ef-core-performance

name	ef-core-performance
description	Best practices for high-performance Entity Framework Core (EF Core 8/9): fast read queries AND fast writes/inserts, plus high-throughput, data-intensive apps (concurrency, parallel/background jobs, streaming large datasets, resiliency, optimistic concurrency, bulk insert). ALWAYS-ON for EF Core projects: if the codebase references any EF Core package (Microsoft.EntityFrameworkCore*, Npgsql.EntityFrameworkCore.PostgreSQL, Pomelo.EntityFrameworkCore.MySql, Oracle/MySql EF providers) or contains a DbContext, DbSet properties, OnModelCreating, or a Migrations folder, consult and apply this skill before writing, generating, reviewing, refactoring, or fixing ANY data-access code in that project — even when performance, N+1, slow SaveChanges, or bulk insert are not explicitly mentioned. All fixes must preserve original behavior (same output). Also use for "review my EF Core query", "why is this insert slow", and "scale my EF Core app".
license	MIT

High-Performance EF Core (Queries + Inserts)

This skill optimizes Entity Framework Core data access for throughput and low latency, covering the read path (queries), the write path (inserts, updates, bulk operations), and the demands of high-throughput, data-intensive applications (concurrency, parallel processing, resiliency, streaming). It targets EF Core 8 and 9; notes call out where a feature needs a specific version.

The golden rule of EF Core performance: EF generates SQL, and the database does the work. Most "EF is slow" problems are really "EF was told to do the wrong thing." So the first move is always to see the SQL EF actually runs, then fix the cause — not to reach for a bulk library or raw SQL prematurely.

Offline & multi-agent use

This skill is fully self-contained: all guidance lives in this file and in references/, so it works with no internet access and no external tools — read the local files and apply them directly. For a fast lookup table and source attribution see references/offline-reference.md. Operational and correctness depth (migrations, inheritance mapping, testing, async bugs, observability, pagination, EF-version shifts, read replicas, non-relational providers) is in references/correctness-and-ops.md. To install the skill in other LLM coding agents (Claude Code, Cursor, GitHub Copilot, Cline, Windsurf, Codex, OpenCode, Gemini CLI, and others) and to force it on for an EF Core repo, see references/INSTALL.md.

Context & token management (apply the skill cheaply)

Applying this skill should not blow up the context window or token cost. Work narrowly.

Context management:

• Progressive disclosure. This SKILL.md body is enough for almost every fix. Open references/offline-reference.md only for a fast lookup, and references/INSTALL.md only when installing — do not load all references up front.
• Read narrowly, not the whole solution. Detect EF Core from a single *.csproj / Directory.Packages.props, then open only the files in scope: the specific DbContext, the query/repository/method under review, and its entity config. Skip unrelated projects, generated migrations, and *ModelSnapshot.cs unless the task needs them.
• Sample, don't dump. When inspecting SQL or query logs, read a representative slice (the N+1 pattern, the one slow statement) rather than pasting the entire log.
• Reference by path/line; don't echo large files back into the conversation.

LLM / token management:

• Scope to what was asked. Fix the hot path the user named; don't silently refactor the whole data layer. Unsolicited rewrites cost tokens and risk behavior changes.
• Emit targeted edits, not full-file rewrites. Prefer minimal diffs to the lines that change.
• Reuse the decision tables here instead of re-deriving the reasoning each turn, and reuse facts already established this session (detected provider, tracking default) rather than re-checking.
• Batch related fixes into one pass so files aren't re-read repeatedly.
• Report concisely: name the practice applied, the before/after, and the behavior-equivalence note — skip restating unchanged code.

Activation — apply automatically on any EF Core project

This skill is not opt-in. Before writing or changing data-access code, check whether the project uses EF Core, and if it does, apply this skill's practices by default.

