Quick answer

Workflow versioning assigns executable rules to a running customer journey. Start new journeys on the new version and preserve existing ones while their original path remains safe and supported. Migrate selected waiting states only after proving that required data, pending events, accepted terms and recovery still work. Test the customer’s next payment, confirmation or consultation action; a successful deployment alone does not establish continuity.

Workflow versioning: preserve existing journeys unless migration is justified

Your release can be green while a customer’s issued payment link is broken. The deployment tests the application you have just shipped; that customer is testing a promise made by the application you replaced. Runtime workflow versioning must connect those two realities.

A workflow version is an executable set of rules. A running instance is one customer’s journey through those rules. A commitment is what that customer has already been offered or granted: an issued payment request, a reserved appointment, or paid access. Store commitments separately from the rule definition so changing the rules does not silently rewrite them.

For a custom application, my proposed policy is to start new journeys on v2 and let existing journeys finish on supported v1 paths. This fits when v1 remains correct, its dependencies remain compatible, and the team can support both cohorts. Change that decision when v1 contains a material defect, an identified requirement must affect existing cases, or a dependency cannot preserve the old behavior. Migration then needs a reason and a continuity test; a newer version number supplies neither.

Side-view of a standing engineer comparing a printed older form with its replacement on a wall-mounted touchscreen in a naturally lit device lab.

Why a version label does not protect the next action

A journey can still say v1 while calling newly deployed code. Trace its next action through the form, application handler, API, background worker, and any subprocess it invokes. Preserving a diagram or Git revision does not preserve that execution path.

In a database-backed application, record each journey’s rule version and route its transitions accordingly. Make that selection visible to operators. Retain data needed by occupied states or provide a tested adapter; removing an old field can break customers still waiting on the old path.

Prefer a shared handler when it can clearly support both inputs. Suppose v2 requires customerId while an older case has only email. Retain the valid legacy route while enforcing the new input for v2. If identity is essential to the next action, reconcile the incomplete case rather than inventing an identity.

Keep distinct form or handler references when behaviors conflict, until the old dependency clears. The api sdk workflow integration architecture guide helps locate those shared boundaries before deciding which code can serve both versions.

Camunda’s documented definition-versioning default keeps existing instances on their starting definition and starts new instances on the latest one. Supporting services can still change underneath them; subprocess bindings also need review. Temporal addresses a different boundary: Pinned Workflows stay on one Worker Deployment Version, while Auto-Upgrade Workflows require replay safety through patching, so updated code can still reproduce their recorded execution. Independent Activities can route outside a pinned deployment. Neither mechanism removes the dependency review, and neither is an automatic feature of an ordinary custom application.

A release engineer hands an older printed booking confirmation across a service counter as an operator compares it with a paper reservation folder.

Use a version-transition matrix to select the cohort

Choose a disposition for each occupied state and its pending event. Migration here means deliberately moving a running journey while preserving proven commitments and pending work. Cancelling it and recreating it is a separate recovery operation, with references and completed actions to reconcile.

Consider an illustrative coaching website releasing new qualification criteria and a fix for a missing future booking reminder. Assume 40 open v1 journeys: 12 awaiting qualification, 10 awaiting payment, 8 awaiting booking confirmation, and 10 awaiting consultation. The new criteria apply only to new customers; prices and appointment terms do not change. The lead initially wants to move everybody so there is only one version to support. The state review gives a more useful answer.

