Bhupesh, That is true, but in reality watermarks do not solve a design problem in the DAG where data is getting mixed up. All the watermarks do is to convey "start" and "end" within the stream. The start and end control tuples should have the physical operator id, + a monotonically increasing number. Both these are inserted by engine and are not user supplied, i.e. engine takes up the guarantees of idenfying these watermarks. This concept is same as our current start-window and end-window (which has worked well).
Today Apex does not have watermarks, and lets say I am sending "start something", "end something" through another port. I will still need to not mix data in a transform operator down stream. That problem exist today and will continue. Putting filename on every tuple is too much of a performance hit. Secondly a lot of batch operations are not file related (i.e. file to file), they are collection of "data" split into part files (due to performance reason) and grouping/dimensions/event time/... are done based on internals of the file. In case of file to file copy, user should be expected to route the data properly (parallel partition?). Event-time based watermarks needs a separate thread. I am certain that engine will need to be event-time aware, and will need to take this into account for proper layout. Thks Amol *Follow @amolhkekre* *Join us at Apex Big Data World-San Jose <http://www.apexbigdata.com/san-jose.html>, April 4, 2017!* [image: http://www.apexbigdata.com/san-jose-register.html] <http://www.apexbigdata.com/san-jose-register.html> On Sat, Feb 18, 2017 at 8:17 AM, Bhupesh Chawda <bhup...@datatorrent.com> wrote: > Amol, agreed. We can address event time based watermarks once file batch is > done. > Regarding, file batch support: by allowing to partition an input (file) > operator, we are implicitly mixing multiple batches. Even if the user does > not do any transformations, we should be able to write the correct data to > right files at the destination. > > ~ Bhupesh > > > _______________________________________________________ > > Bhupesh Chawda > > Software Engineer > > E: bhup...@datatorrent.com | Twitter: @bhupeshsc > > www.datatorrent.com | apex.apache.org > > > > On Sat, Feb 18, 2017 at 12:26 PM, Amol Kekre <a...@datatorrent.com> wrote: > > > Thomas, > > The watermarks we have in Apex (start-window and end-window) are working > > good. It is fine to take a look at event time, but basic file I/O does > not > > need anything more than start and end. Lets say they are start-something, > > end-something. The main difference here is that the tuples are user > > generated, other than that they should follow similar principle as > > start-window & end-window. The commonality includes > > - dedup of start-st and end-st > > - First start-st passes through > > - Last end-st passes through > > - Engine indentifies them with chronologically increasing number and > source > > > > The only main difference is that an emit of these is user controlled and > > cannot be guaranteed to happen as such. BTW, part files are rarely done > > based on event time, they are almost always split by size. A vast > majority > > of batch cases have hourly files bound by arrival time and not event > time. > > > > Bhupesh, > > Attaching file names to tuples does not scale. If user mixes two batches, > > then the user would need to handle the transformations. Post file batch > > support, we should look at event time support. Unlike file based batches, > > event time will overlap each other, i.e. at a given time at least two (if > > not more) event times will be active. I think the engine will need to be > > event time aware. > > > > Thks > > Amol > > > > > > > > *Follow @amolhkekre* > > *Join us at Apex Big Data World-San Jose > > <http://www.apexbigdata.com/san-jose.html>, April 4, 2017!* > > [image: http://www.apexbigdata.com/san-jose-register.html] > > <http://www.apexbigdata.com/san-jose-register.html> > > > > On Wed, Feb 15, 2017 at 9:07 PM, Thomas Weise <t...@apache.org> wrote: > > > > > I don't think this should be designed based on a simplistic file > > > input-output scenario. It would be good to include a stateful > > > transformation based on event time. > > > > > > More complex pipelines contain stateful transformations that depend on > > > windowing and watermarks. I think we need a watermark concept that is > > based > > > on progress in event time (or other monotonic increasing sequence) that > > > other operators can generically work with. > > > > > > Note that even file input in many cases can produce time based > > watermarks, > > > for example when you read part files that are bound by event time. > > > > > > Thanks, > > > Thomas > > > > > > > > > On Wed, Feb 15, 2017 at 4:02 AM, Bhupesh Chawda < > bhup...