Redaguoti

Bendrinti naudojant


Advanced Performance Topics

DbContext pooling

A DbContext is generally a light object: creating and disposing one doesn't involve a database operation, and most applications can do so without any noticeable impact on performance. However, each context instance does set up various internal services and objects necessary for performing its duties, and the overhead of continuously doing so may be significant in high-performance scenarios. For these cases, EF Core can pool your context instances: when you dispose your context, EF Core resets its state and stores it in an internal pool; when a new instance is next requested, that pooled instance is returned instead of setting up a new one. Context pooling allows you to pay context setup costs only once at program startup, rather than continuously.

Note that context pooling is orthogonal to database connection pooling, which is managed at a lower level in the database driver.

The typical pattern in an ASP.NET Core app using EF Core involves registering a custom DbContext type into the dependency injection container via AddDbContext. Then, instances of that type are obtained through constructor parameters in controllers or Razor Pages.

To enable context pooling, simply replace AddDbContext with AddDbContextPool:

builder.Services.AddDbContextPool<WeatherForecastContext>(
    o => o.UseSqlServer(builder.Configuration.GetConnectionString("WeatherForecastContext")));

The poolSize parameter of AddDbContextPool sets the maximum number of instances retained by the pool (defaults to 1024). Once poolSize is exceeded, new context instances are not cached and EF falls back to the non-pooling behavior of creating instances on demand.

Benchmarks

Following are the benchmark results for fetching a single row from a SQL Server database running locally on the same machine, with and without context pooling. As always, results will change with the number of rows, the latency to your database server and other factors. Importantly, this benchmarks single-threaded pooling performance, while a real-world contended scenario may have different results; benchmark on your platform before making any decisions. The source code is available here, feel free to use it as a basis for your own measurements.

Method NumBlogs Mean Error StdDev Gen 0 Gen 1 Gen 2 Allocated
WithoutContextPooling 1 701.6 us 26.62 us 78.48 us 11.7188 - - 50.38 KB
WithContextPooling 1 350.1 us 6.80 us 14.64 us 0.9766 - - 4.63 KB

Managing state in pooled contexts

Context pooling works by reusing the same context instance across requests; this means that it's effectively registered as a Singleton, and the same instance is reused across multiple requests (or DI scopes). This means that special care must be taken when the context involves any state that may change between requests. Crucially, the context's OnConfiguring is only invoked once - when the instance context is first created - and so cannot be used to set state which needs to vary (e.g. a tenant ID).

A typical scenario involving context state would be a multi-tenant ASP.NET Core application, where the context instance has a tenant ID which is taken into account by queries (see Global Query Filters for more details). Since the tenant ID needs to change with each web request, we need to go through some extra steps to make it all work with context pooling.

Let's assume that your application registers a scoped ITenant service, which wraps the tenant ID and any other tenant-related information:

// Below is a minimal tenant resolution strategy, which registers a scoped ITenant service in DI.
// In this sample, we simply accept the tenant ID as a request query, which means that a client can impersonate any
// tenant. In a real application, the tenant ID would be set based on secure authentication data.
builder.Services.AddHttpContextAccessor();
builder.Services.AddScoped<ITenant>(sp =>
{
    var tenantIdString = sp.GetRequiredService<IHttpContextAccessor>().HttpContext.Request.Query["TenantId"];

    return tenantIdString != StringValues.Empty && int.TryParse(tenantIdString, out var tenantId)
        ? new Tenant(tenantId)
        : null;
});

As written above, pay special attention to where you get the tenant ID from - this is an important aspect of your application's security.

Once we have our scoped ITenant service, register a pooling context factory as a Singleton service, as usual:

builder.Services.AddPooledDbContextFactory<WeatherForecastContext>(
    o => o.UseSqlServer(builder.Configuration.GetConnectionString("WeatherForecastContext")));

Next, write a custom context factory which gets a pooled context from the Singleton factory we registered, and injects the tenant ID into context instances it hands out:

public class WeatherForecastScopedFactory : IDbContextFactory<WeatherForecastContext>
{
    private const int DefaultTenantId = -1;

    private readonly IDbContextFactory<WeatherForecastContext> _pooledFactory;
    private readonly int _tenantId;

    public WeatherForecastScopedFactory(
        IDbContextFactory<WeatherForecastContext> pooledFactory,
        ITenant tenant)
    {
        _pooledFactory = pooledFactory;
        _tenantId = tenant?.TenantId ?? DefaultTenantId;
    }

    public WeatherForecastContext CreateDbContext()
    {
        var context = _pooledFactory.CreateDbContext();
        context.TenantId = _tenantId;
        return context;
    }
}

Once we have our custom context factory, register it as a Scoped service:

builder.Services.AddScoped<WeatherForecastScopedFactory>();

Finally, arrange for a context to get injected from our Scoped factory:

builder.Services.AddScoped(
    sp => sp.GetRequiredService<WeatherForecastScopedFactory>().CreateDbContext());

As this point, your controllers automatically get injected with a context instance that has the right tenant ID, without having to know anything about it.