Detect EF Core by scanning for any of:

• A *.csproj, Directory.Packages.props, or packages.config referencing an EF Core package: Microsoft.EntityFrameworkCore, ...SqlServer, ...Sqlite, ...InMemory, ...Cosmos, ...Design, Npgsql.EntityFrameworkCore.PostgreSQL, Pomelo.EntityFrameworkCore.MySql, Oracle.EntityFrameworkCore, MySql.EntityFrameworkCore.
• A class deriving from DbContext, any DbSet<T> property, or using Microsoft.EntityFrameworkCore;.
• A Migrations/ folder, *ModelSnapshot.cs, or OnModelCreating / IEntityTypeConfiguration<T>.

If any of these are present, follow this skill whenever you author, generate, review, refactor, or fix data-access code in the project — even if the user only asked for a feature, a bug fix, or "just add this query" and never mentioned performance. When you apply a practice from here, briefly say which one and why, so the change is reviewable.

(Triggering is ultimately a runtime heuristic, not a hard lock — this section maximizes the chance the skill is consulted, and once consulted it governs the work.)

Non-negotiable: fixes must preserve behavior (same output)

Every change this skill motivates is a performance/clarity refactor, never a behavior change. The optimized code must produce the same observable result as the original: the same rows, the same values, the same ordering, the same null/empty handling, the same number and type of side effects (rows inserted/updated/deleted), and the same thrown-exception contract. Speed is the only thing allowed to change.

Before proposing a fix, verify equivalence against this list:

• Result set: same entities/columns, same filtering, same null vs empty results.
• Ordering: preserve any OrderBy; if the original relied on incidental ordering, add an explicit OrderBy rather than silently changing it.
• Duplicates/identity: if switching to AsNoTracking(), use AsNoTrackingWithIdentityResolution() where the original tracking collapsed duplicate references, so object graphs stay identical.
• Write effects: an insert/update/delete refactor must affect exactly the same rows. ExecuteUpdate/ExecuteDelete and bulk-copy paths bypass the change tracker, so they skip interceptors, SavingChanges events, computed/generated values, and in-context cache updates — only use them where those were not relied upon.
• Transactions: keep the same atomicity. Don't split one transactional SaveChanges into several statements that could partially apply.
• Concurrency: preserve concurrency-token checks; don't drop optimistic-concurrency behavior in the name of speed.

If a faster approach would change any of the above, do NOT apply it silently. Either keep behavior identical, or call out the exact difference and let the user opt in. When practical, confirm equivalence by comparing the generated SQL and/or running the existing tests before and after.

When to use

• Queries are slow, return too much data, or emit far more SQL statements than expected.
• N+1 query patterns appear in logs (one query, then one more per row).
• SaveChanges / SaveChangesAsync is slow when inserting or updating many rows.
• The user needs to import / seed / bulk-insert thousands+ of rows.
• Database CPU/IO or app memory is high and EF Core is in the path.
• The app runs under high concurrency / load: parallel or background data jobs, connection-pool exhaustion, transient/deadlock failures, or concurrency conflicts.
• The user asks to review or refactor EF Core repository / query / persistence code.

When NOT to use

• The code uses Dapper, raw ADO.NET, or another ORM (not EF Core).
• The bottleneck is purely database-side schema (missing indexes, bad data types, lock contention) with no EF-side fix — diagnose, then hand off to DBA work.
• The user wants help designing the schema/model from scratch with no perf angle.

Step 0 — Always measure first: see the real SQL

Never optimize blind. Turn on query logging and read what EF emits.

optionsBuilder
    .UseSqlServer(connectionString)              // or .UseNpgsql(...)
    .LogTo(Console.WriteLine, LogLevel.Information)
    .EnableSensitiveDataLogging()                // shows parameter values — DEV ONLY
    .EnableDetailedErrors();                      // DEV ONLY

Or via appsettings.json:

{ "Logging": { "LogLevel": {
    "Microsoft.EntityFrameworkCore.Database.Command": "Information"
} } }

What to look for: a loop of nearly identical SELECTs (N+1), one giant cartesian JOIN returning duplicated rows, SELECT * when only two columns are needed, or a "client evaluation" warning meaning EF pulled data into memory to filter it in C#. For production diagnosis prefer EF interceptors, dotnet-trace, or the database's own query store rather than EnableSensitiveDataLogging.

