go-schedule

Distributed scheduling: evidence and decision

go-schedule manages independent daemons. Remote control, All Systems, cross-daemon search and portable bundles do not currently provide shared execution ownership or automatic host failover. Clustered execution remains an open product proposal under #20.

On 2026-10-31, the owner explicitly deferred clustered execution during S127. The proposal remains open, and no clustered capability, final rejection or permanent product restriction is implied. No coordination backend or clustered execution semantics have been selected. See the S127 decision and issue dispositions.

The earlier S125 assessment on 2026-10-28 found insufficient recorded real-use evidence to activate architecture and prototype work. That historical evidence-specific hold remains valid as an assessment of the available record; S127 adds the owner's explicit direction to defer the effort.

Evidence register

SourceWhat it establishesWhat remains unknown
Owner proposal on #20, 2026-07-24Product interest in node activity reporting, master/worker job claims, node reliability and optional result callbacks. The fleet-key suggestion is a design idea, not approved security architecture.Actual workload, affected operator, machine/task counts, recovery target, failure tradeoffs and why available alternatives are insufficient.
Connected control center completion and v1.5.0 publication, 2026-09-23Independent-daemon overview, target-specific search/actions and portable configuration shipped.Adoption, operating experience and unmet shared-ownership requirements. Publication alone does not establish those findings.
Demand issue #186 and decision issue #189, inspected 2026-10-28The existing sequence requires actual scenarios, failure semantics and a go/no-go decision. Neither issue contained a real-use report at this assessment.The workload and operator evidence required by #186.
Public issue search and relevant delivery discussions, inspected 2026-10-28Open/closed searches for cluster, distributed, failover and ownership, plus discussions on #20/#146/#174/#183/#184/#185/#282, returned roadmap, design and delivery material rather than a concrete post-release unmet workload. Repository Discussions were disabled.Private use and unreported needs were not assessed. Search results cannot establish that there are no users or no demand.

Available approaches and their limits

NeedAvailable approachLimit relevant to the decision
Operate several machines from one interfaceRemote profiles and All Systems, cross-daemon search and deliberately selected target actions.Each daemon remains authoritative for its own work; a successful action on one target does not create shared ownership across targets.
Reuse automation definitions on another machineBundle export, validation, comparison, preview and apply, with explicit target identity and target-local bindings.Bundles move intent. They do not elect a leader, claim runs, synchronize runtime state or perform automatic failover.
Resume a local schedule after downtimeThe task catch-up policy can run one catch-up invocation or skip the missed window.Recovery occurs on that daemon. Another machine does not acquire ownership while it is unavailable.
Obtain an established scheduler's failover capabilityReview the project's release-bound comparison and the vendor's supported topology, edition and recovery behavior for the actual workload.A documented failover scheduler is not proof that in-flight jobs, arbitrary side effects or every topology recover automatically. Product selection needs the same workload and failure-tradeoff inputs.

The v1.7.0 comparison source, reviewed 2026-10-27, records the official high-availability references for JAMS, ActiveBatch, Control-M, Azure Logic Apps and Resolve, and leaves unsupported conclusions about ScriptRunner and OpCon unknown. Reuse that bounded review when evaluating alternatives; it is not evidence that those systems fail to meet a particular user's needs. That comparison is now published with v1.7.0; its earlier review date remains historical evidence, not a new workload finding.

Supply a concrete workload

Add a report to #186, or provide a private report with a publishable summary if its operational details are sensitive. Do not include credentials, personal data, private addresses or secret-bearing command lines. Record unknown values explicitly instead of inventing estimates or treating the form as an eligibility test for using go-schedule.

Use this outline for each actual scenario:

  1. Source and operator: report date, contact or attributable role, and whether this is observed use, a planned deployment or a design idea. Identify which details may be published.
  2. Job and value: what the job does, who needs its result, the failure or limitation experienced, and a concrete example of the impact.
  3. Current deployment: operating systems, host count, task count, schedule frequency, typical/maximum execution time and required network reachability. State unknowns.
  4. Ownership need: whether work must move between machines, share a queue, survive loss of its current host or stay on one named machine. Identify local files, devices, accounts and other placement constraints.
  5. Recovery expectations: acceptable start delay, outage duration and recovery time; whether running work can be interrupted, restarted or resumed; how an uncertain result is resolved.
  6. Side effects: whether replay is idempotent, how duplicate execution is detected, and which cases tolerate duplicate, missed, delayed or uncertain runs. Explain any irreversible external effect.
  7. Failure cases: required behavior for process crash, host loss, partition, clock disagreement and recovery of an old node. Distinguish observed failures from requirements not yet exercised.
  8. Alternatives tried: independent remote daemons, portable bundles, local catch-up, external orchestration or another scheduler; what each accomplished and the specific remaining gap.
  9. Operating cost: available operator skills, deployment/backup/upgrade expectations and acceptable dependency or support burden. Identify any unmet constraint instead of assigning an invented budget.
  10. Requested outcome: the smallest useful capability and the evidence that would demonstrate its value. Attach sanitized incident or job references where available.

Decision sequence

  1. #186 assesses the supplied actual scenarios, bounds, alternatives and maintenance cost. It must explicitly recommend continued exploration before #187 proceeds.
  2. #187 defines ownership, placement, partitions, clocks, duplicate/missed runs, recovery, uncertainty and single-node behavior. Maintainer review must authorize the prototype before technology selection.
  3. #188 evaluates an isolated, unshipped prototype using established coordination and security components against the approved semantics. It compares a simpler alternative and records operational cost.
  4. #189 records the findings-based go/no-go result for #20. It may document why the process stopped before the later stages. A go result creates a separate implementation roadmap; a stop result records alternatives and the disposition of unstarted work.

S127 records the owner's explicit deferral. #186, #189 and #20 remain open and incomplete; #187 and #188 remain open, deferred, gated and unperformed. Coordinator #146 remains open for its unresolved #20 outcome. No issue closes merely because its deferral is documented.

A later owner revisit or explicit scope revision is required before this effort resumes. An actual workload report can inform that revisit, but does not automatically restart the decision chain. Preserve the sequence above unless the owner explicitly revises it, including maintainer approval of semantics before prototype technology selection. There is no resumption date or scheduled reassessment. The owner may retain the deferral, resume the staged assessment, revise its scope or separately stop the proposal. No such future choice is preselected here.

Edit this page on GitHub

On this page