The full source code for this sample is available here.

Note

Although EF Core takes care of resetting internal state for DbContext and its related services, it generally does not reset state in the underlying database driver, which is outside of EF. For example, if you manually open and use a DbConnection or otherwise manipulate ADO.NET state, it's up to you to restore that state before returning the context instance to the pool, e.g. by closing the connection. Failure to do so may cause state to get leaked across unrelated requests.

Compiled queries

When EF receives a LINQ query tree for execution, it must first "compile" that tree, e.g. produce SQL from it. Because this task is a heavy process, EF caches queries by the query tree shape, so that queries with the same structure reuse internally-cached compilation outputs. This caching ensures that executing the same LINQ query multiple times is very fast, even if parameter values differ.

However, EF must still perform certain tasks before it can make use of the internal query cache. For example, your query's expression tree must be recursively compared with the expression trees of cached queries, to find the correct cached query. The overhead for this initial processing is negligible in the majority of EF applications, especially when compared to other costs associated with query execution (network I/O, actual query processing and disk I/O at the database...). However, in certain high-performance scenarios it may be desirable to eliminate it.

EF supports compiled queries, which allow the explicit compilation of a LINQ query into a .NET delegate. Once this delegate is acquired, it can be invoked directly to execute the query, without providing the LINQ expression tree. This technique bypasses the cache lookup, and provides the most optimized way to execute a query in EF Core. Following are some benchmark results comparing compiled and non-compiled query performance; benchmark on your platform before making any decisions. The source code is available here, feel free to use it as a basis for your own measurements.

Method NumBlogs Mean Error StdDev Gen 0 Allocated
WithCompiledQuery 1 564.2 us 6.75 us 5.99 us 1.9531 9 KB
WithoutCompiledQuery 1 671.6 us 12.72 us 16.54 us 2.9297 13 KB
WithCompiledQuery 10 645.3 us 10.00 us 9.35 us 2.9297 13 KB
WithoutCompiledQuery 10 709.8 us 25.20 us 73.10 us 3.9063 18 KB

To use compiled queries, first compile a query with EF.CompileAsyncQuery as follows (use EF.CompileQuery for synchronous queries):

private static readonly Func<BloggingContext, int, IAsyncEnumerable<Blog>> _compiledQuery
    = EF.CompileAsyncQuery(
        (BloggingContext context, int length) => context.Blogs.Where(b => b.Url.StartsWith("http://") && b.Url.Length == length));

In this code sample, we provide EF with a lambda accepting a DbContext instance, and an arbitrary parameter to be passed to the query. You can now invoke that delegate whenever you wish to execute the query:

await foreach (var blog in _compiledQuery(context, 8))
{
    // Do something with the results
}

Note that the delegate is thread-safe, and can be invoked concurrently on different context instances.

Limitations

  • Compiled queries may only be used against a single EF Core model. Different context instances of the same type can sometimes be configured to use different models; running compiled queries in this scenario is not supported.
  • When using parameters in compiled queries, use simple, scalar parameters. More complex parameter expressions - such as member/method accesses on instances - are not supported.

Query caching and parameterization

When EF receives a LINQ query tree for execution, it must first "compile" that tree, e.g. produce SQL from it. Because this task is a heavy process, EF caches queries by the query tree shape, so that queries with the same structure reuse internally-cached compilation outputs. This caching ensures that executing the same LINQ query multiple times is very fast, even if parameter values differ.

Consider the following two queries:

var post1 = context.Posts.FirstOrDefault(p => p.Title == "post1");
var post2 = context.Posts.FirstOrDefault(p => p.Title == "post2");

Since the expression trees contains different constants, the expression tree differs and each of these queries will be compiled separately by EF Core. In addition, each query produces a slightly different SQL command:

SELECT TOP(1) [b].[Id], [b].[Name]
FROM [Posts] AS [b]
WHERE [b].[Name] = N'post1'

SELECT TOP(1) [b].[Id], [b].[Name]
FROM [Posts] AS [b]
WHERE [b].[Name] = N'post2'

Because the SQL differs, your database server will likely also need to produce a query plan for both queries, rather than reusing the same plan.

