[GitHub] [hudi] vinothchandar commented on a change in pull request #2181: [HUDI-911] Add Blog about Hudi-Spark decoupling and Flink integration…

[GitBox <gi...@apache.org>]

vinothchandar commented on a change in pull request #2181:
URL: https://github.com/apache/hudi/pull/2181#discussion_r508732124



##########
File path: docs/_posts/2020-10-15-apache-hudi-meets-apache-flink.md
##########
@@ -0,0 +1,196 @@
+---
+title: "Apache Hudi meets Apache Flink"
+excerpt: "The design and latest progress of the integration of Apache Hudi and Apache Flink."
+author: wangxianghu
+category: blog
+---
+
+Apache Hudi is a data lake framework developed and open sourced by Uber. Hudi joined the Apache incubator for incubation in January 2019, and was promoted to the top Apache project in May 2020. It is one of the most popular data lake frameworks.

Review comment:
       "originally developed" or "created at Uber" ? Hudi is very much an Apache project. 

##########
File path: docs/_posts/2020-10-15-apache-hudi-meets-apache-flink.md
##########
@@ -0,0 +1,196 @@
+---
+title: "Apache Hudi meets Apache Flink"
+excerpt: "The design and latest progress of the integration of Apache Hudi and Apache Flink."
+author: wangxianghu
+category: blog
+---
+
+Apache Hudi is a data lake framework developed and open sourced by Uber. Hudi joined the Apache incubator for incubation in January 2019, and was promoted to the top Apache project in May 2020. It is one of the most popular data lake frameworks.
+
+## 1. Why decouple
+
+Hudi has been using Spark as its data processing engine since its birth. If users want to use Hudi as their data lake framework, they must introduce Spark into their platform technology stack. 
+A few years ago, using Spark as a big data processing engine can be said to be very common or even natural. Since Spark can either perform batch processing or use micro-batch to simulate streaming, one engine solves both streaming and batch problems. 
+However, in recent years, with the development of big data technology, Flink, which is also a big data processing engine, has gradually entered people's vision and has occupied a certain market in the field of computing engines. 
+In the big data technology community, forums and other territories, the voice of whether Hudi supports Flink has gradually appeared and has become more frequent. Therefore, it is a valuable thing to make Hudi support the Flink engine, and the first step of integrating the Flink engine is that Hudi and Spark are decoupled.
+
+In addition, looking at the mature, active, and viable frameworks in the big data, all frameworks are elegant in design and can be integrated with other frameworks and leverage each other's expertise. 
+Therefore, decoupling Hudi from Spark and turning it into an engine-independent data lake framework will undoubtedly create more possibilities for the integration of Hudi and other components, allowing Hudi to better integrate into the big data ecosystem.
+
+## 2. Challenges
+
+Hudi's internal use of Spark API is as common as our usual development and use of List. Since the data source reads the data, and finally writes the data to the table, Spark RDD is used as the main data structure everywhere, and even ordinary tools are implemented using the Spark API. 
+It can be said that Hudi is a universal data lake framework implemented by Spark. Hudi also leverages deep Spark functionality like custom partitioning, in-memory caching to implement indexing and file sizing using workload heuristics. 
+For some of these, Flink offers better out-of-box support (e.g using Flink’s state store for indexing) and can in fact, make Hudi approach real-time latencies more and more. 
+
+In addition, the primary engine integrated after this decoupling is Flink. Flink and Spark differ greatly in core abstraction. Spark believes that data is bounded, and its core abstraction is a limited set of data. 
+Flink believes that the essence of data is a stream, and its core abstract DataStream contains various operations on data. Hudi has a streaming first design (record level updates, record level streams), that arguably fit the Flink model more naturally. 
+At the same time, there are multiple RDDs operating at the same time in Hudi, and the processing result of one RDD is combined with another RDD. 
+This difference in abstraction and the reuse of intermediate results during implementation make it difficult for Hudi to use a unified API to operate both RDD and DataStream in terms of decoupling abstraction.
+
+## 3. Decoupling Spark
+In theory, Hudi uses Spark as its computing engine to use Spark's distributed computing power and RDD's rich operator capabilities. Apart from distributed computing power, Hudi uses RDD more as a data structure, and RDD is essentially a bounded data set. 
+Therefore, it is theoretically feasible to replace RDD with List (of course, it may sacrifice performance/scale). In order to ensure the performance and stability of the Hudi Spark version as much as possible. We can keep the setting the bounded data set as the basic operation unit. 
+Hudi's main operation API remains unchanged, and RDD is extracted as a generic type. The Spark engine implementation still uses RDD, and other engines use List or other bounded  data set according to the actual situation.
+
+### Decoupling principle
+1) Unified generics. The input records `JavaRDD<HoodieRecord>`, key of input records `JavaRDD<HoodieKey>`, and result of write operations `JavaRDD<WriteStatus>` used by the Spark API use generic `I,K,O` instead;
+
+2) De-sparkization. All APIs of the abstraction layer must have nothing to do with Spark. Involving specific operations that are difficult to implement in the abstract layer, rewrite them as abstract methods and introduce Spark subclasses.
+
+For example: Hudi uses the `JavaSparkContext#map()` method in many places. To de-spark, you need to hide the `JavaSparkContext`. For this problem, we introduced the `HoodieEngineContext#map()` method, which will block the specific implementation details of `map`, so as to achieve de-sparkization in abstraction.
+
+3) Minimize changes to the abstraction layer to ensure the original function and performance of Hudi;
+
+4) Replace the `JavaSparkContext` with the `HoodieEngineContext` abstract class to provide the running environment context.
+
+In addition, some of the core algorithms in Hudi, like rollbacks, has been redone without the need for computing a workload profile ahead of time, which used to rely on Spark caching. 

Review comment:
       may be link to marker based rollback PR?

##########
File path: docs/_posts/2020-10-15-apache-hudi-meets-apache-flink.md
##########
@@ -0,0 +1,196 @@
+---
+title: "Apache Hudi meets Apache Flink"
+excerpt: "The design and latest progress of the integration of Apache Hudi and Apache Flink."
+author: wangxianghu
+category: blog
+---
+
+Apache Hudi is a data lake framework developed and open sourced by Uber. Hudi joined the Apache incubator for incubation in January 2019, and was promoted to the top Apache project in May 2020. It is one of the most popular data lake frameworks.

Review comment:
       also just denote that Hudi means `Apache Hudi`. Can we say "Apache Hudi (hudi for short) is a data lake ..." 




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org