Coming soon
Features that are built or being built but not released yet. Nothing on this page works today.
Warnings and limits
- Nothing on this page is available yet. Each item moves to the pages it belongs to when the release that carries it is running.
- Plans can change: do not depend on an item before it ships.
Not released
The features below exist in Portal's development branches, or are being built, and are not part of the release running at portal.clusy.io.
Your own buckets for outputs
Register an S3 bucket in your AWS account, or a Cloudflare R2 bucket, and Portal saves your jobs' outputs there instead of in its own storage, with a choice between failing a save the bucket refuses and falling back to Portal's storage. It adds portal storage add, use, remove and policy, and the MCP tools storage, storage_add, storage_remove, storage_use and pull.
Logs kept with no expiry
Each job's log is compacted into one object when the job ends, and kept, instead of each part being deleted 30 days after it was written. It comes with a new layout of Portal's storage, by class of data.
Continue an ended job
A class of storage for checkpoints, and a command that continues a job from the checkpoints Portal keeps after an end it decided (they are kept 7 days already).
A check that a spot job's state can be carried
Before a job keeps running on a spot machine, Portal will check at the job's next pause point whether its state can be carried (nothing is serialized, so it takes milliseconds) and say so in the job's events, instead of finding out when a notice arrives.
Scheduling by state size
Portal will choose spot offers by whether a job's state can be saved within the provider's notice, and place jobs whose state is too large on machines with a longer notice or none.
Take the next best GPU on Modal
An option for a Modal job to take the next GPU type that has capacity when the one it asked for has none, instead of ending as no_capacity.
A time limit for suspended jobs
A suspended job will end after a time limit of its own instead of waiting for its checkpoint to expire, and its checkpoint is kept 7 days after that end, as after any end Portal decides.
Deleted runs in Usage
The Usage page will say how many deleted runs a period's totals include.