Skip to content

Mirror canonical cloud task lifecycle and receipts #893

Description

@jack-champagne

Important

Problem — The cloud client and service disagree about liveness labels, terminal fallback, telemetry shapes, and abort support. A task can finish remotely while the client keeps polling or presents no iteration stream.
Approach — Consume the canonical task timeline/status response and make the client mirror a task deterministically through success, failure, abort, and lost-worker terminal paths. Preserve the local run-directory protocol so the Inspector sees equivalent evidence from local and remote work.
Scope — in: remote status parsing, abort contract alignment, event/usage/result mirroring, compatibility fixtures, task provenance display. out: new cloud scheduling or worker infrastructure.
Assumptions — The server timeline and status vocabulary from the parent telemetry slice are available before implementation begins.

Acceptance Criteria

  • The client settles a remote task correctly for completed, failed, aborted, and lost-worker terminal responses without an indefinite poll.
  • Remote timeline events, iterations, pulse metadata, result/usage summary, and terminal receipt are mirrored into the local run contract without duplicate records.
  • A requested abort has one documented idempotent behavior, and the final status poll remains authoritative.
  • Compatibility fixtures prove the client accepts the deployed legacy shape during the server transition and the canonical shape after it.
  • Executor parity tests cover the same terminal and artifact cases for local and remote runs.

Testing Decisions

  • Extend existing cloud-client, remote-executor, telemetry, run-directory, and executor-parity fixtures.
  • Use fake cloud responses for transition coverage; add one live smoke only after the server telemetry slice deploys.

Key Decisions

  • The client never derives scientific completion from a frame or log line; canonical status and terminal evidence decide settlement.
  • Mirroring preserves compatibility with the existing Inspector rather than creating a second remote-only artifact model.

Constraints & Invariants

  • Cloud credentials and authorization errors remain configuration failures, not synthesized run results.
  • This slice does not alter explicit researcher routing.

Prior Art

  • The existing remote executor, cloud client, run-directory writer, and runner receipt contracts.

Source

  • Part of harmoniqs/aws-infra#291.
  • Blocked by harmoniqs/aws-infra#292.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    afkImplementable without human interactionarea:cliarea:schematrack:cloudTrack C — company compute / cloud solve-service

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions