1. LandOffer
  2. Blog
  3. All articles
  4. A Power Automate Flow Failed: Check What Already Ran Before You Resubmit

Productivity & Workflows

A Power Automate Flow Failed: Check What Already Ran Before You Resubmit

Use an actual offline ledger model to distinguish failed execution from completed external effects, unknown notification outcomes, and evidence required for an authorized bounded recovery.

9 min readLandOffer team

Before resubmitting a failed Power Automate cloud flow, inspect the individual actions and reconcile any external changes they may already have made. A failed run can coexist with a created record, an uploaded file, or a posted notification. Identify what is confirmed, what did not run, and what remains unknown before choosing a recovery scope.

This guide from LandOffer.ai uses current Microsoft documentation and an original offline ledger exercise. The example run, project record, connection repair, and notification outcomes are synthetic. We executed the local decision model, but did not access Power Automate, repair a real connection, query a real destination, send a notification, or resubmit a flow.

An analyst compares a failed-run record with an existing project card and a notification envelope marked Check first.

Find the exact failed run and the first relevant error

Start with the right environment, flow, and run. Record its identifier, start time, trigger data, and the business item it concerns. Similar timestamps alone may not distinguish repeated submissions or two requests processed close together.

Microsoft's cloud-flow troubleshooting guide describes selecting a failed run from its history and opening the failed step to inspect the error. It also documents adding trigger-output columns to the history view to identify particular runs. Use an available business identifier to connect the execution to its intended record.

Expand the relevant actions, including loop iterations and branches. Record their actual states, inputs, outputs, errors, and available retry information. A downstream action may be skipped because an earlier dependency failed. An error-handling branch may still have run. Keep each branch and iteration separate so its state remains traceable.

Microsoft's error-handling guidance describes run-after conditions for success, failure, timeout, and skipped states, as well as action retry policies. Read the applicable configuration in the flow you are investigating. A manually requested resubmission and an action's configured retry policy are different mechanisms.

If you cannot see the run or its protected details, ask the authorized owner for the required evidence. Missing access is a reason to keep the conclusion open, not to reconstruct an output from the action's name. Keep credentials and sensitive payloads out of an incident note.

Build an inventory of effects outside the run history

For each action capable of changing another system, identify the expected destination and a stable identifier. Use the returned item ID to open the destination record and inspect its relevant fields. For messages, distinguish a stored message record from proof that a recipient read it.

Check the business key as well as the generated ID. A repeated request may have created two records with different IDs. Search using the intended request identifier and inspect any matches before deciding which record to reuse. A title-only search can miss a duplicate or match an unrelated request.

Treat a timeout or missing response as an unresolved outcome until you reconcile it with destination evidence. Do not convert “the caller did not receive a response” into “the destination made no change.” This is a recovery precaution, not a claim that every timed-out action completed. If the destination cannot be inspected, keep that effect unknown.

Microsoft's trigger troubleshooting documentation explicitly discusses duplicated action effects and recommends designing for duplicate inputs. It gives duplicate checks and Dataverse key constraints as examples. That supports checking actual effects rather than assuming one run label establishes an exactly-once business outcome.

Completed asset: the failed-run effect inventory

In our fictional project-intake process, request OPS-207 should create one project record, post one notification referencing it, and mark the record notified. The synthetic run R-17 has an overall Failed state. Its destination ledger contains project ID 482 and no notification entries.

Here is the completed inventory. The destination evidence is from our local synthetic files, not a SharePoint, Teams, or other live connector readback.

Action Synthetic run evidence Synthetic destination evidence Recovery implication
Read request Succeeded; request key OPS-207 Intended request identified Preserve that business key in the recovery record
Create project record Succeeded; output ID482 Exactly one project with key OPS-207, ID482 Reuse the existing record; do not create another
Post notification Failed; 401 Unauthorized; one recorded attempt Complete fixture audit has zero notification entries Repair the modeled connection before a notification-only recovery
Mark notified Skipped after notification failure Project482 has notified=false Update only after the missing notification is accounted for

The “zero notifications” conclusion depends on a complete audit of the relevant synthetic destination. In real work, an empty search with the wrong channel, key, date window, or account is insufficient. Record the scope and source of a negative finding. If those are incomplete, classify the result as unknown instead.

Our model keeps the created project despite the Failed run state. A deliberately blind replay appends a second project with the same request key. The local duplicate check then blocks recovery. This demonstrates the stipulated ledger's behavior; it is not a Power Automate connector test or a claim that every resubmission creates duplicates.

Separate run resubmission from a supported partial recovery

The current Power Automate testing guide documents selecting a run instance and choosing Resubmit. It warns that resubmission can duplicate records or emails, and recommends checking input validity and changes to the flow or environment. The guidance does not promise that only failed business effects will be repeated.

Consequently, treat the documented run-level operation as a replay that requires an effect review. Do not assume that choosing Resubmit skips every earlier successful write. Establish what your actual flow and available recovery implementation will execute before authorizing it.

