Don't clear ChangeTracker on EnsureCreated/Migrate retry; skip seeder if it already ran

[Copilot]

Follow-up to #38211. Calling ChangeTracker.Clear() from inside EnsureCreated[Async] and Migrate[Async] on retry is too invasive — it wipes user-tracked state as a side effect of a transient failure elsewhere. The original concern (avoiding double-seeding when the execution strategy retries) is better addressed by simply not re-invoking the seeder once it has already completed successfully.

• CosmosDatabaseCreator.EnsureCreatedAsync: dropped the Retrying flag and the ChangeTracker.Clear() call; added a Seeded StrongBox that gates SeedDataAsync so the async seeder runs at most once across retries.
• RelationalDatabaseCreator.EnsureCreated / EnsureCreatedAsync: same treatment — removed the retry branch that cleared the change tracker, and short-circuit the execution-strategy callback when Seeded is already set.
• Migrator.MigrateImplementation / MigrateImplementationAsync: same treatment — removed the retry branch that cleared the change tracker, and skip the seeder when it has already completed. MigrationExecutionState.SeedingAttempted is renamed to SeedingCompleted (and the API baseline updated) to reflect that the flag is only set after a successful seed.

The seeded/completed flag is set only after the seeder (and, for the relational path, the surrounding transaction commit) succeeds, so a failure inside the seeder itself will still be retried.

[AndriySvyryd]

@copilot Also apply this logic in Migrate(Async)Implementation

[Copilot]

@copilot Also apply this logic in Migrate(Async)Implementation

Done in 4cdf8d2. Removed the ChangeTracker.Clear() branch from both MigrateImplementation and MigrateImplementationAsync, and renamed MigrationExecutionState.SeedingAttempted to SeedingCompleted (set only after a successful seed) so the seeder is skipped on retry once it has completed.

[Copilot]

Pull request overview

Follow-up to #38211 that walks back the ChangeTracker.Clear() calls introduced into EnsureCreated[Async] and Migrate[Async] retry paths (which were too invasive on user-tracked state) and instead prevents double-seeding by tracking whether the seeder has already completed.

Changes:

• Replace the per-retry ChangeTracker.Clear() branches in RelationalDatabaseCreator and CosmosDatabaseCreator with a Seeded StrongBox flag that short-circuits the seeder on retries.
• In Migrator, gate seeder invocation on !state.SeedingCompleted and set the flag after the seeder runs; rename MigrationExecutionState.SeedingAttempted → SeedingCompleted and update the public API baseline.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/EFCore.Relational/Storage/RelationalDatabaseCreator.cs	Replaces Retrying/ChangeTracker.Clear() with a Seeded flag short-circuit in sync/async EnsureCreated.
src/EFCore.Relational/Migrations/MigrationExecutionState.cs	Renames SeedingAttempted to SeedingCompleted and updates the XML doc.
src/EFCore.Relational/Migrations/Internal/Migrator.cs	Gates sync/async seeder execution on !SeedingCompleted and sets the flag after seeding.
src/EFCore.Relational/EFCore.Relational.baseline.json	Updates API baseline for the property rename.
src/EFCore.Cosmos/Storage/Internal/CosmosDatabaseCreator.cs	Removes retry-clear logic and adds Seeded flag guarding SeedDataAsync.

Comments suppressed due to low confidence (1)

src/EFCore.Relational/Migrations/Internal/Migrator.cs:339

• Same issue as in MigrateImplementation: state.SeedingCompleted = true is set before the transaction commit on line 348. If the async commit fails transiently, the retry will skip the seeder despite the seeded rows having been rolled back. The flag should be set after CommitAsync succeeds, consistent with the PR description.

                if (!state.SeedingCompleted)
                {
                    await seedAsync(context, state.AnyOperationPerformed, cancellationToken).ConfigureAwait(false);
                    state.SeedingCompleted = true;
                }