You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@dolphinscheduler.apache.org by GitBox <gi...@apache.org> on 2020/12/24 11:54:03 UTC
[GitHub] [incubator-dolphinscheduler] lenboo opened a new issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
lenboo opened a new issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304
Backgroud:
refer: https://github.com/apache/incubator-dolphinscheduler/issues/4083
At present, the problems of master:
There are many polling, that result in unnecessary time-cost
The distributed lock is used when the command is taken, that result in the bottleneck of concurrency
Too many threads(nProcessInstance*nTaskInstances) are used, that result in the waste of system resources
Polling database result in database query pressure bottleneck
Solution：
##1. Command processing
`The API sends command to the scheduler through netty. If the sending fails, try sending again three times`
*Add API, master and scheduler communication process
*API receives the execution workflow command and sends commnad message to the scheduler (synchronization). If it fails, it will try sending again three times and throw up the failure message if it fails three times.
*The API receives the pause / stop command and sends a pause / stop message to the master (synchronization). If it fails, it will try to send it again three times and throw a failure message upward if it fails three times
##2. Add scheduler function
`In Ha mode, the received command message queue is processed and sent to the master according to the policy (priority, load...)`
-After receiving the command message, save the command to the DB, cache it to the local cache queue, and reply to the message sender
-Ha's scheduler is implemented in the master service and elects the active process
-By monitoring the ZK path, it is found that after active is down, stand by begins to compete for active
-After the stand by competition reaches active, scan the command table, and then start to process the command message
-Command cache queue processing, according to the (priority, load...) policy, send the CMD to the master, and the master marks the host of the CMD
-Timing function (choose one from master timing)
-The active scheduler is triggered regularly to generate command messages and cache them to the local and DB. The problem is that the tasks will be lost during the active election process.
##3. Add master polling status (subworkflow, dependent task refactoring)
`After the master performs the workflow / task, it sends the result to other master (check whether it is needed) and the dependency and sub workflow need this function`
-(do not do this scheme for the time being) the worker / Master sends the task and workflow execution results to the master, and the master receives the task execution results
-Perform the next task according to the next task relationship
-Send the task result to the specified master according to the dependency
-A polling thread is added inside the master, which is responsible for polling the status (dependency and subworkflow) of the task / workflow required by the workflow of the master.
-Master timing function
-The master triggers regularly, generates command message, and sends it (asynchronously) to active. If acitve receives the task save DB successfully, but fails to reply to the message in time, it hangs up, and two timed CMDS will be generated
##4. Fault tolerant process change
-Master down
-Find the workflow instance that the master is responsible for. Each workflow generates a fault-tolerant command and sends the fault-tolerant commnad to the active scheduler for active processing (1. Save the CMD to DB 2. Modify host = null)
-Find the task instance that the master is responsible for and check whether the worker of the task instance is alive + task start time > worker start time. (the worker hangs up first, the master hangs up later), and changes the task status to "fault tolerant"
-Find the unfinished command and send it to the scheduler for reprocessing
-If the send fails, always try again
-When the master submits a task, it is found that the task is in progress, and it needs to inform the worker of the master to which the task belongs
-If the DAG does not receive the message, it is not responsible for processing the result
-Worker down
-The original scheme: the master who grabs the distributed lock will change the status of all tasks executed on this worker to "fault tolerant"
-New solution: worker hang up - > each master only tolerates its own tasks
##5. Reconstruct the communication function
`Communication and service separation, communication has the function`
-- > send message to the receiver and send message to the sender
-Asynchronous: send thread send message, send thread cache message, receive message and process message, reply to sender command after processing message, remove cache message after receiving command
##6. Master execution process modification
`Master is responsible for all workflow resolution, fault tolerance, status monitoring and task sending`
-The master receives a CMD from the scheduler
-[PI list processing thread] generates PI processing class, caches process instance, and then sends it to PI thread pool for processing. If there is a state change, corresponding processing is performed (sending the next task / modifying workflow state, etc.)
-[PI thread pool] PI processing, parsing parameters, parsing DAG, finding start task, generating start task instance to DB, sending start task data to queue, and the thread ends.
-The master task status queue receives the task / workflow status from the worker / API / master, adds the status data to the queue, and the status queue is processed by the state processing thread pool. The task state of the same workflow can only be processed by one thread.
-The master sends the task thread, gets the task instance from the queue and sends it to the corresponding worker?
-Master timeout monitoring time wheel. All tasks / workflows that need timeout monitoring need to join the timeout monitoring time wheel. If a timeout occurs, the timeout policy will be triggered (1. Alarm information 2. Modify the task status to failure
================================================================
背景：
目前发现master的问题：
出现比较多的轮询，造成不必要的耗时
取command的时候使用了分布式锁，造成并发数量的瓶颈
线程使用过多(nProcessInstance*nTaskInstances) ，造成系统资源浪费的现象
轮询数据库，数据量大的情况下，造成数据库查询压力瓶颈
方案：
## 1. command处理
`api将command通过netty发送给scheduler, 发送失败则重试三次发送`
* 增加api和master、scheduler通信流程
* api收到执行工作流命令，向scheduler（同步）发送commnad消息，如果失败，重试三次发送，三次失败向上抛出失败信息。
* api收到暂停/停止命令，向master（同步）发送暂停/停止消息，如果失败，重试三次发送，三次失败向上抛出失败信息
## 2. 新增scheduler功能
` HA模式，处理收到的command消息队列, 根据策略(优先级，负载...)，发给master`
- 收到command消息, 将command保存到db中，同时缓存到本地缓存队列，回复消息发送方
- HA的Scheduler，在master服务中实现, 选举Active流程
- 通过监听zk路径，发现active宕机以后，stand-by开始竞争active
- stand-by 竞争到active以后，扫描完command表, 再开始处理command消息
- command缓存队列处理，根据(优先级，负载...)策略, 把cmd发送给master,master标记cmd的host.