---

Part A — Fast queries (the read path)

A1. Kill N+1 — the single biggest EF Core query killer

N+1 happens when related data is loaded one parent at a time inside a loop (classic with lazy loading enabled).

// BAD: 1 query for orders + N queries (one per order) for items
var orders = await db.Orders.ToListAsync();
foreach (var o in orders)
    total += o.Items.Count;          // each access hits the DB

Fix it by telling EF what you need up front:

// Option 1 — Include (one JOIN). Good for a single, small related set.
var orders = await db.Orders.Include(o => o.Items).ToListAsync();

// Option 2 — Split query. Avoids cartesian explosion across multiple/large Includes.
var orders = await db.Orders
    .Include(o => o.Items)
    .Include(o => o.Payments)
    .AsSplitQuery()
    .ToListAsync();

// Option 3 — Projection (BEST when you don't need full entities). Fetches only
// the columns you use, no tracking, no extra rows.
var summaries = await db.Orders
    .Select(o => new OrderSummary {
        OrderId   = o.Id,
        Total     = o.Items.Sum(i => i.Price),
        ItemCount = o.Items.Count
    })
    .ToListAsync();

Single vs split query — pick by data shape:

Scenario	Use
One Include, small child set	Single query (default)
Multiple Includes (cartesian risk)	AsSplitQuery()
Include of a large child collection	AsSplitQuery()
Need a consistent snapshot in one transaction	Single query

Split queries run several round-trips, so without an explicit transaction the child rows can drift if the data changes between them — choose single query when consistency matters more than the cartesian cost.

A2. Use AsNoTracking() for read-only queries

Change tracking exists so EF can detect edits for SaveChanges. If you are only displaying data, that bookkeeping is pure overhead (memory + CPU).

var products = await db.Products
    .AsNoTracking()
    .Where(p => p.IsActive)
    .ToListAsync();

For read-heavy apps, make it the default and opt back in only where you write:

services.AddDbContext<AppDbContext>(o => o
    .UseSqlServer(cs)
    .UseQueryTrackingBehavior(QueryTrackingBehavior.NoTracking));

Use AsNoTrackingWithIdentityResolution() when a no-tracking query can return the same entity multiple times (e.g. many parents sharing one child) and you want a single in-memory instance instead of duplicates.

Note: projecting to a DTO with Select is already effectively untracked for the projected shape, so AsNoTracking() adds little on top of a pure projection.

A3. Project to exactly what you need

Pulling whole entities to use two fields wastes I/O, memory, and tracking. Shape the query to the DTO/view model. This also sidesteps over-fetching wide tables and large columns (e.g. varchar(max), blobs) you never display.

A4. Compile queries on hot paths

For queries executed very frequently with the same shape, cache the compiled plan to skip EF's per-call expression compilation.

private static readonly Func<AppDbContext, int, Task<Order?>> GetOrderById =
    EF.CompileAsyncQuery((AppDbContext db, int id) =>
        db.Orders.Include(o => o.Items).FirstOrDefault(o => o.Id == id));

var order = await GetOrderById(db, orderId);

A5. Common query traps

Trap	Problem	Fix
ToList() then Where() in memory	Loads the whole table	Filter first: .Where(...).ToListAsync()
Count() > 0 to test existence	Counts every row	await q.AnyAsync()
.Select(...) after .Include(...)	Include is ignored once you project	Drop the Include; project the related data directly
.ToList() inside a Select	Triggers nested per-row queries	Keep projecting with Select all the way down
Non-translatable method in Where	Falls back to client evaluation	Use EF.Functions.Like(...), or rework the predicate
Unbounded result set	Memory blowup, slow first byte	Page with .Skip()/.Take(), or stream with AsAsyncEnumerable()
OFFSET paging deep into a table	Slow on large offsets	Prefer keyset/seek paging on an indexed sort key

A6. Drop to raw SQL when LINQ can't express it well

var rows = await db.Orders
    .FromSqlInterpolated($@"
        SELECT o.* FROM Orders o
        JOIN (SELECT OrderId, SUM(Price) Total FROM OrderItems
              GROUP BY OrderId HAVING SUM(Price) > {minTotal}) t
          ON o.Id = t.OrderId")
    .AsNoTracking()
    .ToListAsync();

Use FromSqlInterpolated / FromSql (parameterized) — never build SQL by string-concatenating user input into FromSqlRaw; that is a SQL-injection hole.

