Skip to main content

ADR 0013: Runtime Upgrade Improvements

Changelog

  • 2022-01-25: Initial version

Status

Accepted

Context

Currently major runtime updates incur at least one epoch worth of downtime for the transition period. This is suboptimal, and can be improved to allow seamless runtime updates, with some changes to the runtime descriptor and scheduler behavior.

Decision

Implement support for seamless breaking runtime upgrades.

Implementation

Runtime descriptor related changes:

// Runtime represents a runtime.
type Runtime struct { // nolint: maligned
// Deployments specifies the runtime deployments (versions).
Deployments []*VersionInfo `json:"deployments"`

// Version field is relocated to inside the VersionInfo structure.

// Other unchanged fields omitted for brevity.
}

// VersionInfo is the per-runtime version information.
type VersionInfo struct {
// Version of the runtime.
Version version.Version `json:"version"`

// ValidFrom stores the epoch at which, this version is valid.
ValidFrom beacon.EpochTime `json:"valid_from"`

// TEE is the enclave version information, in an enclave provider specific
// format if any.
TEE []byte `json:"tee,omitempty"`
}

The intended workflow here is to:

  • Deploy runtimes with the initial Deployment populated.

  • Update the runtime version via the deployment of a new version of the descriptor with an additional version info entry. Sufficient nodes must upgrade their runtime binary and configuration by the ValidFrom epoch or the runtime will fail to be scheduled (no special handling is done, this is the existing "insufficient nodes" condition).

  • Aborting or altering pending updates via the deployment of a new version of the descriptor with the removed/ammended not-yet-valid Deployments is possible in this design, but perhaps should be forbidden.

  • Altering exisiting Deployments entries is strictly forbidden, except the removal of superceded descriptors.

  • Deploying descriptors with Deployments that will never be valid (as in one that is superceded by a newer version) is strictly forbidden.

The existing node descriptor is a flat vector of Runtime entries containing the runtime ID, version, and TEE information, so no changes are required.

On transition to an epoch where a new version takes effect, the consensus layer MAY prune the descriptor's Deployments field of superceded versions.

The only scheduler and worker side changes are to incorporate the runtime version into scheduling, and to pick the correct deployed version of the runtime to use, both on a once-per-epoch-per-runtime basis.

Consequences

Positive

  • Seamless runtime upgrades will be possible.

  • The code changes required are relatively minimal, and this is likely the simplest possible solution that will work.

Negative

  • It may be overly simplistic.