@datatorrent.com > > > > > > wrote: > > > > > > > For better understanding the use case for control tuples in batch, I > > am > > > > creating a prototype for a batch application using File Input and > File > > > > Output operators. > > > > > > > > To enable basic batch processing for File IO operators, I am > proposing > > > the > > > > following changes to File input and output operators: > > > > 1. File Input operator emits a watermark each time it opens and > closes > > a > > > > file. These can be "start file" and "end file" watermarks which > include > > > the > > > > corresponding file names. The "start file" tuple should be sent > before > > > any > > > > of the data from that file flows. > > > > 2. File Input operator can be configured to end the application > after a > > > > single or n scans of the directory (a batch). This is where the > > operator > > > > emits the final watermark (the end of application control tuple). > This > > > will > > > > also shutdown the application. > > > > 3. The File output operator handles these control tuples. "Start > file" > > > > initializes the file name for the incoming tuples. "End file" > watermark > > > > forces a finalize on that file. > > > > > > > > The user would be able to enable the operators to send only those > > > > watermarks that are needed in the application. If none of the options > > are > > > > configured, the operators behave as in a streaming application. > > > > > > > > There are a few challenges in the implementation where the input > > operator > > > > is partitioned. In this case, the correlation between the start/end > > for a > > > > file and the data tuples for that file is lost. Hence we need to > > maintain > > > > the filename as part of each tuple in the pipeline. > > > > > > > > The "start file" and "end file" control tuples in this example are > > > > temporary names for watermarks. We can have generic "start batch" / > > "end > > > > batch" tuples which could be used for other use cases as well. The > > Final > > > > watermark is common and serves the same purpose in each case. > > > > > > > > Please let me know your thoughts on this. > > > > > > > > ~ Bhupesh > > > > > > > > > > > > > > > > On Wed, Jan 18, 2017 at 12:22 AM, Bhupesh Chawda < > > > bhup...@datatorrent.com> > > > > wrote: > > > > > > > > > Yes, this can be part of operator configuration. Given this, for a > > user > > > > to > > > > > define a batch application, would mean configuring the connectors > > > (mostly > > > > > the input operator) in the application for the desired behavior. > > > > Similarly, > > > > > there can be other use cases that can be achieved other than batch. > > > > > > > > > > We may also need to take care of the following: > > > > > 1. Make sure that the watermarks or control tuples are consistent > > > across > > > > > sources. Meaning an HDFS sink should be able to interpret the > > watermark > > > > > tuple sent out by, say, a JDBC source. > > > > > 2. In addition to I/O connectors, we should also look at the need > for > > > > > processing operators to understand some of the control tuples / > > > > watermarks. > > > > > For example, we may want to reset the operator behavior on arrival > of > > > > some > > > > > watermark tuple. > > > > > > > > > > ~ Bhupesh > > > > > > > > > > On Tue, Jan 17, 2017 at 9:59 PM, Thomas Weise <t...@apache.org> > > wrote: > > > > > > > > > >> The HDFS source can operate in two modes, bounded or unbounded. If > > you > > > > >> scan > > > > >> only once, then it should emit the final watermark after it is > done. > > > > >> Otherwise it would emit watermarks based on a policy (files names > > > etc.). > > > > >> The mechanism to generate the marks may depend on the type of > source > > > and > > > > >> the user needs to be able to influence/configure it. > > > > >> > > > > >> Thomas > > > > >> > > > > >> > > > > >> On Tue, Jan 17, 2017 at 5:03 AM, Bhupesh Chawda < > > > > bhup...@datatorrent.com> > > > > >> wrote: > > > > >> > > > > >> > Hi Thomas, > > > > >> > > > > > >> > I am not sure that I completely understand your suggestion. Are > > you > > > > >> > suggesting to broaden the scope of the proposal to treat all > > sources > > > > as > > > > >> > bounded as well as unbounded? > > > > >> > > > > > >> > In case of Apex, we treat all sources as unbounded sources. Even > > > > bounded > > > > >> > sources like HDFS file source is treated as unbounded by means > of > > > > >> scanning > > > > >> > the input directory repeatedly. > > > > >> > > > > > >> > Let's consider HDFS file source for example: > > > > >> > In this case, if we treat it as a bounded source, we can define > > > hooks > > > > >> which > > > > >> > allows us to detect the end of the file and send the "final > > > > watermark". > > > > >> We > > > > >> > could also consider HDFS file source as a streaming source and > > > define > > > > >> hooks > > > > >> > which send watermarks based on different kinds of windows. > > > > >> > > > > > >> > Please correct me if I misunderstand. > > > > >> > > > > > >> > ~ Bhupesh > > > > >> > > > > > >> > > > > > >> > On Mon, Jan 16, 2017 at 9:23 PM, Thomas Weise <t...@apache.org> > > > wrote: > > > > >> > > > > > >> > > Bhupesh, > > > > >> > > > > > > >> > > Please see how that can be solved in a unified way using > windows > > > and > > > > >> > > watermarks. It is bounded data vs. unbounded data. In Beam for > > > > >> example, > > > > >> > you > > > > >> > > can use the "global window" and the final watermark to > > accomplish > > > > what > > > > >> > you > > > > >> > > are looking for. Batch is just a special case of streaming > where > > > the > > > > >> > source > > > > >> > > emits the final watermark. > > > > >> > > > > > > >> > > Thanks, > > > > >> > > Thomas > > > > >> > > > > > > >> > > > > > > >> > > On Mon, Jan 16, 2017 at 1:02 AM, Bhupesh Chawda < > > > > >> bhup...