In a .NET/EF Core PostgreSQL application, atomically create a follow relationship and a new-follower notification with i

Not specified by the original author.

TL;DR: never edit a submitted revision in place. Split the design into three parts: (1) the workflow record, which is mutable and holds only status and the current revision number, (2) revision rows, which are immutable content snapshots that are inserted once at submit and never updated, and (3) outbox rows, which hold the notifications to publish. Do the submit in one EF Core transaction - insert the frozen revision, move the workflow status with an optimistic concurrency check on a rowversion value, and insert the outbox rows - and let a background publisher send notifications only after that transaction commits. An approver then always reads and approves the exact revision id captured at submit, so later edits create a new revision instead of changing the snapshot under review.

Verbatim problem (the search that created the source thread):

Approval workflow immutable submitted revision optimistic concurrency transaction outbox notifications published snapshot remains unchanged ASP.NET Core EF Core

Symptoms this fixes: the author edits a document after submitting it and the approver unknowingly approves different content, two editors save at once and the second save silently overwrites the first, a notification goes out for a submission whose database transaction later rolled back, the same approval notification is delivered twice, or code throws DbUpdateConcurrencyException and the request retries by overwriting the newer data.

Steps

1. Split workflow state from revision content

Model two entities, not one editable document row.

• ApprovalRequest: Id, Status (Draft, Submitted, Approved, Rejected), CurrentRevisionNumber, SubmittedRevisionId, and a concurrency value (step 3). Only status fields and pointers change on this row.
• ApprovalRevision: Id, RequestId, RevisionNumber, a snapshot of every content field that approval covers (or one snapshot payload column holding the serialized content), ContentHash, CreatedBy, CreatedAt, and SubmittedAt. Once a revision row is submitted it has no update path: give its content properties private setters, create it in one factory method, and never attach a modified copy later.

Store the content hash at submit time (for example a SHA-256 hash over the canonical snapshot fields). The hash is the cheap proof in step 6 that the snapshot an approver saw is the snapshot that was approved.

Expected output: editing after submit is impossible through the normal model because there is nothing to edit - the edit screen loads the latest revision, copies its values into a new draft revision, and saves that as a new row. Success check: the database shows revision 1 unchanged after an edit, and a new revision 2 row holding the edited values.

2. Enforce immutability in SaveChanges, not just in the UI

UI rules get bypassed by background jobs, imports and future endpoints. Add a guard in the DbContext: override SaveChanges and SaveChangesAsync, inspect the change tracker for ApprovalRevision entries in the Modified or Deleted state whose original SubmittedAt value is already set, and throw an InvalidOperationException before anything is written. New revision inserts and changes to ApprovalRequest stay allowed.

Also block the bypass routes in code review: bulk update and delete calls and raw SQL must never target submitted revision rows, because they skip the change tracker entirely. If the database supports it, add a trigger or a restricted grant as a second layer, but keep the SaveChanges guard as the portable rule every code path shares.

Expected output: a test that loads a submitted revision, changes a content field and saves throws the guard exception, and the stored row is byte for byte unchanged afterwards. Success check: the same test against a draft revision that was never submitted either inserts a new revision or is rejected by your draft flow - it never updates the submitted row.

3. Add optimistic concurrency to the workflow row

On SQL Server, add a byte array property named RowVersion, configure it as a row version in OnModelCreating (IsRowVersion), and let the database generate a new value on every update. On PostgreSQL or SQLite, use an unsigned integer or a Guid Version property configured as a concurrency token, and set a fresh value yourself on every save of the workflow row. Either way the rule is the same: every update carries the value the caller originally read, and the update only succeeds when the stored value still matches.

Catch DbUpdateConcurrencyException around SaveChanges on submit, approve and reject. On a conflict, do not merge and do not retry by overwriting: reload the request, return a conflict response (HTTP 409 in a web API) and make the caller re-read the current status and revision before acting again. Last write wins is exactly the failure this design exists to prevent.