Affected stateDefault dispositionCompatibility requirementsMigration eligibilityCustomer-visible effectRollback constraint
Not startedStart on v2Coherent intake, validation and downstream handlersNo running instance to migrateNew rules disclosed at entryReturn future starts to v1 only if it accepts their data and terms
Awaiting qualificationPreserve v1Readable old answers, forms and decision inputsSufficient inputs plus an explicit decision on changed criteriaCommunicate revised questions or decisions before applying themVersion reversal cannot unsend a decision
Awaiting paymentPreserve issued request and termsSame reference, amount, currency, expiry, callback routing and duplicate handlingOutstanding payment and callback retain their meaningNo replacement request or changed price solely for migrationReconcile actual payment independently of code rollback
Awaiting booking confirmationPreserve unless mapping is provenSame reservation, slot, time zone, original expiry and confirmation eventTested waiting-state mapping with intact reservation and deadlineSame appointment and one logical confirmationReleased slots and sent confirmations need corrective action
Awaiting consultationPreserve booked service and accessAppointment, roles, entitlement, access checks, reminders and completion handlingRequired data exists or is deliberately supplied; commitments survive testingSame session, time and authorized participantsReconcile issued links, reminders and completed calls separately
Proposed version-transition policy for this release

The qualification and issued-payment cohorts stay on supported v1 paths: 12 + 10 = 22 preserved journeys. The booking and consultation cohorts are candidates because they could benefit from the reminder fix without changing the service: 8 + 10 = 18 candidates. One booking case lacks its reservation reference, leaving at most 17 for rehearsal. These are classification counts, not a migration success rate. Every remaining candidate still has to pass its state-specific tests.

The operator must reconcile the incomplete case against booking evidence, supply the verified reservation reference, and retest. Until then, retain a safe supported path or pause its affected transition for repair. A plausible appointment is not a reservation.

Within this example, Maya is one of the seven booking candidates with an intact reservation. Her payment of 80 units is already recognized. Her consultation is reserved for 24 November 2026 at 10:00 UTC, and the original confirmation deadline is 20 November at 12:00 UTC. The proposed reminder is due on 23 November at 10:00 UTC. Eligibility means retaining all of those facts, not just reaching a state called “awaiting confirmation.”

The proposed application must durably capture events, find journeys through their existing external references, and recognize repeated deliveries through stable event identifiers. Process migration and events one at a time per journey, checking its current state before acting. Record outbound work for controlled delivery. These are requirements to implement and rehearse, not benefits obtained by naming a version.

If Maya confirms first, stop the stale migration and reclassify her now-advanced journey. If migration happens first, the handler reads the migrated state and confirms there. Repeated delivery must create no additional booking or logical confirmation. Rehearse both orderings and delayed delivery; do not overwrite progress to fit the migration plan.

The customer message remains: “Your consultation remains reserved for 24 November 2026 at 10:00 UTC. Please confirm by 20 November at 12:00 UTC. Your payment of 80 units is already recorded; no further payment is needed.” In the example’s successful continuity test, Maya confirms through the existing booking route before that deadline. The application recognizes the reservation once, retains her appointment and paid entitlement, and schedules the reminder for 23 November at 10:00 UTC. The repeated event changes nothing. She receives the reminder and attends the original consultation without paying again.

For Camunda users, the waiting point must also support the proposed element mapping; changing an active element’s type is outside the documented mapping scope. Explicitly mapping relevant catch events can preserve event subscriptions and timer duration. Unmapped subscriptions may instead be removed or created. A booking deadline should survive by deliberate mapping and verification, not by assuming a similarly drawn timer means the same thing.

Overhead view of hands reconciling a payment receipt and appointment slip beside a phone and an open reservation folder with stapled papers.

Release, recover, and retire using continuity evidence

Expand only after customers’ next events reconcile with the commitments they already hold. A successful migration command proves less than a preserved booking confirmation reaching the right appointment. Keep an operator available through those events, with authority to stop expansion.

  1. Inventory open states, commitments, pending events and dependencies; complete the transition matrix for the actual release.
  2. Deploy compatible supporting code. Rehearse preservation and selected migration on resettable production-like data before moving production customers.
  3. Limit new v2 starts, then migrate a small eligible cohort if needed. Assign a release owner, a business owner for customer-visible changes, and an exception operator; one person may hold several roles.
  4. Reconcile the cohort’s next events, amounts, deadlines, appointments and access before expanding. Size the cohort to the operator coverage actually available.

