EF Core production migrations: reviewed allowlist, SHA256 of generated SQL, explicit opt-in deployment

TL;DR: never generate or apply EF Core migrations at production deploy time. Generate the idempotent SQL once in CI from a frozen, pinned migration set, treat that exact file (or the migration bundle built from it) as the immutable release artifact, and gate production on four checks that must all pass in one dedicated job: the migration IDs are on a reviewed allowlist, the SHA256 of the artifact about to run matches the SHA256 recorded at approval time, a human has explicitly opted in and approved, and a fresh verified backup exists. Rehearse the exact artifact on staging first, then apply it from a single runner under the migrator advisory lock with per-migration transactions, and verify the history table afterwards.

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

EF Core production migrations reviewed allowlist SHA256 generated SQL explicit deployment opt in backup transaction advisory lock immutable release artifact approval staging

Symptoms this fixes: production running SQL nobody reviewed because it was generated on the deploy machine, a regenerated script differing from the reviewed one after a tool update, two app instances racing migrations at startup, a deploy applying migrations as an unnoticed side effect of an app release, and no clean answer to what exactly ran, who approved it, or how to recover when step three of five fails.

Steps

1. Freeze the migration set and pin the tooling

Lock the release to an explicit list of migration IDs (the output of dotnet ef migrations list at the release commit) and pin the dotnet-ef tool version and the EF Core package versions in the repo (a tool manifest, not a globally installed latest). After this point the migration files do not change: any fix means a new migration appended to the list and a new artifact, never an edit to a frozen migration. Regenerating SQL with a different EF Core minor version can change the generated statements, which is exactly what the hash gate in step 4 exists to catch, so the generator version is part of the release record.

Expected output: a release record listing the commit, the from migration (the last migration production is known to be on), the to migration, and every migration ID in between. Success check: the list in the record matches dotnet ef migrations list at that commit exactly, with nothing pending left out.

2. Generate the idempotent SQL once, in CI, as the artifact

In CI, at the release commit, generate the script for exactly that range with the idempotent flag, for example: dotnet ef migrations script FROMMIGRATION TOMIGRATION --idempotent --output artifacts/migrations.sql (substitute the real migration names for the bracket-free placeholders FROMMIGRATION and TOMIGRATION). Idempotent means each migration is guarded by a history-table check, so re-running the script after an interrupted deploy skips what already applied instead of failing on existing objects. Normalize line endings to LF at generation time and record the byte size alongside the file, so the hash in step 4 is stable across checkouts. This file, not the migration source files, is what gets reviewed, hashed, approved and executed. The alternative artifact form is a migration bundle (dotnet ef migrations bundle), a self-contained executable built from the same frozen set; if you use a bundle, the review in step 3 reviews the SQL the bundle will run, generated the same way, and the bundle binary is the hashed artifact.

Expected output: one migrations.sql (and optionally one bundle) in the build artifacts, produced by the pinned tooling. Success check: generating again from a clean checkout of the same commit with the pinned tooling produces a byte-identical file.

3. Review the generated SQL against an allowlist

A human reviewer reads the generated SQL, not just the C sharp migration code, because the SQL is what production will execute and raw SQL operations inside migrations are invisible in a model diff. The review checks two things. First, the migration ID allowlist: the release record lists the exact migration IDs permitted in this release, and a CI check extracts the IDs the script will record in the history table and fails the build if the sets differ in either direction. Second, a statement review: destructive or irreversible operations (dropped tables or columns, type narrowing, large table rewrites, raw SQL blocks) are flagged and need explicit sign-off named in the approval, and anything the reviewer cannot explain blocks the release. Data-moving migrations get extra scrutiny for lock duration and batching.

Expected output: an approved allowlist file in the release record and a review note naming any destructive statements and their sign-off. Success check: deliberately adding an unlisted migration to a test build makes the allowlist check fail before anything reaches staging.