A7. Project on the way in too — DTOs as a security boundary

Projection isn't only a read optimization; for writes it's a safety boundary. Binding an HTTP request body straight onto an EF entity (TryUpdateModel/auto-mapping the whole entity) is over-posting / mass assignment: a caller can set fields they shouldn't, e.g. IsAdmin, Balance, OwnerId, or a foreign key. Accept a request DTO with only the fields a client may set, then copy those explicitly onto the entity. This also keeps sensitive/PII columns out of read payloads — project entities to response DTOs rather than serializing entities directly. Treat "what the client can read/write" as a deliberate, narrow mapping in both directions.

---

Part B — Fast inserts and bulk writes (the write path)

EF Core's tracker is built for correctness of small unit-of-work changes, not for shoveling 100k rows. Insert performance falls off a cliff when people fight that design. Escalate through these tiers — only go further when a tier is too slow.

B1. Tier 1 — Batch with one SaveChanges, not one per row

The most common insert mistake is saving inside the loop.

// BAD: N round-trips, tracker churns on every save
foreach (var dto in dtos) {
    db.Customers.Add(new Customer { Name = dto.Name });
    await db.SaveChangesAsync();      // <-- inside the loop
}

// GOOD: stage everything, save once. EF batches the INSERTs automatically.
db.Customers.AddRange(dtos.Select(d => new Customer { Name = d.Name }));
await db.SaveChangesAsync();          // one call, batched statements

AddRange + a single SaveChanges lets EF group inserts into batched commands (SQL Server defaults to batching many rows per round-trip; the batch size is provider-tunable via MaxBatchSize). This alone fixes most "insert is slow" cases up to a few thousand rows.

B2. Tier 2 — Turn off the tracker's per-add cost for large batches

When adding many entities in one context, DetectChanges can run repeatedly and dominate CPU. For a bounded bulk operation, disable auto-detect, add everything, then save:

db.ChangeTracker.AutoDetectChangesEnabled = false;
try {
    db.Customers.AddRange(bigList);   // no per-add change scan
    await db.SaveChangesAsync();
}
finally {
    db.ChangeTracker.AutoDetectChangesEnabled = true;
}

Also: use a fresh, short-lived DbContext for a big import and don't let one context accumulate hundreds of thousands of tracked entities — tracking state grows and slows down. For very large jobs, chunk the input (e.g. 5k–10k rows), and either clear the context (db.ChangeTracker.Clear()) or new up a context per chunk.

B3. Tier 3 — Set-based updates/deletes without loading rows (EF Core 7+)

To change or remove many rows, do not load them into memory just to modify and save. Use ExecuteUpdate / ExecuteDelete, which emit a single set-based SQL UPDATE/DELETE and bypass the tracker entirely.

// One UPDATE statement, no entities loaded
await db.Orders
    .Where(o => o.Status == OrderStatus.Pending && o.CreatedUtc < cutoff)
    .ExecuteUpdateAsync(s => s
        .SetProperty(o => o.Status,      OrderStatus.Expired)
        .SetProperty(o => o.UpdatedUtc,  DateTime.UtcNow));

// One DELETE statement
await db.AuditLogs
    .Where(a => a.CreatedUtc < retentionCutoff)
    .ExecuteDeleteAsync();

Caveat: these run immediately and outside SaveChanges, so they don't fire SaveChanges interceptors/SavingChanges events and won't update entities already tracked in the current context. Wrap in an explicit transaction if you combine them with other writes.