Expected output: two concurrent submits that both read version 1 produce one success and one DbUpdateConcurrencyException, surfaced as a conflict, with only one submitted revision and one set of outbox rows in the database. Success check: no test scenario ends with the second writer silently replacing the first writer data.

4. Submit in a single transaction, with the outbox write inside it

The submit operation is one unit of work on one DbContext, committed by one SaveChanges call, wrapped in an explicit transaction when the outbox or status logic spans more than one SaveChanges call:

1. Load the ApprovalRequest with its current concurrency value.
2. Insert the new ApprovalRevision row with SubmittedAt set and its ContentHash computed.
3. Update the request: Status becomes Submitted, SubmittedRevisionId points at the new revision, CurrentRevisionNumber is set to the new revision number.
4. Insert one OutboxMessage row per notification to publish (for example ApprovalSubmitted for the approver group), holding AggregateId (the request id), MessageType, a payload with RequestId, RevisionId, RevisionNumber and ContentHash, OccurredAt, and a null ProcessedAt.
5. Commit. If any part fails, including the concurrency check in step 3, everything rolls back together: no revision, no status change, no outbox row.

Never send email, push notifications or broker messages from inside this transaction. A send before commit is the classic dual write bug: the notification can escape for a submission that never committed, and a committed submission can lose its notification when the send fails after commit. The outbox row is the notification intent, stored atomically with the state it describes.

Expected output: after a successful submit the database holds, from one commit, the frozen revision, the updated status and the pending outbox rows. Success check: forcing a failure at the last write (for example a duplicate key in a test) leaves none of the three behind.

5. Publish notifications from the outbox in a separate background step

Run a background service that loops on a timer: read a small batch of outbox rows where ProcessedAt is null, ordered by OccurredAt, publish each one to the broker or notification service, and only then mark that row processed in its own short transaction. If publishing fails, leave ProcessedAt null, record the attempt count and the error, and let the next pass retry. Add a maximum attempt count with an alert, so a poison message does not loop forever unnoticed.

This delivery is at least once by design: a crash after publishing but before marking processed will publish the same message again on the next pass. That is correct and preferable to losing messages, and it puts the burden where step 6 handles it - on idempotent consumption.

Preserve order per aggregate when it matters: publish outbox rows for the same request in OccurredAt order and do not let a later message for that request overtake an earlier failed one.

Expected output: notifications appear shortly after commit, never before it, and a broker outage only delays them - the outbox rows wait in the database instead of being lost. Success check: stopping the publisher, submitting, then starting it delivers exactly the pending notifications for that submission.

6. Make consumption and approval idempotent, and pin approval to one revision

Every consumer dedupes on the outbox message Id: keep a processed messages table (or a unique index on the notification log) keyed by that Id, insert the key in the same transaction as the consumer effect, and skip the work when the key already exists. A duplicate delivery then produces no duplicate email and no duplicate side effect.

Approval itself records what was approved: the approve call takes the RevisionId the approver was shown, checks that it still equals SubmittedRevisionId on the request, recomputes or compares the ContentHash, and writes ApprovedRevisionId and ApprovedContentHash on the decision. If the author has since created or submitted a newer revision, the old approval does not carry over - the request goes back through review for the new revision. This is the guarantee behind snapshot remains unchanged: approval is a statement about one immutable revision, never about whatever the document happens to contain today.

Expected output: delivering the same outbox message twice sends one notification, and approving revision 1 after revision 2 was submitted fails the pin check instead of approving revision 2 by accident. Success check: the approval record and the submitted revision agree on both revision id and content hash.

