Flink engine
Overview of Flink
what is big data
Refers to a collection of data that cannot be acquired, stored, managed, and processed with conventional software tools within a certain period of time.
The difference between batch computing and stream computing
Why stream computing is needed
The real-time nature of big data brings greater value, eg: real-time recommendation, data monitoring
Features of Flink
Exactly-Once
High throughput and low latency, real-time fast
High fault tolerance
Flow batch integration
Streaming/Batch SQL
Flink overall framework
Flink layered framework
SDK layer: support SQL/Table, DataStream(java), Python
Execution engine layer: provide a unified DAG (directed acyclic graph) to describe the pipeline of data processing; the scheduling layer converts the DAG into tasks in a distributed environment; transfer data between tasks through Shuffle
State storage layer: store operator state information
Resource scheduling layer: Flink can support deployment in multiple environments
Flink overall framework
A Flink cluster mainly includes two core components: JM (JobManager), TM (TaskManager)
JM is responsible for the coordination of the entire task, including: scheduling tasks, triggering and coordinating tasks to do checkpoints, and coordinating fault-tolerant recovery. The core has the following three components:
Dispatcher: Receive the job, pull up the JobManager to execute the job, and resume the job after the Job Master hangs up;
Job Master: Manage the entire life cycle of a job, apply for a slot from the Resource Manager, and schedule the task to the corresponding TM;
Resource Manager: Responsible for the management and scheduling of slot resources, Task manager will register with RM after pulling up;
TM is responsible for executing each task of a DataFlow Graph and the buffer and data exchange of data streams.
How Flink achieves stream-batch integration
When stream computing and batch computing are independent:
High labor costs: the logic of the batch and flow systems is similar, but they need to be developed twice;
Data link redundancy: the calculation content itself is consistent, and two sets of logically similar links are used to process, resulting in a certain waste of resources;
Inconsistent data caliber: The two sets of links will more or less produce errors, which will cause troubles for the business side.
为什么可以实现流批一体:
站在 Flink 的角度,Everything is Streams,无边界数据集是一种数据流,可以按照时间分成一个个有界数据集;
而批计算可以看作是流计算的特例,其有界数据集也是一种特殊数据流。
因此,无论是无边界数据集还是有界数据集,Flink都可以支持,并且从API到底层处理都是统一的,实现了流批一体。
流批一体的 Scheduler 层
Scheduler 主要负责将作业的 DAG 转化为在分布式环境中可以执行的 Task;
EAGER模式(Streaming 场景):申请一个作业所需要的全部资源,然后同时调度这个作业的全部 Task,所有的 Task 之间采取 Pipeline 的方式进行通信;
LAZY模式(Batch 场景):先调度上游,等待上游产生数据或结束后再调度下游,类似 Spark 的 Stage 执行模式。
流批一体的 Shuffle Service 层
Shuffle:在分布式计算中,用来连接上下游数据交互的过程叫做 Shuffle。
为了统一在Streaming和Batch模式下的Shuffle架构,Flink实现了一个Pluggable的Shuffle Service框架,抽象出一些公共模块,详情如下
经过在 DataStream 层、Scheduler层、Shuffle Service 层进行改造和优化,Flink已经可以方便地解决流和批场景的问题。