[HDFS-15547] Dynamic disk-level tiering

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: datanode
• Assignee: Leon Gao
• Reporter: Leon Gao
• Created: 29/Aug/20 00:20
• Updated: 07/May/21 12:52

Description

This is a proposal for a new use case based on archival storage, to allow configuring DISK and ARCHIVE storage types on the same device (filesystem) to balance disk IO for disks with different density.

The proposal is to mainly solve two problems:

1) The disk IO of ARCHIVE disks is underutilized. This is normal in many use cases where the data hotness is highly skewed.

2) Over the years, as better/cheaper hard drives showing on the market, a large production environment can have mixed disk densities. For example, in our prod environment, we have 2TB, 4TB, 8TB, and 16TB disks. When putting all different HDDs into the cluster, we should be able to utilize disk capacity and disk IO efficiently for all of them.

When moving blocks from DISK to ARCHIVE, we can prefer the same disk and simply rename the files instead of copying.

Issue Links

• duplicates: HDFS-16009 HDFS tiered storage support

Sub-Tasks

• Allow configuring DISK/ARCHIVE storage types on same device mount: Resolved
• Use Hardlink to move replica between DISK and ARCHIVE storage if on same filesystem mount: Resolved
• Allow configuring DISK/ARCHIVE capacity for individual volumes: Resolved
• Add metrics for how blocks are moved in replaceBlock: Resolved
• RefreshVolume fails when replacing DISK/ARCHIVE vol on same mount: Resolved