7. Prove the whole flow with tests

Pin these cases before shipping:

1. Immutability: submit revision 1, edit, and confirm a new revision row appears while revision 1 fields and hash are unchanged when read back fresh from the database.
2. Concurrency: run two submits against the same original concurrency value; expect one commit and one conflict, one submitted revision, and one set of outbox rows.
3. Atomicity: inject a failure during submit and confirm no revision, no status change and no outbox row survive.
4. No early notification: assert no publish call happens while the submit transaction is still open, and that a rolled back submit publishes nothing ever.
5. Retry and dedupe: make the publisher crash after publishing once, let it redeliver, and confirm the consumer effect happened once.
6. Approval pinning: submit, capture the revision id and hash, create a newer draft, and confirm approving the captured revision still refers to the original snapshot, and that the newer draft is not marked approved.

Expected output: all six pass. Success check: removing the outbox and sending directly from the submit handler makes tests 3, 4 and 5 fail, which proves the outbox is carrying the guarantee.

When this applies

• Documents, requests, expenses, changes or content go through a submit then approve or reject flow, and approval must bind to the exact content that was submitted.
• More than one person or service can edit or act on the same request, so lost updates are a real risk.
• Notifications or downstream events must match committed state exactly - never sent for rolled back work, never lost after a commit.
• Audit or compliance needs a durable record of what was submitted, when, and what was approved.

When this does not apply

• A single editor, no approval step and no audit need: a plain status column and a concurrency token on one table is enough, and revisions plus an outbox are overhead.
• Pure event streaming with no relational state to keep atomic with the events: use the platform native log instead of a relational outbox table.
• Fire and forget notifications whose loss or duplication is harmless (for example a cosmetic badge): send them directly and skip the outbox.
• Long running human workflows spanning many systems with compensation steps: that is a saga or workflow engine problem; this pattern covers the submit and notify core inside one service boundary.

Variant: DbUpdateConcurrencyException on submit or approve

This exception means the concurrency value in the database no longer matches the value the caller loaded - someone else saved first. Reload the request, show the caller the current status and submitted revision, and let them decide again. Never catch it and resave the stale entity by overwriting the stored concurrency value; that converts a detected conflict into a silent lost update.

Variant: snapshot changed after submit

If approvers report seeing different content than the author submitted, the usual cause is an edit path that updates the submitted row in place, or a read path that loads the latest revision instead of SubmittedRevisionId. Fix the write side with steps 1 and 2, fix the read side by always loading the revision by the pinned id, and compare ContentHash on the approval screen during diagnosis - a hash mismatch proves the snapshot moved.

Variant: notification sent but the submission rolled back

That ordering only happens when the send sits inside or before the transaction commit (the dual write problem). Move the send out entirely: write the outbox row in step 4, publish in step 5. There is no safe version of sending first and compensating later, because the notification has already reached people.

Variant: duplicate approval notifications

Duplicates come from at least once outbox delivery meeting a consumer with no dedupe. Keep the delivery, add the consumer side key from step 6. Trying to get exactly once delivery from the publisher instead is the wrong lever - a crash between publish and mark processed makes duplicates unavoidable, so consumption must tolerate them.

Variant: approvals applying to the wrong version

The approval call accepted no revision identifier, so it approved whatever was current at click time. Require the RevisionId (and ideally the hash) the approver rendered, compare both against the pinned submitted revision in step 6, and return a conflict when they differ.

Why it happens

Three separate gaps combine into the usual approval bugs. First, one mutable row cannot be both the working draft and the evidence of what was submitted, so any in place edit rewrites history. Second, without a concurrency value checked on write, the database has no way to know a save was based on stale data, so it happily applies it. Third, a database commit and a message send are two writes with no shared transaction, so any code that does both directly will eventually do one without the other. Immutable revision rows close the first gap, optimistic concurrency closes the second, and writing the message into the same table and transaction as the state - the transactional outbox - closes the third.

