apache / incubator-uniffle

Uniffle is a high performance, general purpose Remote Shuffle Service.
https://uniffle.apache.org/
Apache License 2.0
387 stars 149 forks source link

[FEATURE] Introduce dynamic disk selection for one partition in LOCALFILE #306

Open zuston opened 2 years ago

zuston commented 2 years ago

Code of Conduct

Search before asking

Describe the feature

Introduce the dynamic disk selection for one partition in LOCALFILE, especially when some disks reached the high-watermark.

Motivation

Currently, when mounting multiple disks in shuffle-server and using the MEMORY_LOCALFILE storage type, one app has a very large partition data and its selected disk reaches the high watermark, it will preserve the data into the memory. This will make some app hard to get the memory.

Describe the solution

Maybe we should let the its partial big partition data store to the another disk.

Additional context

No response

Are you willing to submit PR?

jerqi commented 2 years ago

If you use MEMORY_LOCALFILE_HDFS, it won't be a big problem.

zuston commented 2 years ago

If you use MEMORY_LOCALFILE_HDFS, it won't be a big problem.

Yes. This is an optional solution.

jerqi commented 2 years ago

Actually we ever use raid disk to solve this problem when we only have localfile mode.

maobaolong commented 6 days ago

@jerqi As the raid disk has been tested and shown that there is performance degradation while using raid0/LVM. I will pick this issue up to implement a multiply disk store approach for one partition.