EF Core Interview Questions (Concurrency, Filters, SaveChanges)

EF Core interview questions at senior level are not “what is Include.” Interviewers ask whether you can keep a clinic or marketplace database correct when two users save the same row, when a tenant filter is missing, and when a background job shares a DbContext.

This URL is interview narration. SQL-shaped performance (N+1, split queries, cartesian explosion, parameter sniffing, tracking vs identity resolution) lives in the EF Core hub. Do not retell those posts here. If the prompt is “the dashboard is slow,” start at EF Core SQL performance and come back when they ask about correctness.

---

Scenario 1: Two clinicians save the same encounter

Prompt: Two Angular tabs load the same encounter. Both PATCH status. Last write wins. How do you make that a conflict the API can show, not silent data loss?

Weak answer

“Add a lock” with no row version, or “use a transaction” as if wrapping two independent HTTP requests in one TransactionScope were possible.

Strong answer

Optimistic concurrency with a RowVersion / xmin / timestamp column that EF maps as a concurrency token. Each GET returns the token. Each PUT sends it back. The second save throws DbUpdateConcurrencyException. Map that to 409 with ProblemDetails, not 500.

modelBuilder.Entity<Encounter>()
    .Property(e => e.RowVersion)
    .IsRowVersion();

try
{
    await _db.SaveChangesAsync(cancellationToken);
}
catch (DbUpdateConcurrencyException)
{
    throw new ConcurrencyConflictException(encounterId);
}

Prove it: two integration tests against a real SQL database (or Testcontainers). Same Id, different RowVersion. One succeeds, one 409s. In-memory provider is a liar for this.

Refuse: pessimistic UPDLOCK on every read for a SPA. Clinicians leave tabs open. You lock the row for the lifetime of a coffee.

---

Scenario 2: Global query filter that leaked a tenant

Prompt: HasQueryFilter(e => e.ClinicId == _tenant.ClinicId) is on Encounter. A nightly job emailed clinic B a CSV of clinic A’s encounters. How?

Weak answer

“Filters are always applied.” They are not.

Strong answer

I list the escape hatches out loud:

1. IgnoreQueryFilters() on a report query someone copied from a debug session
2. Raw SQL / FromSqlRaw that never saw the filter
3. A DbContext constructed with a default tenant (Guid.Empty or first clinic in the table) because the hosted service has no HTTP scope
4. Navigation includes from an unfiltered parent (Clinic loaded without a filter, then .Encounters accessed in memory)

Fix the job: create a scope, set tenant from the job payload, or use an explicit Where(e => e.ClinicId == clinicId) on exports and log if IgnoreQueryFilters appears in a PR.

I do not pretend a filter is a security boundary. Authorization still lives in policies. The filter is a seatbelt.

---

Scenario 3: SaveChanges inside a foreach

Prompt: Import 4,000 fee-schedule rows. The code does Add + SaveChangesAsync per row. The request times out. What do you change, and what do you not change?

Weak answer

“Use AddRange and one SaveChanges” as the entire answer, or “bulk insert library” without saying when.

Strong answer

One SaveChanges per HTTP request is the default. Per-row save is N round-trips plus N change-tracker snapshots.

For a true bulk import I want ExecuteInsert / a bulk copy path, not 4,000 tracked entities. For a small upsert I batch:

_db.FeeSchedules.AddRange(rows);
await _db.SaveChangesAsync(cancellationToken);

When the rows already exist, I prefer ExecuteUpdate for a set-based column change, not load-mutate-save:

await _db.FeeSchedules
    .Where(f => f.ClinicId == clinicId && f.Year == year)
    .ExecuteUpdateAsync(
        s => s.SetProperty(f => f.IsActive, false),
        cancellationToken);

Refuse: SaveChanges in a loop “so we can return partial success” without a strategy. Either a transaction with a report of failures, or a staging table. Healthcare fee files that half-apply are worse than a 500.

---

Scenario 4: Same entity, two tracked instances

Prompt: A PUT loads the encounter, then _db.Update(dtoMappedEntity) and SaveChanges throws: another instance with the same key is already being tracked. What happened?

Weak answer

“Call AsNoTracking on the PUT.” Tracking on a write path is not the bug.

Strong answer

You loaded a tracked instance, then attached a second instance with the same primary key (mapped from the Angular body). EF cannot have two tracked copies.

Pick one:

1. Load, mutate the tracked entity, save
2. Or attach/update without a prior tracked load (ExecuteUpdate, or Update on a disconnected graph with no previous Find)

var encounter = await _db.Encounters.FirstAsync(e => e.Id == id, cancellationToken);
encounter.Status = body.Status;
await _db.SaveChangesAsync(cancellationToken);

Identity resolution on reads (AsNoTrackingWithIdentityResolution) is a different article: AsNoTracking vs identity resolution. Do not mix that story into a PUT.

---

Scenario 5: DbContext in a singleton cache

Prompt: A ReportCache singleton injects AppDbContext. Intermittent “cannot access a disposed object” and, once, the wrong clinic’s numbers. Diagnose.

Weak answer

“Make everything singleton” or “make everything scoped.”

Strong answer

This is a captive dependency. DbContext is scoped. A singleton holds the first request’s context (or a disposed one). Tenant data rides along.

Fix: cache is singleton or scoped; context stays scoped. If the cache must be singleton, inject IServiceScopeFactory and open a scope per refresh. Full lifetime rules: DI lifetimes. If the exception is actually “Unable to resolve service for type,” that is the registration post.

---

Scenario 6: Migrations in production

Prompt: App Service swap. API 500s: pending model changes or could not apply migration. How do you ship schema?

Weak answer

“Migrate() in Program.cs” as a religion, or “never migrate in prod” without an alternative.

Strong answer

I separate schema from app start:

• Generate SQL (dotnet ef migrations script) and run it in the release pipeline before the new bits take traffic
• Or a dedicated migrator job with a lock so two instances do not race
• Database.Migrate() at startup is acceptable for a single-instance internal tool, not for a slot swap with three instances

Pending model changes at runtime means someone edited entities and skipped dotnet ef migrations add. That is a CI gate, not a try/catch.

---

Rapid-fire (30 seconds)

Prompt	Answer I want
Why not the InMemory provider for concurrency tests?	No real transactions, no row versions you can trust
EnsureCreated vs migrations	EnsureCreated skips the migration history; do not mix in one database
Compiled queries	Hot path with a stable shape; not a substitute for indexes
Returning IQueryable from a repository	Callers can still IgnoreQueryFilters and Include the world
AsNoTracking on a write	You will reattach; usually the wrong default for PUT

---

How this differs from the other EF posts

• This page: correctness under concurrent writes, tenant filters, SaveChanges shape
• EF Core SQL performance: N+1, projections, indexes
• N+1 vs Include vs AsSplitQuery: which SQL you meant
• Cartesian explosion: two collection Includes
• Parameter sniffing: one tenant fast, another slow

Related reading

• ASP.NET Core interview scenarios
• C# expert-level interview questions
• Interview questions hub
• EF Core topic hub

Preparing an EF Core loop for a healthcare or marketplace API, or sitting one? Contact me. Bring a SaveChanges failure, not a definition of DbSet.