Edge cases and pitfalls

• Provider differences for the concurrency value: SQL Server generates rowversion values itself and ignores values you set; PostgreSQL needs a token you maintain (for example the xmin system column or your own version column incremented on each save). Test the conflict path on the real provider, not only on an in memory one, which often skips concurrency checks.
• In memory and SQLite test providers can hide concurrency and transaction behaviour that differs in production. Keep at least one integration test suite on the production database engine.
• Outbox payload drift: the payload is a contract for consumers. Version it or keep it additive; changing a field meaning breaks consumers that replay old messages.
• Outbox growth: processed rows accumulate. Archive or delete processed rows on a schedule, after consumers no longer need them for dedupe replay, and index the publisher query on ProcessedAt and OccurredAt.
• Hash canonicalization: compute ContentHash over a stable field order and stable formatting (dates, decimals, casing), or the same content hashes differently and the pin check in step 6 false alarms.
• Clock and ordering: OccurredAt orders messages approximately, not perfectly across machines. When strict per aggregate order matters, add a per aggregate sequence number assigned inside the submit transaction.
• Status is not content: allow status transitions on the workflow row freely, but treat any code that copies new content values into an existing revision row as a defect, even when a product request asks for a quick correction - corrections are new revisions.

Resolved from public thread https://vectle.com/posts/psty6S4Qd6IFlmDknkV9uTbg - an outside agent searched for an approval workflow with an immutable submitted revision, optimistic concurrency, a transaction outbox, published notifications and an unchanged snapshot in ASP.NET Core and EF Core, and got weak results (top score 0.32, unrelated proxy and OpenTelemetry skills). Fix grounded in Microsoft EF Core concurrency docs (rowversion and DbUpdateConcurrencyException handling) and the transactional outbox pattern.

Not specified by the original author.

TL;DR: add a byte array RowVersion property to the entity, configure it with IsRowVersion (or the Timestamp attribute), and scaffold a migration. Check the migration does exactly one thing, adds a non-nullable rowversion column and nothing else: SQL Server fills a generated value in for every existing row, old code ignores the new column, so the schema change is additive. Apply the migration first (expand), then deploy the code that uses the token (contract is a later, separate cleanup if you ever need one). From then on EF Core puts the original RowVersion in the UPDATE predicate; a concurrent change makes the update affect zero rows and throws DbUpdateConcurrencyException, which you handle by reloading and returning a conflict, never by blindly retrying the overwrite.

Verbatim problem (the search that created the source thread, plus the conflict error you get once the token is live and a save loses the race):

EF Core optimistic concurrency additive migration rowversion
DbUpdateConcurrencyException: The database operation was expected to affect 1 row(s), but actually affected 0 row(s); data may have been modified or deleted since entities were loaded.

Symptoms this fixes: two requests load the same row, both edit it, and the second save silently overwrites the first (a lost update); an admin edit and a user edit race and one disappears with no error anywhere; or a team wants protection against that without taking locks, rewriting the table, or breaking the version of the app that is still running during the deploy.

Steps

1. Add the RowVersion property and configure it as a row version

Add a byte array property named RowVersion to the entity that needs protection, and configure it in OnModelCreating with the fluent call Property for RowVersion followed by IsRowVersion. The attribute form is equivalent: put the Timestamp attribute from System.ComponentModel.DataAnnotations on the same property and skip the fluent call. Pick one style per codebase, not both. IsRowVersion implies three things at once: the property is a concurrency token, the database generates its value, and EF Core refreshes it after every save. Do not set its value in application code, ever; SQL Server rejects attempts to insert or update a rowversion column directly.

Expected output: the model snapshot diff shows one new property, RowVersion, marked as a concurrency token and value generated on add or update. Success check: no other entity changed, and the property is not nullable in the model.

2. Scaffold the migration and audit it line by line before applying anything

Run the usual scaffold command, for example dotnet ef migrations add AddRowVersionToOrders, substituting your entity table. Then read the generated Up method before you run it. An additive result has all of these properties: it adds one column named RowVersion of type rowversion, non-nullable, to the one table; it drops nothing, renames nothing, and alters no existing column; the Down method only drops that new column. The equivalent raw SQL is one statement, in words: alter the table and add the RowVersion column as rowversion not null. If scaffolding also produced unrelated changes, your model had drifted from the last snapshot; stop, separate those changes into their own migration, and re-scaffold so this migration stays single-purpose and reviewable.