B4. Tier 4 — True bulk insert for very large volumes

EF's batched INSERTs are fine into the low thousands, but for tens of thousands to millions of rows, use a bulk-copy path. EF Core has no built-in BulkInsert; choose one of:

• A maintained bulk-extensions library (e.g. community EFCore.BulkExtensions for SQL Server/PostgreSQL/SQLite, or a provider-specific bulk package). These add BulkInsert/BulkUpdate/BulkMerge that use the database's fast path. Vet the library's license and maintenance before adding it to a project.
• The provider's native bulk API directly, which is the fastest and dependency-light option:
  • SQL Server: SqlBulkCopy (stream a DataTable/IDataReader).
  • PostgreSQL: Npgsql binary COPY (BeginBinaryImport).

// SQL Server native bulk copy — fastest path for huge inserts
using var bulk = new SqlBulkCopy(connectionString) {
    DestinationTableName = "Customers",
    BatchSize = 10_000
};
bulk.ColumnMappings.Add(nameof(Customer.Name), "Name");
await bulk.WriteToServerAsync(dataReader);

These bypass EF's tracking, validation, and value generation, so generated keys, shadow properties, and concurrency tokens are your responsibility. Use them for ingestion/seeding/ETL, not for ordinary domain writes.

B5. Insert decision guide

Volume / situation	Recommended approach
A handful of rows	Add/AddRange + one SaveChangesAsync
Hundreds–low thousands	AddRange + one SaveChanges (let EF batch)
Thousands, tracker is the cost	Disable AutoDetectChangesEnabled, chunk, fresh context
Bulk modify/remove existing rows	ExecuteUpdateAsync / ExecuteDeleteAsync
Tens of thousands → millions	SqlBulkCopy / Npgsql COPY / bulk-extensions library

---

Part C — Context & connection hygiene (helps reads and writes)

• DbContext lifetime is short and scoped. Register with AddDbContext (per-request/scoped). Never share one context across threads or cache it for the app lifetime — DbContext is not thread-safe and accumulates tracked state.
• Pool contexts on hot paths. AddDbContextPool<T>(...) reuses context instances to cut per-request allocation/setup. Don't store request state in the context if you pool it.
• Always use the async APIs (ToListAsync, SaveChangesAsync, etc.) in server apps so threads aren't blocked on I/O.
• Index for the query, not the table. EF won't add the index your Where/ OrderBy/join needs — confirm covering indexes exist on the columns you filter and sort by. Configure with HasIndex(...) and verify in migrations.
• Use the right key/value-generation strategy for high-insert tables. Default identity columns serialize on insert hotspots less than some alternatives; for distributed inserts consider sequential GUIDs or HiLo to reduce index fragmentation and contention.

---

Part D — High-throughput, data-intensive applications

When an EF Core app serves heavy concurrent load or churns through large datasets, the bottleneck shifts from "one slow query" to resiliency, concurrency, and keeping the connection pool healthy. Apply these on top of Parts A–C.

D1. Connection resiliency — survive transient faults under load

Cloud/managed databases drop connections, throttle, and time out far more often under load. Enable the retrying execution strategy so transient failures auto-retry with backoff instead of bubbling up as errors.

services.AddDbContext<AppDbContext>(o => o
    .UseSqlServer(cs, sql => sql.EnableRetryOnFailure(
        maxRetryCount: 5,
        maxRetryDelay: TimeSpan.FromSeconds(10),
        errorNumbersToAdd: null)));   // Npgsql: o.UseNpgsql(cs, n => n.EnableRetryOnFailure())

Gotcha: a retrying strategy cannot wrap a user-initiated transaction implicitly, because a retry must re-run the entire unit. When you open an explicit transaction, run it inside the strategy so the whole block is retried atomically:

var strategy = db.Database.CreateExecutionStrategy();
await strategy.ExecuteAsync(async () =>
{
    await using var tx = await db.Database.BeginTransactionAsync();
    // ... multiple SaveChanges / ExecuteUpdate calls ...
    await tx.CommitAsync();
});