Run one preserved v1 payment or booking fixture through the new deployment and inspect its commitment and next state. Cover every occupied state, missing data, duplicates, a pending timer and incompatible active work. The workflow automation testing checklist extends coverage; workflow automation monitoring helps specify operator evidence. Stop expansion for unexplained state loss, unmatched events, duplicate actions, changed commitments or unavailable recovery coverage.

In Camunda Operate 8.8, select the source version and intended running instances, choose the target version, map supported nodes, preview, then execute. Reconcile each instance afterward: the preview cannot prove payment-reference or worker compatibility.

For a harmful old path, the release owner isolates the transition while retaining event capture, deploys a correction and proves eligibility. If neither path is safe, the operator handles controlled repair. Camunda rejects invalid per-instance migration without moving that instance, but a batch is not atomic. Active jobs are not recreated, expressions reevaluated or input mappings reapplied.

Separate stopping new v2 starts, returning eligible journeys and reconciling external effects. For a paid customer whose mapping fails, retain the last confirmed state and payment evidence. Pause progression; assign the operator to restore one booking entitlement without another charge, recording the intended outcome and correction. Version reversal cannot undo received money, issued links or sent reminders.

Retire v1 after open journeys, jobs, timers, late events and recovery dependencies reconcile. Zero open journeys with two unresolved payment events is insufficient. Give the retained path an owner and review date; retain readable history and release mappings under the business’s retention policy. Temporal’s Drained worker status does not authorize deleting everything: query-related code can still run within retention. Consider shorter independent phases when long journeys prevent retirement.

Hardware security key beside a recovery sheet and sealed receipt folder on a worn metal surface in strong window light.

Carry continuity requirements into consultation delivery

When the journey ends in a website consultation, the release boundary includes entry to the call. A valid booking is insufficient if the paid customer can no longer reach it. Carry the same appointment, recognized payment, authorized participants and session-entry path into the integration decision.

Bring the completed matrix and Maya’s existing-booking scenario to a custom-development discussion. Ask the proposed delivery route to demonstrate that she reaches her original consultation with the correct access and that completion returns to the right journey. That gives an embed or a deeper integration a concrete requirement to meet, rather than treating a working call window as proof of workflow continuity.

A coach adjusts a fixed webcam while holding a headset in a daylight consultation room with acoustic panels, chairs and a wall display.

Choose the call integration around the existing booking

Website Video Calling compares simple embeds with API/SDK integration and describes custom website video experiences with scheduling, authentication, roles and paid-session access through Scrile Stream. Use the guide to compare those routes against Maya’s preserved appointment and access requirement, then bring the transition matrix to a custom-development discussion. The useful result is a concrete integration requirement; the guide does not establish workflow migration or rollback support.

Frequently asked questions

Is workflow versioning the same as source control?

No. Source control records changes; runtime versioning determines which executable rules a waiting customer actually receives after deployment.

Can a shared handler serve old and new journeys?

Yes, if it explicitly supports both input shapes and preserves their commitments. Conflicting behavior calls for separate references until the old dependency clears.

Does being a migration candidate mean a journey is eligible?

No. Candidate selection identifies a possible benefit. Required data, supported state mapping, pending events and recovery still need to pass rehearsal.

Should confirmations be disabled during migration?

Keep event capture available. Process each journey’s events and migration in a controlled order, then recheck its current state before applying either operation.

What should be tested before widening the release?

Run preserved and selected migrated journeys through their next customer action, including duplicate delivery, missing data and pending timers.

Does rolling back code undo a payment?

No. Preserve evidence of the external payment and reconcile the resulting entitlement separately. A version reversal does not refund or erase it.

Can old workers be removed when no journeys remain open?

Only after pending work, unresolved late events and recovery dependencies are reconciled. Retain historical records and any code still needed during retention.

Does a video integration provide workflow migration?

Do not assume so. Evaluate the call integration against booked-customer access requirements; design and verify workflow migration separately.