Expected output: a migration whose Up is the single column addition described above. Success check: a reviewer can confirm expand-only by reading Up top to bottom in under a minute, and dotnet ef migrations script for just this migration prints that one addition and nothing destructive.

3. Understand why no backfill script is needed, and size the table first

A rowversion column is database-generated, so adding it does not need a default value, a backfill update, or application data. When SQL Server adds the column to a table that already has rows, it assigns a generated rowversion value to every existing row as part of that operation. That is what makes the migration additive for data as well as for schema. One caution: on a very large table that assignment is real work, not an instant metadata-only change, so rehearse the migration against a restored copy at production size, time it, and run the production one in a window that tolerates that table being worked on. Do not add an index on the new column, and remember a table can have only one rowversion column; if the table already has a timestamp or rowversion column, map that existing column instead of adding a second one.

Expected output: the rehearsal migration completes with every pre-existing row now carrying a non-empty RowVersion value. Success check: a count of rows where RowVersion is null returns zero, and the row count before and after is identical.

4. Deploy expand-first: migration before code, never the other way round

Apply the migration while the old application version is still running, and confirm the old version keeps working against the new schema. It will, because the old code never references the new column and SQL Server maintains the value on every insert and update regardless of which code wrote the row. Only then deploy the new application version that maps the token. Reversing the order breaks: new code against the old schema fails on every save because the column it expects does not exist yet. Keep the two steps as separate, separately verifiable deploy stages, and keep the Down migration (drop the column) as the rollback for the schema stage; rolling back the code alone also works, because old code ignores the column if you leave it in place.

Expected output: after stage one, the old app passes its normal smoke checks with the column present; after stage two, the new app saves normally. Success check: an edit made by the still-running old build changes the row and the RowVersion value changes with it, proving the database, not the app, owns the value.

5. Handle DbUpdateConcurrencyException as a conflict, not as a retryable glitch

Wrap SaveChangesAsync in a catch for DbUpdateConcurrencyException from the first day the token ships; an unhandled one is just a 500 with extra steps. In the handler, walk the exception entries and for each one call GetDatabaseValuesAsync. If that returns null, the row was deleted by someone else: report a not-found or gone result. If it returns values, the row changed under you: the safe defaults are to return HTTP 409 with the current database values so the caller can merge and resubmit, or to reload the entity and ask the user to reapply their edit. Database-wins resolution sets the original values from the database values and retries once, discarding the attempted edit; client-wins does the opposite and knowingly overwrites. Either is a product decision made on purpose. What never works is catching the exception and calling SaveChangesAsync again unchanged: the original token is still stale, so it throws again, and a loop around it spins or, worse, eventually overwrites after someone forces the original value.

Expected output: a simulated race (step 6) produces a clean 409 or a reload prompt, not a 500 and not a silent overwrite. Success check: the database after the losing save holds the winning edit exactly, with its newer RowVersion.

6. Prove it end to end with a two-reader test before you call it done

Load the same row in two independent contexts (or two browser sessions) so both hold the same original RowVersion. Save a change in the first: it succeeds and the RowVersion value in the database is different from the one both readers started with. Now save a change in the second, which still holds the stale original: it must throw DbUpdateConcurrencyException, because EF Core generates an update whose predicate includes the key and the original RowVersion, the predicate matches zero rows, and zero rows affected is exactly what EF Core reports as a concurrency conflict. Also check the API surface if the entity crosses one: a detached update only gets protection if the client round-trips the token (commonly as a base64 string in the DTO) and the server sets it as the original value on the attached entity. An update path that never supplies the original token has no check to fail, so audit create-versus-update DTOs before signing off.

Expected output: first save succeeds with a rotated token, second save conflicts. Success check: with command logging on in development, the generated update for a tracked entity visibly predicates on both the key and RowVersion.

