•

Concurrent Approval Updates: Stop Stale Decisions

Two steel rods approach a floating socket; one is seated while a slate boundary blocks the offset rod, above a painted chalk landscape.
What’s in this article?

    Keep an older review screen from overwriting a newer decision. Define which decision attempt may take effect, what a stale response means and how operators reconcile an uncertain save.

    Two reviewers open the same request. One rejects it because required evidence is missing. The other, still looking at the earlier screen, clicks approve. If the later save simply replaces the earlier one, the workflow can release work that its current decision should block.

    Concurrent approval updates are an operating-system problem, not just an awkward notification sequence. The control must connect what someone reviewed to the authoritative state when their decision takes effect. This blueprint is a proposed design for process owners and system implementers; its states and tests are not universal software behavior.

    Separate decision policy from update integrity

    Your approval policy determines who may decide and which combination of decisions permits release. Update integrity determines whether an attempted change is still valid when it reaches the record. A workflow needs both.

    A first-response policy does not mean any late response may overwrite the winner. An all-required policy does not mean several people should replace one shared approval field. Preserve each role’s decision where the policy requires it, and validate the aggregate state through a controlled transition.

    Use the broader approval workflow design guide to establish authority and review requirements. The narrower question here is what happens when two valid actors try to change the same decision state at almost the same time.

    A refresh is not a commit safeguard

    Refreshing a screen immediately before submission reduces stale information, but another update can still arrive between that refresh and the save. Ask the implementer to demonstrate that the relevant comparison and state change occur as one protected operation. A separate “check current status” task followed by an unrestricted update leaves the same gap.

    MDN’s If-Match documentation describes conditional requests as a way to prevent lost updates. That is one implementation mechanism, not proof that a particular approval screen uses it or that its check covers all business conditions.

    For a concrete product example, commercetools’ approval-flow guide requires the current flow version with an approval or rejection. A conflicting update produces a 409 response; the client should read the flow again and reconsider the action. Other applications can use different mechanisms and error responses.

    Define the decision attempt as its own work object

    Keep the business request, reviewed submission and decision attempt distinguishable. A request may have several submissions and attempts without having several effective final decisions.

    • Request and submission: identify the business record, review cycle, evidence basis and current approval policy.
    • Attempt: capture a stable attempt identifier, actor, intended action, reason, reviewed basis and expected authoritative revision.
    • Admission result: record whether the attempt took effect, was refused as stale or ineligible, or remains unconfirmed.
    • Effective result: retain the accepted decision, resulting revision and authorized next step separately from an attempted click.

    An opaque revision token need not be a human-friendly submission number. Ask which changes advance it and what record it protects. A token for the approval row alone may not cover an attachment replaced elsewhere. Define how changes to evidence, roles or policy invalidate a pending decision; do not assume a row version covers the whole review basis.

    Use an admission boundary before release

    At the decision boundary, require the system to establish the current record, eligible actor, permitted action, applicable review basis and expected revision under its demonstrated consistency model. Refuse an attempt that no longer meets those conditions. Record the result before initiating downstream work.

    Proposed decision-attempt state model
    StateAccountable ownerExit condition and evidence
    PreparedReviewerIdentified submission, intended decision and reviewed basis are available; no effective decision is implied.
    Submitted; result pendingReview operatorOriginal attempt is tracked until the authoritative service establishes accepted, refused or still unknown.
    AcceptedDecision service ownerSaved decision and resulting revision are confirmed. Release still depends on the full business policy.
    Refused; reassessment neededReviewer, coordinated by operatorCurrent facts, authority and decision history are reread. Reviewer abandons the old intent or makes a fresh eligible decision.
    Outcome unknownSystem operatorOriginal attempt is reconciled against authoritative state; affected release remains held until its disposition is known.
    Handoff acceptedExecution ownerReceiver accepts the confirmed authorized action with its evidence and identifiers; delivery outcome remains separately tracked.

    Do not convert a refused attempt into an effective rejection of the business request. The reviewer may still be authorized to act after reassessment, or the request may already have reached a terminal state. Likewise, a successful save of one reviewer’s decision is not necessarily final approval.

    Reassess stale decisions; reconcile unknown ones

    A confirmed conflict needs a new decision basis

    Show the reviewer what changed without exposing information outside their permissions. Identify the accepted decision, current submission and action now permitted. Preserve the refused attempt as history, but do not display it as an approval that took effect.

    Do not automatically fetch the new revision and replay the old approval. That would attach yesterday’s reasoning to today’s facts. If the reviewer still wishes to act, they must reassess and submit a fresh decision under the current policy. A new attempt cannot revive authority that has been revoked.

    A timeout does not tell you whether the save failed

    When acknowledgement is missing, the operator searches for the original attempt and reads the current authoritative decision. A retry is safe only through a demonstrated mechanism that preserves the original attempt’s identity and prevents duplicate effects. If that mechanism is unavailable, hold the affected release and obtain an operator disposition rather than clicking again.

    Revision checks and duplicate-attempt handling solve different problems. The first detects changed state. The second establishes whether a repeated submission represents an action already processed. Neither, by itself, proves that a notification, booking or other external action happened only once.

    Make the execution handoff explicit

    The sender is the review operator; the receiver is the named delivery owner. Ready-to-send means the effective decision is confirmed, the complete release policy passes and no relevant uncertainty remains. Include request/submission IDs, accepted decision reference, authorized scope, current release state, remaining conditions and the execution reference.

    Acceptance means the receiver can access the evidence, understands the permitted action and owns the next task. Until acceptance, the sender retains coordination responsibility. An email saying “approved” is not the release record.

