apache / incubator-uniffle

Uniffle is a high performance, general purpose Remote Shuffle Service.
https://uniffle.apache.org/
Apache License 2.0
370 stars 143 forks source link

[FEATURE] Support stage recompute for Spark clients #477

Open advancedxy opened 1 year ago

advancedxy commented 1 year ago

Code of Conduct

Search before asking

Describe the feature

Once Uniffle's shuffle data read fails, the spark client has its chance to recompute the whole stage.

Motivation

In a distributed cluster with large number of shuffler servers, it's common for some node to go down, such as:

  1. node crash/maintenance due to hardware failures or security patches
  2. Pod eviction if deployed in a k8s environment
  3. vm/spot instance eviction if deployed in a cloud environment

Uniffle already has provide a mechanism to overcome this issue: the quorum protocol. But it requires multiple replica of the same shuffle data, which increases the network traffic and memory pressure on shuffle server. And the E2E performance may be degraded due to the replication.

I'd like to provide a new way to alleviate the potential node failure(in rare chance). Once the whole stage could be recompute, the Spark App could be resilient to shuffle server node failure.

Describe the solution

TBD.

A design doc would be added later

Additional context

No response

Are you willing to submit PR?

advancedxy commented 1 year ago

I wrote up a design doc for this issue: https://docs.google.com/document/d/1OGswqDDQ52rpw5Lat-FpEGDfX1T6EFXYLC7L0G5O4WE/edit?usp=sharing

@zuston @xianjingfeng @jerqi could you mind do some design review?

jiafuzha commented 1 year ago

@advancedxy I have some comments in the design doc. Please grant me 'comment' permission.

advancedxy commented 1 year ago

@advancedxy I have some comments in the design doc. Please grant me 'comment' permission.

Changed the default permission from viewer to commenter. Please try to refresh it and check whether you have comment permission or not.

I thought viewer already has the comment permission

jerqi commented 1 year ago

@YutingWang98 Uniffle is a remote shuffle service. We have supported stage retry. Maybe you have interest about it.

connorlwilkes commented 1 year ago

What is remaining on this particular issue Spark side and are there any docs on how this is enabled?

jerqi commented 1 year ago

What is remaining on this particular issue Spark side and are there any docs on how this is enabled?

Wait for the pr https://github.com/apache/incubator-uniffle/pull/1129