Some search results describe a different product. Microsoft's Azure Logic Apps run-history guide documents rerunning an entire workflow or starting from a specific action. Its action-start option has restrictions, including sequential workflows, a completed run, and forty or fewer actions. Loop and branch cases have further exclusions. That Azure capability is not evidence that the same control is available for your Power Automate cloud flow.

Likewise, “recover notification only” is a desired recovery boundary, not a button we observed. It may require an existing approved recovery flow, a properly designed duplicate guard, or an authorized manual completion procedure. If no supported route can limit effects appropriately, hold the recovery and involve the flow owner rather than inventing an action-level resubmission feature.

Repair the cause without assuming that repair settles replay safety

For the fictional 401 Unauthorized, the connection owner first needs to restore the required access. Microsoft's troubleshooting guide describes fixing the affected connection and verifying credentials. That repair addresses authentication; it does not remove the project already created by an earlier action.

Record which connection and destination the action uses, who owns it, and what evidence shows the repair worked. Use the approved connection identity; ask its owner to resolve missing access rather than substituting a personal account. An authorization problem may require an administrator or resource owner, not a broader credential.

Check the inputs again before replaying anything. The request may have changed or been cancelled. The destination record may have been edited by a teammate. Preserve the original run reference and document any deliberate use of current data so the recovery does not silently apply stale instructions.

Also establish the flow definition or recovery implementation being used. Microsoft warns that changes since the original run can affect resubmission. Its troubleshooting guidance discusses retrying after configuration correction, while the testing guide recommends considering version consistency. Identify the actual applicable behavior instead of assuming a rerun necessarily uses either the original or latest definition in every environment.

Completed asset: a bounded recovery and result record

Our local exercise models the connection changing from expired to valid, then evaluates a notification-only recovery using project482. This is a data-model transition, not observed reauthentication. The selected boundary excludes creating another project and requires a complete notification audit.

The completed recovery record is:

Original incident: R-17, request OPS-207. Project482 already exists. Notification action failed with a synthetic401 response; final status action was skipped.

Cause treatment: The modeled connection is now valid. In a real incident, the authorized connection owner must supply actual repair evidence before proceeding.

Allowed recovery: Reuse project482. If the notification audit is complete and finds no existing notification for OPS-207, perform only the approved missing-notification operation and subsequent completion update. If an existing notification is found, reuse its evidence. Do not create a project or change the request inputs.

Local outcome: One notification record, N-91, now references project482. The project retains ID482 and its notified flag becomes true. The local after-ledger contains exactly one project and one notification entry for OPS-207. A second evaluation selects the existing notification and adds nothing.

Live acceptance still required: Inspect the actual destination record and notification under the same business key, confirm intended content and recipient or channel, and link the authorized recovery execution. No live operation or recipient delivery was established by this exercise.

We also modeled a timed-out notification with an incomplete destination audit. The decision remains Hold even after the connection is valid. In a separate synthetic reconciliation, an existing notification N-90 is found; the model selects reuse rather than sending another. Fixing a connection does not answer that independent outcome question.

Candidate recovery Actual local model result Decision boundary
Recreate the project without reconciliation Two projects share OPS-207 Reject and investigate duplicates
Continue with expired connection Recovery held Repair evidence still missing
Continue with unknown notification outcome Recovery held Reconcile destination first
Recover confirmed missing notification using ID482 One project, one notification, notified=true Matches this synthetic case's stipulated result
Evaluate recovery again after N-91 exists No extra notification appended Reuses the found record in the local model

The executable exercise recorded fifteen assertions and retained before, rejected, unknown, repaired-candidate, and after ledgers. It validates this explicit local decision logic. It does not establish platform idempotency, atomicity under concurrent runs, connector behavior, or exactly-once delivery. A lookup followed by a write can still require stronger protections when multiple real executions race.

An effect ledger separates an existing project from a missing notification, with an unknown-outcome hold card kept distinct from the recovery path.

Verify recovery against the business result

After an authorized real recovery, inspect both the execution and the destination. Confirm the expected number of records for the business key, the retained project identity, the notification's reference and content, and any final status field. If a status says notified but the message cannot be found, investigate that inconsistency rather than using the status as its own proof.

Keep one unresolved effect visible instead of closing an incident because the newest run is green. An expected catch branch, a skipped loop iteration, or a destination that is temporarily unavailable can leave important work unfinished. Record who will inspect the missing destination evidence and what observation will close that gap.

If recovery has already produced a duplicate, preserve its identifiers and ask the responsible owner which record and downstream effects to retain. Deleting an extra project is not automatically sufficient if notifications, links, or other processes now refer to it. Follow the approved correction process with its own evidence.

For the next failed run, start an effect inventory before selecting Resubmit. The completion record should name existing effects, the authorized recovery performed, the destination evidence, and any unresolved outcome. For further workplace reading, visit LandOffer.ai.

Sources and Further Reading