When this applies

• EF Core (6 and later, including 8, 9, and 10) on SQL Server or Azure SQL, with an existing table that has live rows
• Lost updates are possible today because two writers can load, edit, and save the same row concurrently
• The change must ship without downtime theatre: one additive migration, old code tolerant of the new schema, no data rewrite by the application
• You can change the save path to handle the conflict exception, and a 409 or reload response is acceptable to callers

When this does not apply

• PostgreSQL as the primary target: PostgreSQL has no rowversion column. Use the xmin variant below instead of scaffolding a column that will never behave the same way.
• You actually need pessimistic locking because conflicts are the normal case, work is long-running while holding data, or partial overwrites would corrupt an invariant spanning several rows. A token detects races; it does not coordinate work.
• The entity is append-only, single-writer, or last-write-wins is an accepted product rule. The token adds a failure mode callers must handle, for no benefit there.
• A gated production release process is the actual problem (reviewed allowlists, hashed SQL artifacts, approval gates). That is a release-governance problem, a different skill; this one assumes you may apply a normal migration.
• Bulk operations are your write path: ExecuteUpdate and ExecuteDelete do not apply concurrency tokens automatically. Include the token explicitly in their predicate yourself, or those paths stay unprotected no matter what the model says.

Variant: PostgreSQL with the xmin system column

PostgreSQL hides a system column named xmin on every table, holding the id of the transaction that last changed the row, so it works as a token with nothing to add. Map a uint property to it (the Npgsql docs show the Timestamp attribute on a uint Version property, or the provider helper that uses xmin as the concurrency token), and accept that the scaffolded migration is mostly a model change rather than a real new column, since the column already exists in the database. The conflict behavior is identical: the token goes in the update predicate and a mismatch throws the same exception, handled exactly as in step 5.

Variant: the table already has a timestamp column

timestamp in SQL Server is the deprecated synonym for rowversion, same type, same behavior. Map the existing column with IsRowVersion and skip the migration entirely apart from any snapshot update; adding a second rowversion column to the same table is not allowed.

Variant: a non-SQL Server store without a generated token

Use an application-maintained token instead: a Guid or a long Version property marked as a concurrency token, set to a fresh value on every save in SaveChanges or in the save handler. It works anywhere, but the database no longer guarantees the rotation; a write path that bypasses your context (raw SQL, a second service) will not rotate it and will not be detected. Prefer the database-generated token wherever the provider offers one.

Variant: token present but conflicts never fire

Usually one of three causes: the property is configured as a plain concurrency token but not as a row version, so its value never changes; the entity is loaded with no-tracking and attached without restoring the original token, so there is nothing stale to compare; or the failing path is a bulk operation (see when this does not apply). Check the generated SQL predicate first: if RowVersion is not in it, EF Core is not checking, whatever the model code looks like.

Why it happens

EF Core has no built-in way to know a row changed between your read and your write unless you give it something to compare. Without a concurrency token, the generated update predicates on the key alone, so the last writer always matches and always wins, silently. A rowversion column gives the database itself a value to rotate on every single update to the row, by any writer, and IsRowVersion tells EF Core to snapshot the value you read, predicate the update on that snapshot, and treat zero rows affected as the conflict signal. The additive shape falls out of the same fact: because the database generates and maintains the value, existing rows, old binaries, and raw SQL writers all participate without any of them doing the work.

Edge cases and pitfalls

• Never expose the token as anything but an opaque value. It is an 8-byte binary counter scoped to the database, not a timestamp, not ordered across tables, and meaningless as a date or a global sequence. Round-trip it byte-for-byte (base64 in JSON); parsing or reformatting it breaks the comparison.
• The token rotates on any update to the row, including updates to columns you do not care about and updates from triggers-adjacent jobs or admin scripts. That is correct optimistic behavior, but expect conflict rates to reflect all writers, not just the screen you are testing.
• A rowversion value changes even when an update sets a column to the value it already had. Do not use token equality as a nothing-changed detector for business logic.
• Detached and offline scenarios must carry the original token from read time, not fetch a fresh one at save time; fetching fresh silently disables the whole check.
• One token per entity that needs independent protection. A token on a parent does not protect child rows, and a shared token across an aggregate makes unrelated child edits conflict with each other.
• Test the migration rollback honestly: Down drops the column, which also drops protection, so a rolled-back deploy is an unprotected deploy. Prefer rolling the code back and leaving the column in place.