- 定时功能(和master定时二选一)
-active scheduler定时触发，生成command消息,缓存到本地和DB.问题是在active选举过程中会造成任务丢失。
## 3. 增加master轮询状态（子工作流、依赖任务重构）
` master执行完工作流/任务，（检查是否需要）发送结果给别的master，依赖和子工作流需要这个功能`
-（暂时不做这个方案） worker/master把任务、工作流执行结果发送给master，master收到任务执行结果以后
- 根据任务关系执行下一个任务
- 根据依赖关系，将任务结果发送到指定的master
- master内部增加一个轮询线程，负责轮询master负责的工作流需要的任务/工作流的状态（依赖和子工作流）。
- master定时功能
- master定时触发，生成command消息，（异步）发送至active，acitve如果收到任务保存db成功，但是未及时回复消息，就挂了，就会产生两个定时cmd
## 4. 容错流程改变
- master挂掉
- 找到master负责的工作流实例，每个工作流生成容错的command, 将容错commnad发送给active-scheduler，由active处理（1. 保存cmd到db 2.修改host=null）
- 找到master负责的任务实例，检查任务实例的worker是否活着+任务开始时间>worker启动时间，就容错。(worker先挂，master随后挂)，就把任务状态改为“容错”
- 找到还未处理完的command,发到scheduler进行重新处理
- 如果发送失败，则一直重试
- master提交任务的时候，发现任务正在执行，需要向worker通知任务所属的master.
- master收到任务执行结果，发现如果不是自己负责的DAG，就不处理这个消息
- worker挂掉
- 原方案：抢到分布式锁的master将在此worker上执行的所有任务状态改为“容错”
- 新方案：worker挂掉->每个master只容错自己的任务
## 5. 重构通信功能
` 通信和业务剥离, 通信具有的功能`
- 同步： 发送线程send消息（阻塞） -> 接收方收到消息，并且处理消息（存db,写数据等）->接收方返回消息给发送方 -> 发送方解除阻塞
- 异步： 发送线程send消息->发送线程缓存消息-> 接收方收到消息，并处理->处理完消息回复发送方command -> 发送方收到command，移除缓存消息
## 6. master执行流程修改
` master负责所有工作流的解析，容错，状态监控，任务发送 `
- master从scheduler收到cmd
- [PI列表处理线程]生成PI处理类，并缓存流程实例，然后发送到PI线程池处理，发现有状态改变，做相应的处理（发送下一个任务/修改工作流状态等）
- [PI线程池] PI处理，解析参数，解析DAG, 找到开始任务，生成开始任务实例到db,发送开始任务数据到queue，线程结束。
- master任务状态队列，从worker/api/master收到任务/工作流状态，将状态数据加入到队列中，由[状态处理线程池]处理状态队列。同一个工作流的任务状态只能用一个线程处理。
- master发送任务线程，从queue获取任务实例，发送给对应的worker. 要全局spi?
- master超时监控时间轮，所有需要超时监控的任务/工作流，都需要加入超时监控时间轮，如果发生超时，则触发超时策略（1.告警信息 2.修改任务状态为失败）
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] licongyang commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
licongyang commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-753871370
Even if the Scheduler uses HA mde, it also breaks down distributed decentralization.It's going to be under all the pressure that the master was under before, and it's likely to be introduced migration，data consistency ,maintenance,etc. It can take an optional plug-in, and event if it's not introduced, it won't block the entire process. Prioritize the most commonly used and simplest scenarios over the more complex ones.
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] lenboo commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
lenboo commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-754547778
> Even if the Scheduler uses HA mde, it also breaks down distributed decentralization.It's going to be under all the pressure that the master was under before, and it's likely to be introduced migration，data consistency ,maintenance,etc. It can take an optional plug-in, and event if it's not introduced, it won't block the entire process. Prioritize the most commonly used and simplest scenarios over the more complex ones.
1. master is also distributed decentralization;
2. scheduler can take the command distribution task, master does not need a distributed lock, which can relieve the master pressure;
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] CalvinKirs commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
CalvinKirs commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-751315889
full of expectation
On 12/26/2020 11:51，dailidong<no...@github.com> wrote：
we will discuss the architecture refactoring draft at 7 o'clock tonight (Beijing time), Everyone is welcome to join the meeting.
To join the meeting on Google Meet, click this link: meet.google.com/ttu-pzft-gxy
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub, or unsubscribe.
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] lenboo commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
lenboo commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-754547585
1. master is also distributed decentralization;
2. scheduler can take the command distribution task, master does not need a distributed lock, which can relieve the master pressure;
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] lenboo removed a comment on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
lenboo removed a comment on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-754547585
1. master is also distributed decentralization;
2. scheduler can take the command distribution task, master does not need a distributed lock, which can relieve the master pressure;
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] wen-hemin commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
wen-hemin commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-751161893
Good job.
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] dailidong commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
dailidong commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-751315134
we will discuss the architecture refactoring draft at 7 o'clock tonight (Beijing time), Everyone is welcome to join the meeting.
To join the meeting on Google Meet, click this link: meet.google.com/ttu-pzft-gxy
----------------------------------------------------------------
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
[GitHub] [incubator-dolphinscheduler] lenboo commented on issue #4304: [Improvement] Propose for master refactor: less thread consumption, less time consumption
Posted by GitBox <gi...@apache.org>.
lenboo commented on issue #4304:
URL: https://github.com/apache/incubator-dolphinscheduler/issues/4304#issuecomment-751344406
----------------------------------------------------------------
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