Keep¶

Keep is a distributed content-addressable storage system designed for high performance in I/O-bound cluster environments.

Notable design goals and features include:

• High scalability
• Node-level redundancy
• Maximum overall throughput in a busy cluster environment
• Maximum data bandwidth from client to disk
• Minimum transaction overhead
• Elimination of disk thrashing (commonly caused by multiple simultaneous readers)
• Client-controlled redundancy

Design¶

The above goals are accomplished by the following design features.

• Data is transferred directly between the client and the physical node where the disk is installed.
• Data collections are encoded in large (≤64 MiB) blocks to minimize short read/write operations.
• Each disk accepts only one block-read/write operation at a time. This prevents disk thrashing and maximizes total throughput when many clients compete for a disk.
• Storage redundancy is directly controlled, and can be easily verified, by the client simply by reading or writing a block of data on multiple nodes.
• Data block distribution is computed based on the MD5 digest of the data block being stored or retrieved. This eliminates the need for a central or synchronized database of block storage locations.

Components¶

The Keep storage system consists of data block read/write services, SDKs, and management agents.

The responsibilities of the Keep service are:

• Write data blocks
• When writing: ensure data integrity by comparing client-supplied MD5 digest to client-supplied data
• Read data blocks (subject to permission, which is determined by the system/metadata DB)
• Send read/write/error event logs to management agents

The responsibilities of the SDK are:

• When writing: split data into ≤64 MiB chunks
• When writing: encode directory trees as manifests
• When writing: write data to the desired number of nodes to achieve storage redundancy
• After writing: register a collection with Arvados
• When reading: parse manifests
• When reading: verify data integrity by comparing locator to MD5 digest of retrieved data