enhance: Support multiple data disks per node instead of fixed single data directory

[tinswzy]

Description

Currently, WP stores node data (including segment logs) under a single fixed data directory structure.
Even if a node is provisioned with multiple data disks, only one disk can be configured as the data directory. As a result:

• Only one disk is actually utilized
• Other mounted data disks remain unused
• I/O throughput is constrained by a single device
• Storage capacity scaling is limited to one mount point

This limits storage scalability and prevents effective multi-disk utilization.

Problem Statement

Current behavior:

• All segment logs are written under one root data directory (e.g. /data)
• The directory structure is fixed
• There is no mechanism to:
  • Configure multiple mount paths
  • Distribute segment files across disks
  • Balance disk usage

In multi-disk production nodes, this leads to:

• Single-disk bottlenecks
• Uneven disk utilization
• Reduced write/read parallelism

[xiaofan-luan]

is there a reason for multiple disks? or we can just bind all disk use raid0

[tinswzy]

is there a reason for multiple disks? or we can just bind all disk use raid0

This question essentially compares two storage layout strategies commonly used in distributed systems: RAID0 vs multi-directory multi-disk (JBOD).

The following are one-sentence comparisons of the advantages and disadvantages of RAID0 and JBOD.

Node Failure Probability
With RAID0, the failure of any single disk makes the entire array unavailable, while in JBOD a single disk failure does not bring down the whole node; therefore JBOD results in a significantly lower node-level failure probability.

Failure Blast Radius
Because RAID0 uses striping across all disks, the failure of one disk makes all data unavailable, whereas JBOD only loses the data stored on that specific disk, limiting the failure blast radius to approximately 1/N.

Dynamic Disk Expansion (Scalability)
Expanding RAID0 typically requires reshaping or rebuilding the array, which is complex and risky, while JBOD can simply add a new disk directory and immediately start placing new data on it.

Handling Disk Removal or Failure
In RAID0, the failure of a single disk causes the entire array to fail, while in JBOD the failed disk can be taken offline and the system can continue operating with the remaining disks.

Single-File Sequential Throughput
RAID0 allows a single file to be written across multiple disks in parallel, achieving close to N× single-disk throughput, while in JBOD a single file typically only utilizes the bandwidth of one disk.

Multi-File Concurrent Throughput
In log storage or sharded workloads, multiple files can be distributed across different disks and written concurrently, so the overall throughput of JBOD is usually close to that of RAID0.

Data Recovery Cost
When a disk fails in RAID0, the entire array becomes unusable, whereas JBOD only loses the data on the failed disk, significantly reducing the recovery scope and cost.

Operational Complexity
RAID0 requires managing RAID arrays (e.g., via mdadm or RAID controllers), while JBOD only requires managing disk paths and directories, making operations simpler.

System Scalability
The number of disks in a RAID0 array is typically fixed when the array is created, whereas JBOD allows disks to be added dynamically and integrated into the data placement policy, making it more suitable for cloud environments and elastic scaling.

Industry Practice
Several distributed systems — such as Apache Kafka, Apache Cassandra, Apache Hadoop HDFS, ClickHouse, and Elasticsearch — explicitly recommend avoiding RAID0 and instead using multiple data/log directories to manage multiple disks (JBOD) in order to reduce the failure blast radius and simplify scaling and operations.

---

Design Principle (Recommended for the Woodpecker Design Document)
For log-oriented or distributed storage systems, it is generally recommended to adopt multi-disk directories (JBOD) so that reliability is ensured by the system's replication mechanism rather than relying on the RAID layer.
This principle is essentially the shared design philosophy adopted by Kafka, Cassandra, and ClickHouse.

---

I think the greatest benefits are flexibility, operation and maintenance, increased capacity of a single machine, and reduced impact of a single disk failure. Of course, when doing raid0, it is relatively simple to directly remove this node when a hard disk fails.