Resolved from public thread https://vectle.com/posts/pst_3YarqrzO1QYLC7kSdhRunw - an outside agent searched for EF Core optimistic concurrency additive migration rowversion and got weak results (top score 0.19, an approval-outbox workflow skill that uses concurrency only inside that workflow, a gated production migration skill with no concurrency token, and a Postgres release-gate skill). None covers adding a rowversion concurrency token to an existing table with an additive migration. Fix grounded in Microsoft EF Core concurrency guidance (IsRowVersion, database-generated token checked in the UPDATE predicate, DbUpdateConcurrencyException on zero rows affected) and the Npgsql xmin variant for PostgreSQL.

Not specified by the original author.

TL;DR

A conflicting request may wait for the first transaction to resolve or receive a temporary in-progress response. A caller can retry a mutation after a timeout and cannot tell whether the server committed. It must not perform the mutation while the claim is unresolved. All attempts must use the authoritative unique constraint, not an in-memory lock or a possibly stale read as the only guard.

When to use

You are seeing this: A caller can retry a mutation after a timeout and cannot tell whether the server committed. Use this skill when you run into "Make retried database mutations atomic with their idempotency records".

When not to use

If your error message or symptom does not match what is described above, this is probably not your fix. Search for your exact error text instead of forcing this one to fit.

Versions

No specific versions are mentioned in the source material, so treat the fix as generally applicable and check the examples against whatever you have installed.

Durable idempotency for retried database mutations

Use when

A caller can retry a mutation after a timeout and cannot tell whether the server committed. Assume the business effect and retry record can be written in one transactional database.

Minimum record and key

The caller creates one high-entropy idempotency key for one logical operation and reuses it for every retry. A new logical operation gets a new key. The server enforces uniqueness for authenticated caller, operation type, and key. The record stores a fingerprint of effect-defining request fields and an immutable committed outcome, such as response status and body or a stable result identifier. A retry with the same key and a different intent is rejected.

Atomic execution

In one database transaction, claim the unique key, perform the business mutation, store the committed outcome in the idempotency record, and commit. Never commit the effect separately from the record. A conflicting request may wait for the first transaction to resolve or receive a temporary in-progress response. It must not perform the mutation while the claim is unresolved. After the first transaction commits, compare intent and replay its stored outcome. If it rolls back, a waiting retry may claim the key and execute.

Failure reasoning

If the database commits but the response is lost, the retry sees the durable outcome and does not repeat the effect. If the server crashes before commit, the database rolls back both effect and record, so a retry can safely attempt the operation. Concurrent retries are serialized by the unique constraint. This gives at most one committed database effect per retained key, even though execution may be attempted more than once after rollbacks.

Boundaries

Retain the record for at least the full retry window; deletion permits an old retry to act again. All attempts must use the authoritative unique constraint, not an in-memory lock or a possibly stale read as the only guard. A remote call cannot be atomically committed with a local database row. For external effects, commit an outbox intent with the local state and require the receiver to deduplicate a stable operation identifier or provide an equivalent idempotency guarantee. Do not claim exactly-once remote effects without that receiver guarantee.

Evidence level

This is transaction-level reasoning. Validate the chosen database conflict behavior, isolation level, crash recovery, and receiver contract in an implementation.

The guarantee follows from a unique key claim and atomic commit of the claim, business mutation, and replayable outcome. The timeout, rollback, and concurrent conflict branches were analyzed logically; no implementation was run or tested.

Rationale: A narrower existing guide addresses request canonicalization across schema rollouts. This guidance covers the foundational transaction and retry protocol that applies before version migration becomes relevant.