Great to hear 2 different viewpoints, and thanks a lot for your input
Michael. For now, our application perform an etl process where it reads
data from kafka and stores it in HBase and then performs basic enhancement
and pushes data out on a kafka topic.
We have a conflict of opinion here as few people want to go with DStreams
stating that it provides the primitive rdd abstraction and functionality is
better and easier than structured streaming. We don't have any event time
requirement and also not using windowing mechanism, some basic grouping,
enhancement and storing.
Thats why the question was directed towards Structured Streaming vs
Also, when you say,
> Structured streaming is a completely new implementation that does not use
> DStreams at all, but instead directly runs jobs using RDDs
I understand it doesn't it doesn't use Dstream but I thought Structured
Streaming runs jobs on RDD's via dataframes and in the future, if the RDD
abstraction needs to be switched, it will be done by removing RDD with
something else. Please correct me if I understood this wrong.
Thanks & Regards
On Thu, Feb 1, 2018 at 12:12 AM, Michael Armbrust <mich...@databricks.com>
> At this point I recommend that new applications are built using structured
> streaming. The engine was GA-ed as of Spark 2.2 and I know of several very
> large (trillions of records) production jobs that are running in Structured
> Streaming. All of our production pipelines at databricks are written using
> structured streaming as well.
> Regarding the comparison with RDDs: The situation here is different than
> when thinking about batch DataFrames vs. RDDs. DataFrames are "just" a
> higher-level abstraction on RDDs. Structured streaming is a completely new
> implementation that does not use DStreams at all, but instead directly runs
> jobs using RDDs. The advantages over DStreams include:
> - The ability to start and stop individual queries (rather than needing
> to start/stop a separate StreamingContext)
> - The ability to upgrade your stream and still start from an existing
> - Support for working with Spark SQL data sources (json, parquet, etc)
> - Higher level APIs (DataFrames and SQL) and lambda functions (Datasets)
> - Support for event time aggregation
> At this point, with the addition of mapGroupsWithState and
> flatMapGroupsWithState, I think we should be at feature parity with
> DStreams (and the state store does incremental checkpoints that are more
> efficient than the DStream store). However if there are applications you
> are having a hard time porting over, please let us know!
> On Wed, Jan 31, 2018 at 5:42 AM, vijay.bvp <bvpsa...@gmail.com> wrote:
>> here is my two cents, experts please correct me if wrong
>> its important to understand why one over other and for what kind of use
>> case. There might be sometime in future where low level API's are
>> and become legacy but for now in Spark RDD API is the core and low level
>> API, all higher APIs translate to RDD ultimately, and RDD's are
>> these are things that are not supported and this list needs to be
>> with the use case you have.
>> From my experience Structured Streaming is still new and DStreams API is a
>> matured API.
>> some things that are missing or need to explore more.
>> watermarking/windowing based on no of records in a particular window
>> assuming you have watermark and windowing on event time of the data, the
>> resultant dataframe is grouped data set, only thing you can do is run
>> aggregate functions. you can't simply use that output as another dataframe
>> and manipulate. There is a custom aggregator but I feel its limited.
>> There is option to do stateful operations, using GroupState where the
>> function gets iterator of events for that window. This is the closest
>> to StateStore a developer could get.
>> This arbitrary state that programmer could keep across invocations has its
>> limitations as such how much state we could keep?, is that state stored in
>> driver memory? What happens if the spark job fails is this checkpointed or
>> Sent from: http://apache-spark-user-list.1001560.n3.nabble.com/
>> To unsubscribe e-mail: user-unsubscr...@spark.apache.org