4. Hash the artifact and bind the approval to the hash

Compute the SHA256 of the exact artifact bytes (sha256sum on Linux CI, Get-FileHash on Windows) and write the digest into the release record. The human approval in step 6 approves that digest, not a version label: the approval reads approved artifact SHA256, migration range, target environment. The deploy job recomputes the digest of the file it downloaded and refuses to run on any mismatch. This closes the gap where an artifact is rebuilt, re-uploaded or swapped between review and deploy while keeping the same file name. Store the artifact in versioned, write-once storage (an artifact feed or bucket path that includes the digest) so the approved bytes cannot be overwritten in place.

Expected output: one digest in the release record, one digest computed at approval, one digest computed in the deploy job, all three equal. Success check: flipping one byte in a staging copy of the script makes the deploy job stop at the hash check with no database connection opened.

5. Make deployment explicit opt-in, never a side effect

Production migration runs only in a dedicated job that is disabled by default and is not part of the normal app deploy pipeline. Turning it on takes two deliberate acts: a pipeline variable or job toggle set for this run only (call it the production migration opt-in), and a manual approval from the named approver recorded against the hash from step 4. The application itself must not migrate on startup in production: no Database.Migrate call in the production startup path, because app startup is implicit, runs once per instance, and races. If the app needs a schema check at startup, make it a read-only check that the expected migration is present in the history table, failing readiness if not, so a skipped migration deploy shows up as an unhealthy app instead of a silently unmigrated database that half works.

Expected output: an app deploy with the opt-in unset changes no schema, and the audit log shows who set the opt-in and who approved, tied to the digest. Success check: deploying the app build alone to staging leaves the staging history table untouched.

6. Rehearse the exact artifact on staging

Before production, restore staging from a recent production-like backup and run the same artifact bytes (verified by the same hash check) through the same job, with the same runner identity model. Time it: the rehearsal duration and the lock footprint are the production estimate, and a migration that locks a hot table for minutes gets reworked (batched, or split into expand and contract steps across two releases) before it gets approved, not during the production window. Confirm the application build that pairs with the new schema works against the migrated staging database, including the previous app version if a rollback of the app alone must stay possible during the window.

Expected output: staging history table ends at the to migration, the paired app passes smoke checks, and the rehearsal timing is recorded. Success check: the artifact digest rehearsed on staging equals the digest approved for production.

7. Back up and preflight immediately before the production run

The production job starts by taking or verifying a fresh backup or snapshot of the production database and proving it is restorable in principle (backup completed, checksum or catalog entry present, restore point named in the run log). A backup that has not completed is a hard stop, not a warning. Preflight then checks, read-only: the current last row of the __EFMigrationsHistory table equals the from migration in the release record (any drift means someone or something migrated outside the gate, and the run stops for reconciliation), the target database is the intended one, and no other migration job is running. Only after backup and preflight pass does the job ask for the final approval confirmation if your platform separates approval from execution.

Expected output: a named restore point and a preflight log line matching the expected from migration. Success check: pointing the job at a database whose history has drifted makes it stop before any DDL runs.

8. Apply from a single runner under lock and transaction, then verify

Run the artifact from exactly one job instance. The EF Core migrator acquires a database lock that protects against concurrent migration application (an application lock on SQL Server, an advisory lock on PostgreSQL when migrations run through the migrator or bundle), which is one reason not to wrap Migrate in your own outer transaction: an external transaction prevents that lock from being acquired, and EF Core 9 turns that pattern into an error by default. If you execute the raw SQL script instead, take the equivalent lock yourself at the top of the session (the provider application-lock or advisory-lock call) and hold it for the whole script. Let each migration run in its own transaction where the provider supports transactional DDL, which is the EF Core default: a failed migration rolls itself back, earlier migrations in the script stay applied and recorded, and the idempotent script can simply be re-run after the fix. After execution, verify: the history table now ends at the to migration with every allowlisted ID recorded exactly once, the artifact digest in the run log matches the approved digest, and the paired app passes its readiness and smoke checks. Record all of it in the release record so the next release has a trustworthy from migration.