D2. DbContext is not thread-safe — use a pooled factory for parallel work

One DbContext instance must never be touched by two operations concurrently (you'll get corruption or InvalidOperationException). For parallel/background processing, give each unit its own context from a pooled factory:

services.AddPooledDbContextFactory<AppDbContext>(o => o.UseSqlServer(cs));

// In a parallel worker / channel consumer:
await Parallel.ForEachAsync(batches, async (batch, ct) =>
{
    await using var db = await factory.CreateDbContextAsync(ct);
    db.Items.AddRange(batch);
    await db.SaveChangesAsync(ct);
});

Pooling reuses context instances to cut per-operation allocation and setup — valuable on hot request paths. Don't stash per-request mutable state on a pooled context, and bound your parallelism: too many concurrent contexts will exhaust the ADO.NET connection pool (default Max Pool Size is 100 on SQL Server). Tune Max Pool Size in the connection string and watch for pool-timeout errors.

D3. Stream large result sets — don't buffer millions of rows

ToListAsync() materializes everything into memory at once. For large reads (reports, ETL, exports), stream row-by-row so memory stays flat:

await foreach (var row in db.Orders.AsNoTracking()
                              .Where(o => o.CreatedUtc >= from)
                              .AsAsyncEnumerable()
                              .WithCancellation(ct))
{
    await ProcessAsync(row);   // one row in flight, not the whole table
}

Combine with AsNoTracking() (or projection) so the tracker doesn't accumulate state across the stream. For chunked writes during a stream, flush every N rows and ChangeTracker.Clear() (or use a fresh context per chunk).

D4. Compiled models — cut cold-start cost for large schemas

Apps with hundreds of entity types pay a noticeable model-building cost on first use (slow cold starts, painful in serverless/autoscaled tiers). Pre-compile the model so startup skips that work:

dotnet ef dbcontext optimize --output-dir Compiled --namespace MyApp.Compiled

optionsBuilder.UseModel(MyApp.Compiled.AppDbContextModel.Instance);

Regenerate the compiled model whenever the entity model changes (wire it into the build). Small models barely benefit — this pays off at scale.

D5. Handle optimistic concurrency under contention

When many writers hit the same rows, use a concurrency token (rowversion / xmin on PostgreSQL) so last-write-wins corruption is caught instead of silently overwriting. Configure the token, then handle the conflict with a retry/merge loop:

// Model: builder.Property(o => o.RowVersion).IsRowVersion();
for (var attempt = 0; attempt < 3; attempt++)
{
    try { await db.SaveChangesAsync(); break; }
    catch (DbUpdateConcurrencyException ex)
    {
        foreach (var entry in ex.Entries)
        {
            var current = await entry.GetDatabaseValuesAsync();
            if (current is null) { /* row deleted — decide policy */ }
            else entry.OriginalValues.SetValues(current);   // reload + reapply, then loop
        }
    }
}

For write-heavy hotspots, also consider set-based ExecuteUpdate (Part B3) which sidesteps load-then-save races entirely.

D6. Timeouts, cancellation, and isolation

• Pass a CancellationToken to every async EF call so overloaded requests can be shed instead of piling up and exhausting the pool.
• Set a command timeout for legitimately long operations: db.Database.SetCommandTimeout(TimeSpan.FromMinutes(2)); — but prefer fixing the query over raising timeouts.
• Pick an isolation level deliberately for hot tables. Lower isolation (e.g. SQL Server read-committed snapshot / READ COMMITTED SNAPSHOT) reduces reader↔ writer blocking under load; specify it on explicit transactions when needed.
• Tune batch size for large writes via o.UseSqlServer(cs, s => s.MaxBatchSize(...)) when the default round-trip batching isn't optimal for your row width.

D7. Caching — EF Core has no second-level cache

EF Core caches compiled query plans, not results. Repeatedly hitting the database for the same rarely-changing data is a common scale bottleneck. Options: distributed cache (IDistributedCache/Redis) in front of read queries you control, or a maintained second-level-cache library. Cache only safely-stale data, and have an explicit invalidation story — a stale cache under load is worse than a slow query.