    If the request changes before execution, validate the applicable release conditions again. For withdrawal and resubmission, use the separate approval recall workflow. If work already started, establish its actual disposition; changing an approval status does not reverse a real-world action.

    Test with two sessions, not a single happy path

    Use a non-production test record and authorized test roles. The process owner defines the expected business outcome; the implementer supplies evidence of the saved state and side effects. The matrix below is an acceptance specification, not a claim these tests have already run.

    Concurrent-decision acceptance matrix
    CaseRequired demonstrationFailure disposition
    Two sessions read the same revision; reject saves before approveThe later stale attempt cannot overwrite rejection. Current policy determines whether any fresh action remains eligible.Hold release; system owner investigates the admission boundary.
    Two sessions approve the same decisionEffective history follows the decision policy; repeated intent does not issue duplicate execution work.Reconcile decision and execution identities before release.
    Evidence changes while a review is openThe old review basis cannot silently authorize changed evidence, including evidence stored outside the decision row.Route back to review with a controlled current basis.
    Reviewer permission is removed after opening the screenSubmission checks current eligibility, not only login-time access.Operator obtains an eligible reviewer; do not restore access merely to finish the click.
    Save succeeds but its acknowledgement is lostOriginal-attempt reconciliation establishes the result without a duplicate decision or downstream task.Keep Outcome unknown with a named resolver and next review target.
    Decision is accepted but delivery handoff failsAccepted decision remains traceable; missing handoff is recovered without recreating the approval.Sender retains ownership until the receiver accepts.

    Unavailable reviewers need an authorized backup, not an overwrite privilege. Missing revision or evidence references block submission until the responsible owner supplies them. Overdue reconciliation escalates to the system and process owners; time elapsed never converts uncertainty into approval.

    Walk through a stale facilities review

    This is a constructed example. A facilities request for a meeting-room setup is open in two eligible reviewers’ sessions at revision 7. Reviewer A rejects it because the readiness evidence is incomplete. The accepted rejection creates revision 8.

    Reviewer B’s earlier approve attempt still references revision 7. The admission control refuses it; the screen explains that the decision changed and offers a current read, not an automatic resubmit. The operator confirms that no delivery task was released.

    The requester later supplies the missing evidence through the controlled review lifecycle. An eligible reviewer assesses that current submission. Only after its required decisions pass does the operator send one confirmed release packet to the facilities owner. If the first rejection’s acknowledgement had been lost, the operator would first reconcile that original attempt—not assume the request remained undecided.

    Measure integrity alongside flow

    • Demand and capacity: decision attempts, distinct requests and reviewer workload; separate repeated attempts from new demand.
    • Flow and queue health: pending-result age, reassessment wait and oldest unknown outcome by accountable owner.
    • Quality: stale-basis refusals, ineligible-action attempts and missing evidence detected before admission.
    • Reliability: overwritten decisions, duplicate execution tasks, unresolved acknowledgements and failed handoffs.
    • Outcome and improvement: evidence-backed releases and later delivery results. Investigate repeated conflicts before assuming faster clicks mean a better process.

    A rise in refused stale attempts can mean the control is detecting a real issue. Review long-open screens, unclear authority and competing entry points with the process owner. Do not remove the safeguard simply to improve a success-rate chart.

    Where Workhint fits

    The Workhint Workflow Engine describes approval routing to defined roles and required details, files or notes before work advances. These are relevant building blocks for the review path.

    Workhint’s Workflow Automation Software page describes forms, conditional routing, stages, permissions and event-based notifications. Those capabilities fit the shared process around reviewers, operators and delivery owners.

    Neither page establishes atomic revision admission, stable-attempt deduplication, coverage of external evidence changes or immutable audit retention. Validate those requirements in the configured system. If enforcement is unavailable, keep a controlled single-writer decision and manual release process, with direct competing update paths restricted by verified permissions. A spreadsheet warning or a coordinator’s presence is not a concurrency lock; defer unattended release until the control is demonstrated.

    Build This in Workhint

    Use this as a configuration and validation brief, not a promise of native concurrency protection:

    Design a shared approval process for [request type] that admits only eligible current decisions and ends in an accepted execution handoff. Define requester, reviewer, operator, process-owner, system-owner and delivery-owner roles with restricted decision and release permissions. Capture request/submission IDs, reviewed evidence, expected revision, stable attempt ID, intended action, reason, admission result and confirmed release reference.

    Model Prepared, Submitted; result pending, Accepted, Refused; reassessment needed, Outcome unknown and Handoff accepted. Keep underlying request states separate. Require current-authority and review-basis validation; identify how the implementer will enforce the atomic boundary. Assign evidence repair, reassessment, unavailable-owner, unknown-result and rejected-handoff tasks. Configure meaningful event notifications, overdue escalation and views for pending decisions, unknown outcomes and unaccepted handoffs. Retain actor, time, attempt and decision history; validate audit retention. Report the flow and integrity measures above. Demonstrate the two-session matrix and external-action behavior before unattended release.

    FAQ

    Does the first saved decision always win?

    No universal rule applies. The current policy determines eligible transitions and finality. A protected update prevents silent overwrite; it does not grant an actor authority or define which business decision is correct.

    Is disabling the approval button enough?

    No. Another session, delayed request or integration may still submit an update. Validate the authoritative decision boundary, not only the visible button behavior.

    Can we automatically retry a stale approval?

    Do not replay the old intent against a freshly fetched revision. Reassess current evidence, authority and state first. Treat unknown acknowledgements separately through original-attempt reconciliation.

    Comments

    Leave a Reply

    Your email address will not be published. Required fields are marked *


    The reCAPTCHA verification period has expired. Please reload the page.