed > messages. We had to re-partition to distribute the processing. > > > > From Spark 1.3.0 release the Spark Direct API for Kafka supported parallel > reads from Kafka streamed to Spark workers. See the “Approach 2: Direct > Approach” in this page: > http://spark.apache.org/docs/1.3.0/streaming-kafka-integration.html. Note > that is also mentions zero data loss and exactly once semantics for kafka > integration. > > > > > > Prajod > > > > *From:* Jordan Pilat [mailto:jrpi...@gmail.com] > *Sent:* 18 June 2015 03:57 > *To:* Enno Shioji > *Cc:* Will Briggs; asoni.le...@gmail.com; ayan guha; user; Sateesh > Kavuri; Spark Enthusiast; Sabarish Sasidharan > *Subject:* Re: Spark or Storm > > > > >not being able to read from Kafka using multiple nodes > > Kafka is plenty capable of doing this, by clustering together multiple > consumer instances into a consumer group. > If your topic is sufficiently partitioned, the consumer group can consume > the topic in a parallelized fashion. > If it isn't, you still have the fault tolerance associated with clustering > the consumers. > > OK > JRP > > On Jun 17, 2015 1:27 AM, "Enno Shioji" 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.sasidha...@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?
Re: RE: Spark or Storm
I am wondering how direct stream api ensures end-to-end exactly once semantics I think there are two things involved: 1. From the spark streaming end, the driver will replay the Offset range when it's down and restarted,which means that the new tasks will process some already processed data. 2. From the user end, since tasks may process already processed data, user end should detect that some data has already been processed,eg, use some unique ID. Not sure if I have understood correctly. bit1...@163.com From: prajod.vettiyat...@wipro.com Date: 2015-06-18 16:56 To: jrpi...@gmail.com; eshi...@gmail.com CC: wrbri...@gmail.com; asoni.le...@gmail.com; guha.a...@gmail.com; user@spark.apache.org; sateesh.kav...@gmail.com; sparkenthusi...@yahoo.in; sabarish.sasidha...@manthan.com Subject: RE: Spark or Storm >>not being able to read from Kafka using multiple nodes > Kafka is plenty capable of doing this.. I faced the same issue before Spark 1.3 was released. The issue was not with Kafka, but with Spark Streaming’s Kafka connector. Before Spark 1.3.0 release one Spark worker would get all the streamed messages. We had to re-partition to distribute the processing. From Spark 1.3.0 release the Spark Direct API for Kafka supported parallel reads from Kafka streamed to Spark workers. See the “Approach 2: Direct Approach” in this page: http://spark.apache.org/docs/1.3.0/streaming-kafka-integration.html. Note that is also mentions zero data loss and exactly once semantics for kafka integration. Prajod From: Jordan Pilat [mailto:jrpi...@gmail.com] Sent: 18 June 2015 03:57 To: Enno Shioji Cc: Will Briggs; asoni.le...@gmail.com; ayan guha; user; Sateesh Kavuri; Spark Enthusiast; Sabarish Sasidharan Subject: Re: Spark or Storm >not being able to read from Kafka using multiple nodes Kafka is plenty capable of doing this, by clustering together multiple consumer instances into a consumer group. If your topic is sufficiently partitioned, the consumer group can consume the topic in a parallelized fashion. If it isn't, you still have the fault tolerance associated with clustering the consumers. OK JRP On Jun 17, 2015 1:27 AM, "Enno Shioji" 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 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 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 wrote: I have a similar scenario where we need to bring data from kinesis to hbase. Data v