Expected output: history table at the to migration, a complete run log (digest, approver, backup name, duration), and a healthy app. Success check: running the same artifact a second time is a no-op that reports zero migrations applied, thanks to the idempotent guards.

When this applies

• EF Core migrations are going to a production database whose downtime or corruption would matter.
• More than one person or agent can propose migrations, so review and provenance of the exact SQL matter.
• App instances are replicated, so startup migration would race, or compliance needs a named approver per schema change.
• Releases are built as artifacts in CI and promoted across staging and production.

When this does not apply

• Local development and throwaway test databases: dotnet ef database update directly is fine there.
• A prototype with a single deployer and no real data, where the process cost exceeds the risk; keep at least the backup and idempotent script habits.
• Non-EF schema changes owned by a DBA team with its own change system: feed this artifact and hash into their system instead of running a parallel gate.
• Emergency hotfix migrations still go through the same gate, just faster; skipping the hash or backup under time pressure removes the protection exactly when mistakes are likeliest.

Variant: bundle instead of a SQL script

With a migration bundle there is no loose SQL file to execute, so the bundle binary is the artifact: hash the binary, review the SQL generated from the same frozen set, and rehearse the bundle itself on staging. Never rebuild the bundle in the deploy job; a rebuild is a new artifact and needs a new hash and approval.

Variant: PostgreSQL and other providers

The flow is provider-neutral, but the lock call differs (advisory lock functions on PostgreSQL, application lock on SQL Server) and transactional DDL coverage differs: some operations cannot run inside a transaction on some providers and need the suppress-transaction option in the migration, which the step 3 review must call out explicitly because those operations do not roll back on failure.

Variant: several DbContexts

Generate, allowlist, hash and approve one artifact per context, and run them in a recorded order in the same job under the same lock discipline. A single combined approval may cover the set, but each artifact keeps its own digest so a change to one context reopens approval only for that artifact.

Variant: expand and contract for zero-downtime releases

For changes the old app version cannot survive (dropped or renamed columns), split the release: expand migrations (additive only) with this deploy, ship the app, then contract migrations in a later gated release once no old instance remains. The allowlist for the expand release must contain no destructive statements; that is a mechanical check the reviewer can enforce.

Why it happens

EF Core makes applying migrations very easy (one command, or one call at startup), so teams drift into applying whatever the current code generates, wherever the deploy happens to run. That skips the properties production actually needs: the SQL was never read by a human, the bytes that ran are not the bytes anyone reviewed, two runners can interleave, a routine app deploy silently changes schema, and when a run fails halfway nobody can say precisely what applied. The gate moves all of that to build time and approval time, where it is cheap, and leaves the production job as a dumb, verified executor of one immutable artifact.

Edge cases and pitfalls

• Hash mismatch from regeneration: regenerating the script at deploy time on a different SDK, EF Core version or operating system can change bytes (line endings, ordering) even when the schema change is identical. That is a process failure, not a nuisance: never regenerate after approval; ship the approved bytes.
• Non-transactional operations: index builds with online options, certain raw SQL, and operations marked to suppress transactions do not roll back. The review must list them, and the backup from step 7 is their only real undo.
• Long locks on big tables: a rehearsal on a small staging database hides lock time. Rehearse against production-scale data volume for any migration touching a large table, and batch backfills separately from schema changes.
• Drift outside the gate: a single manual database update against production invalidates the from migration assumption for every later release. The step 7 preflight catches it; treat a catch as a process incident, not an annoyance to click through.
• Down migrations are not the recovery plan: reverting schema under live data is usually lossy. The recovery ladder is fix-forward with a new migration, then restore from the step 7 backup only if fix-forward is impossible, accepting the data-loss window that restore implies.
• Failed run, then app deploy: if the migration run fails partway, do not deploy the paired app build against a half-migrated schema unless the expand-and-contract split was designed for it. The readiness history check in step 5 exists to make this fail visibly.