---

Part E — Correctness & operations (load the reference when relevant)

Fast but wrong is still wrong, and a fast query path can be undone by a bad migration or a test that can't catch regressions. When the task touches any of the following, read references/correctness-and-ops.md and apply it:

• Migrations in production — no startup auto-migrate races; expand/contract for zero-downtime; idempotent deploy scripts.
• Inheritance mapping — TPH (fast, default) vs TPT (join-heavy, often slow) vs TPC.
• Testing that can catch regressions — avoid the InMemory provider's false confidence; use the real provider (Testcontainers) or SQLite in-memory.
• Async/threading bugs — no sync-over-async (.Result/.Wait()); one operation per DbContext; flow CancellationToken.
• Observability — .TagWith(...) to correlate LINQ↔SQL; OpenTelemetry/metrics.
• Pagination correctness — stable OrderBy; keyset paging for deep pages.
• EF version shifts — Contains/collection translation changed across EF7→8→9→10; re-profile list-filter queries after a major upgrade.
• Read replicas / CQRS routing, and non-relational providers (Cosmos is governed by request units + partition key, not relational query plans).

---

Validation checklist

Before declaring an EF Core path optimized, confirm:

• The optimized code is behavior-equivalent to the original: same rows, values, ordering, null handling, write effects, and transaction atomicity (speed only).
• Query logging was inspected; the SQL count matches expectations (no N+1).
• Read-only queries use AsNoTracking() or project to a DTO.
• Include vs AsSplitQuery matches the data shape (no cartesian blowup).
• No client-side evaluation warnings remain in the logs.
• Inserts/updates of many rows do one SaveChanges, not one per row.
• Bulk modify/delete uses ExecuteUpdate/ExecuteDelete, not load-then-save.
• Very large imports use a real bulk path (SqlBulkCopy / COPY / bulk lib).
• DbContext is scoped/short-lived; not shared across threads.
• Parallel/background work uses a context-per-unit from a pooled factory.
• Transient faults are handled (EnableRetryOnFailure); explicit transactions run inside the execution strategy.
• Large reads stream (AsAsyncEnumerable) instead of buffering whole tables.
• Write hotspots use a concurrency token and handle DbUpdateConcurrencyException.
• CancellationToken flows through async calls; connection pool size is sized for load.
• Filter/sort/join columns are actually indexed.
• Raw SQL is parameterized (FromSqlInterpolated, never concatenated input).
• Request bodies aren't bound straight onto entities (no over-posting); DTOs used.
• Pagination has a stable OrderBy; production migrations don't auto-run on startup.
• Tests run against a provider that can catch regressions (not the InMemory provider).

Common pitfalls

Pitfall	Fix
Lazy loading silently causing N+1	Disable lazy-loading proxies; load explicitly with Include/projection
Forgotten global query filters skewing analysis	Check HasQueryFilter; use IgnoreQueryFilters() when intended
One DbContext kept alive / tracking 100k entities	Scope it; chunk imports; ChangeTracker.Clear() or new context per chunk
Bulk update by load → edit → SaveChanges	ExecuteUpdateAsync / ExecuteDeleteAsync
SaveChanges inside the insert loop	Stage with AddRange, save once
SQL injection via FromSqlRaw + string interpolation	Use FromSqlInterpolated (parameterized)
Expecting EF to add indexes for you	Define indexes in the model; verify in migrations
Bulk-copy path expected to set generated keys / fire interceptors	It won't — handle keys/validation yourself, reserve it for ETL
Sharing one DbContext across parallel tasks	Context-per-unit via AddPooledDbContextFactory
Explicit transaction silently not retried	Run it inside CreateExecutionStrategy().ExecuteAsync(...)
Connection-pool timeout errors under load	Bound parallelism; raise Max Pool Size; shed load with cancellation
Buffering huge result sets with ToList	Stream with AsAsyncEnumerable, page, or project
Expecting EF to cache query results	It caches plans only — add an explicit distributed/second-level cache
Lost updates on contended rows	Add a rowversion token; handle DbUpdateConcurrencyException