@datatorrent.com > > > > >> > > > > > > >> > > wrote: > > > > >> > > > > > > >> > > > Yes, if the user needs to develop a batch application, then > > > batch > > > > >> aware > > > > >> > > > operators need to be used in the application. > > > > >> > > > The nature of the application is mostly controlled by the > > input > > > > and > > > > >> the > > > > >> > > > output operators used in the application. > > > > >> > > > > > > > >> > > > For example, consider an application which needs to filter > > > records > > > > >> in a > > > > >> > > > input file and store the filtered records in another file. > The > > > > >> nature > > > > >> > of > > > > >> > > > this app is to end once the entire file is processed. > > Following > > > > >> things > > > > >> > > are > > > > >> > > > expected of the application: > > > > >> > > > > > > > >> > > > 1. Once the input data is over, finalize the output file > > from > > > > >> .tmp > > > > >> > > > files. - Responsibility of output operator > > > > >> > > > 2. End the application, once the data is read and > > processed - > > > > >> > > > Responsibility of input operator > > > > >> > > > > > > > >> > > > These functions are essential to allow the user to do higher > > > level > > > > >> > > > operations like scheduling or running a workflow of batch > > > > >> applications. > > > > >> > > > > > > > >> > > > I am not sure about intermediate (processing) operators, as > > > there > > > > >> is no > > > > >> > > > change in their functionality for batch use cases. Perhaps, > > > > allowing > > > > >> > > > running multiple batches in a single application may require > > > > similar > > > > >> > > > changes in processing operators as well. > > > > >> > > > > > > > >> > > > ~ Bhupesh > > > > >> > > > > > > > >> > > > On Mon, Jan 16, 2017 at 2:19 PM, Priyanka Gugale < > > > > pri...@apache.org > > > > >> > > > > > >> > > > wrote: > > > > >> > > > > > > > >> > > > > Will it make an impression on user that, if he has a batch > > > > >> usecase he > > > > >> > > has > > > > >> > > > > to use batch aware operators only? If so, is that what we > > > > expect? > > > > >> I > > > > >> > am > > > > >> > > > not > > > > >> > > > > aware of how do we implement batch scenario so this might > > be a > > > > >> basic > > > > >> > > > > question. > > > > >> > > > > > > > > >> > > > > -Priyanka > > > > >> > > > > > > > > >> > > > > On Mon, Jan 16, 2017 at 12:02 PM, Bhupesh Chawda < > > > > >> > > > bhup...@datatorrent.com> > > > > >> > > > > wrote: > > > > >> > > > > > > > > >> > > > > > Hi All, > > > > >> > > > > > > > > > >> > > > > > While design / implementation for custom control tuples > is > > > > >> > ongoing, I > > > > >> > > > > > thought it would be a good idea to consider its > usefulness > > > in > > > > >> one > > > > >> > of > > > > >> > > > the > > > > >> > > > > > use cases - batch applications. > > > > >> > > > > > > > > > >> > > > > > This is a proposal to adapt / extend existing operators > in > > > the > > > > >> > Apache > > > > >> > > > > Apex > > > > >> > > > > > Malhar library so that it is easy to use them in batch > use > > > > >> cases. > > > > >> > > > > > Naturally, this would be applicable for only a subset of > > > > >> operators > > > > >> > > like > > > > >> > > > > > File, JDBC and NoSQL databases. > > > > >> > > > > > For example, for a file based store, (say HDFS store), > we > > > > could > > > > >> > have > > > > >> > > > > > FileBatchInput and FileBatchOutput operators which allow > > > easy > > > > >> > > > integration > > > > >> > > > > > into a batch application. These operators would be > > extended > > > > from > > > > >> > > their > > > > >> > > > > > existing implementations and would be "Batch Aware", in > > that > > > > >> they > > > > >> > may > > > > >> > > > > > understand the meaning of some specific control tuples > > that > > > > flow > > > > >> > > > through > > > > >> > > > > > the DAG. Start batch and end batch seem to be the > obvious > > > > >> > candidates > > > > >> > > > that > > > > >> > > > > > come to mind. On receipt of such control tuples, they > may > > > try > > > > to > > > > >> > > modify > > > > >> > > > > the > > > > >> > > > > > behavior of the operator - to reinitialize some metrics > or > > > > >> finalize > > > > >> > > an > > > > >> > > > > > output file for example. > > > > >> > > > > > > > > > >> > > > > > We can discuss the potential control tuples and actions > in > > > > >> detail, > > > > >> > > but > > > > >> > > > > > first I would like to understand the views of the > > community > > > > for > > > > >> > this > > > > >> > > > > > proposal. > > > > >> > > > > > > > > > >> > > > > > ~ Bhupesh > > > > >> > > > > > > > > > >> > > > > > > > > >> > > > > > > > >> > > > > > > >> > > > > > >> > > > > > > > > > > > > > > > > > > > >
