[
https://issues.apache.org/jira/browse/HDDS-15120?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Ivan Andika updated HDDS-15120:
-------------------------------
Description:
Currently, Ozone supports bucket snapshot which creates a read-only immutable
state of the entire bucket for use cases such as backup, replication,
compliance, etc. These was achieved using the RocksDB checkpoint feature.
However, in the recent rise agentic workloads, there is a need for systems to
implement forking / branching to cater for multi-agents workload. Unlike
snapshot, forks can be mutated. The idea of forking and branch is similar to
Git branch / worktrees where a new "branch" is created based on the base
directory. Multiple agents can fork the same base file system in parallel and
mutate these forks without affecting each other. These forks should also have
zero-copy, similar to snapshot (which should only require O(1) time to create).
Additionally, these forks lifetime can varies (it can be retained for a long
time or discarded quite quickly).
Example systems
* NeonDB branching: https://neon.com/docs/introduction/branching
* Tigris Object Store: https://www.tigrisdata.com/docs/snapshots-and-forks/
(please see the related blogs on the implementations for the idea).
Ozone can consider supporting this feature. With the rise of the storage
compute separation architecture on databases (OLAP or OLTP), the compute /
caching layer can rely on Ozone as the backing store for agentic workloads
since Ozone supports snapshot and forking (they don't need to implement
snapshot and forking or need to write complicated logic to store their forks
state). Ozone can then position itself as the open-source object store /
distributed file system for agentic workloads.
This ticket acts as a way to start a discussion in the community on this
direction. We can start thinking about this (and probably try to start
prototyping some ideas). This might require a radical change of Ozone Manager
design (e.g. might need to introduce versioning, reference counting,
introducing some kind of log subsystems, OM deletions semantic change, etc).
was:
Currently, Ozone supports bucket snapshot which creates a read-only immutable
state of the entire bucket for use cases such as backup, replication,
compliance, etc. These was achieved using the RocksDB checkpoint feature.
However, in the recent rise agentic workloads, there is a need for systems to
implement forking / branching to cater for multi-agents workload. Unlike
snapshot, forks can be mutated. The idea of forking and branch is similar to
Git branch / worktrees where a new "branch" is created based on the base
directory. Multiple agents can fork the same base file system in parallel and
mutate these forks without affecting each other. These forks should also have
zero-copy, similar to snapshot (which should only require O(1) time to create).
Additionally, these forks lifetime can varies (it can be retained for a long
time or discarded quite quickly).
Example systems
* NeonDB branching: https://neon.com/docs/introduction/branching
* Tigris Object Store: https://www.tigrisdata.com/docs/snapshots-and-forks/
(please see the related blogs on the implementations for the idea).
Ozone can consider supporting this feature. With the rise of the storage
compute separation architecture on databases (OLAP or OLTP), the compute /
caching layer can rely on Ozone as the backing store for agentic workloads
since Ozone supports snapshot and forking (they don't need to implement
snapshot and forking or need to write complicated logic to store their forks
state). Ozone can then position itself as the open-source object store /
distributed file system for agentic workloads.
This ticket acts as a way to start a discussion in the community on this
direction. We can start thinking about this (and probably try to start
prototyping some ideas). This might require a radical change of Ozone Manager
design (e.g. might need to introduce versioning, reference counting,
introducing some kind of log subsystems, etc).
> Support bucket forks for agentic workload
> -----------------------------------------
>
> Key: HDDS-15120
> URL: https://issues.apache.org/jira/browse/HDDS-15120
> Project: Apache Ozone
> Issue Type: New Feature
> Reporter: Ivan Andika
> Priority: Major
>
> Currently, Ozone supports bucket snapshot which creates a read-only immutable
> state of the entire bucket for use cases such as backup, replication,
> compliance, etc. These was achieved using the RocksDB checkpoint feature.
> However, in the recent rise agentic workloads, there is a need for systems to
> implement forking / branching to cater for multi-agents workload. Unlike
> snapshot, forks can be mutated. The idea of forking and branch is similar to
> Git branch / worktrees where a new "branch" is created based on the base
> directory. Multiple agents can fork the same base file system in parallel and
> mutate these forks without affecting each other. These forks should also have
> zero-copy, similar to snapshot (which should only require O(1) time to
> create). Additionally, these forks lifetime can varies (it can be retained
> for a long time or discarded quite quickly).
> Example systems
> * NeonDB branching: https://neon.com/docs/introduction/branching
> * Tigris Object Store: https://www.tigrisdata.com/docs/snapshots-and-forks/
> (please see the related blogs on the implementations for the idea).
> Ozone can consider supporting this feature. With the rise of the storage
> compute separation architecture on databases (OLAP or OLTP), the compute /
> caching layer can rely on Ozone as the backing store for agentic workloads
> since Ozone supports snapshot and forking (they don't need to implement
> snapshot and forking or need to write complicated logic to store their forks
> state). Ozone can then position itself as the open-source object store /
> distributed file system for agentic workloads.
> This ticket acts as a way to start a discussion in the community on this
> direction. We can start thinking about this (and probably try to start
> prototyping some ideas). This might require a radical change of Ozone Manager
> design (e.g. might need to introduce versioning, reference counting,
> introducing some kind of log subsystems, OM deletions semantic change, etc).
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
