Keep storage classes

Use cases

  • Partition data among several cloud buckets for legal or financial reasons.
  • Shift data from "hot" to "cool" storage (e.g. SSD to disk) for price/performance tradeoff.
  • Move data from on-line to off-line storage (e.g. Glacier) but maintain provenance.

Requirements

  • arv-put & arv-copy have option to specify storage class.
    • When writing each block, client can specify storage class for the block.
    • Use API to specify that the blocks belonging to a collection should go into a certain storage class.
  • Workbench permits changing storage class on a collection
  • arvados-cwl-runner has options to specify storage classes for intermediate and final output collections.
    • Use API to specify the storage class for the output collection of a container request.
  • TBD: access controls on storage classes, can restrict which users can place collections in which storage class?
  • TBD: rules for de-duplicating blocks across classes? (e.g., if collections with identical data exist in "hot" & "cool" classes, do we really need a copy of the data in "cool" as well as the copy in "hot"?)

Design

A "storage class" is effectively a tagging scheme to specify a group of keep volumes where a block should be preferentially stored.

Generalized from Keep storage tiers (but unlike storage tiers proposal, there is no implied price/performance relationship between classes).

Each keepstore service has access to one or more storage classes. Storage classes are independent. Data assigned to a class may still be sharded among multiple servers. Classes are be identified with labels or uuids instead of integers. The keep services table adds a column which lists which classes are available at which services.

When writing blocks, keepstore recognizes a header X-Keep-Storage-Classes and accepts or denies the block based on whether it can place the block in the designated classes. If not supplied, keepstores should have a default pool. The value of X-Keep-Storage-Classes should be reported in the response.

Each keepstore volume (mount) is associated with a number of storage classes.

Collections may specify a desired set of classes for the blocks in the collection. Keep balance should move blocks to volumes that offer the desired classes. If multiple collections reference the same block and different sets of classes, multiple copies may be required.

Data management policies, such as "move data from hot storage to cool storage after 1 month", should be implemented on top of the keepstore layer with additional tooling/scripts that set storage classes on collections.

Storage classes could be used for moving data into long-term storage (e.g. Glacier, tape backup, etc). As an example, the user would change the storage class to "glacier", which would copy the blocks into offline storage and delete them from the online storage. To retrieve the blocks, the user would change the storage class to "s3". This would fetch the blocks and copy them back to online storage. (TBD: how does the client find out when the data actually becomes available.)

Development tasks

  1. keepstore: configurable class list per volume/mount
  2. keepstore: support x-keep-storage-classes header
  3. apiserver: collections.desired_storage_classes column
  4. keep-balance: compute/report desired and actual storage classes for each block