[Feature]: new architecture of cubefs in the cloud

[bladehliu]

design the new architecture of running CubeFS in the cloud

Goal: "Achieving Compute-Storage Separation without Performance Degradation in the Cloud"

• leverage the object storage like S3 as the cold data store
• benefit from read caching and moreover, optimize write performance, e.g., a particular case is that the WAL appending should be fast enough for database workload
• work together with the k8s cluster in the cloud, as the shared storage solution to support compute-storage separation for the upper layer databases.

Describe the solution

cubefs cloud = its metadata layer (RAM) + its hot data layer (SSD) + the cold data layer in S3,

note the hot data layer = the newly written data (such as the recent WALs) + the frequently accessed data (such as a file just open and read); the cold data layer = the file extents swapped out from the hot data layer

Describe an alternate solution.

compared with direct read/write from/to S3 (and other filesystems on top of it), cubefs cloud has its very own metadata store and hot data layer for better performance to power the compute-storage separation for online database systems.

Anything else?

note by adding this feature, cubefs will become a hybrid cloud storage solution. #2994

related work: open-source filesystems on top of S3

[Zorlin]

Please don't abandon the local storage use case. I don't want to use cloud storage, I want to build cloud storage, and the current architecture of CubeFS is great for me.

Otherwise, I fully support this. Check out JuiceFS, it does something similar.

[xiaochunhe]

Please don't abandon the local storage use case. I don't want to use cloud storage, I want to build cloud storage, and the current architecture of CubeFS is great for me.

Otherwise, I fully support this. Check out JuiceFS, it does something similar.

local storage will not be abandoned, it will continue to be optimized

[Zorlin]

Please don't abandon the local storage use case. I don't want to use cloud storage, I want to build cloud storage, and the current architecture of CubeFS is great for me.
Otherwise, I fully support this. Check out JuiceFS, it does something similar.

local storage will not be abandoned, it will continue to be optimized

Perfect :) Then I fully support this. Let's make CubeFS more flexible 👯

[awzhgw]

that is a good idea. The need for file-level lifecycle management functions is crucial for cost considerations.

[fengjiannan2010]

正在开发的功能在哪个分支上？

[true1064]

正在开发的功能在哪个分支上？

https://github.com/cubefs/cubefs/tree/develop-hybridcloud-lifecycle