A small modification to your queries can change things considerably:

var postTitle = "post1";
var post1 = context.Posts.FirstOrDefault(p => p.Title == postTitle);
postTitle = "post2";
var post2 = context.Posts.FirstOrDefault(p => p.Title == postTitle);

Since the blog name is now parameterized, both queries have the same tree shape, and EF only needs to be compiled once. The SQL produced is also parameterized, allowing the database to reuse the same query plan:

SELECT TOP(1) [b].[Id], [b].[Name]
FROM [Posts] AS [b]
WHERE [b].[Name] = @__postTitle_0

Note that there is no need to parameterize each and every query: it's perfectly fine to have some queries with constants, and indeed, databases (and EF) can sometimes perform certain optimization around constants which aren't possible when the query is parameterized. See the section on dynamically-constructed queries for an example where proper parameterization is crucial.

Note

EF Core's metrics report the Query Cache Hit Rate. In a normal application, this metric reaches 100% soon after program startup, once most queries have executed at least once. If this metric remains stable below 100%, that is an indication that your application may be doing something which defeats the query cache - it's a good idea to investigate that.

Note

How the database manages caches query plans is database-dependent. For example, SQL Server implicitly maintains an LRU query plan cache, whereas PostgreSQL does not (but prepared statements can produce a very similar end effect). Consult your database documentation for more details.

Dynamically-constructed queries

In some situations, it is necessary to dynamically construct LINQ queries rather than specifying them outright in source code. This can happen, for example, in a website which receives arbitrary query details from a client, with open-ended query operators (sorting, filtering, paging...). In principle, if done correctly, dynamically-constructed queries can be just as efficient as regular ones (although it's not possible to use the compiled query optimization with dynamic queries). In practice, however, they are frequently the source of performance issues, since it's easy to accidentally produce expression trees with shapes that differ every time.

The following example uses three techniques to construct a query's Where lambda expression:

  1. Expression API with constant: Dynamically build the expression with the Expression API, using a constant node. This is a frequent mistake when dynamically building expression trees, and causes EF to recompile the query each time it's invoked with a different constant value (it also usually causes plan cache pollution at the database server).
  2. Expression API with parameter: A better version, which substitutes the constant with a parameter. This ensures that the query is only compiled once regardless of the value provided, and the same (parameterized) SQL is generated.
  3. Simple with parameter: A version which doesn't use the Expression API, for comparison, which creates the same tree as the method above but is much simpler. In many cases, it's possible to dynamically build your expression tree without resorting to the Expression API, which is easy to get wrong.

We add a Where operator to the query only if the given parameter is not null. Note that this isn't a good use case for dynamically constructing a query - but we're using it for simplicity:

[Benchmark]
public int ExpressionApiWithConstant()
{
    var url = "blog" + Interlocked.Increment(ref _blogNumber);
    using var context = new BloggingContext();

    IQueryable<Blog> query = context.Blogs;

    if (_addWhereClause)
    {
        var blogParam = Expression.Parameter(typeof(Blog), "b");
        var whereLambda = Expression.Lambda<Func<Blog, bool>>(
            Expression.Equal(
                Expression.MakeMemberAccess(
                    blogParam,
                    typeof(Blog).GetMember(nameof(Blog.Url)).Single()),
                Expression.Constant(url)),
            blogParam);

        query = query.Where(whereLambda);
    }

    return query.Count();
}

Benchmarking these two techniques gives the following results:

Method Mean Error StdDev Gen0 Gen1 Allocated
ExpressionApiWithConstant 1,665.8 us 56.99 us 163.5 us 15.6250 - 109.92 KB
ExpressionApiWithParameter 757.1 us 35.14 us 103.6 us 12.6953 0.9766 54.95 KB
SimpleWithParameter 760.3 us 37.99 us 112.0 us 12.6953 - 55.03 KB

Even if the sub-millisecond difference seems small, keep in mind that the constant version continuously pollutes the cache and causes other queries to be re-compiled, slowing them down as well and having a general negative impact on your overall performance. It's highly recommended to avoid constant query recompilation.

Note

Avoid constructing queries with the expression tree API unless you really need to. Aside from the API's complexity, it's very easy to inadvertently cause significant performance issues when using them.

Compiled models

Compiled models can improve EF Core startup time for applications with large models. A large model typically means hundreds to thousands of entity types and relationships. Startup time here is the time to perform the first operation on a DbContext when that DbContext type is used for the first time in the application. Note that just creating a DbContext instance does not cause the EF model to be initialized. Instead, typical first operations that cause the model to be initialized include calling DbContext.Add or executing the first query.

Compiled models are created using the dotnet ef command-line tool. Ensure that you have installed the latest version of the tool before continuing.

A new dbcontext optimize command is used to generate the compiled model. For example:

dotnet ef dbcontext optimize

The --output-dir and --namespace options can be used to specify the directory and namespace into which the compiled model will be generated. For example:

PS C:\dotnet\efdocs\samples\core\Miscellaneous\CompiledModels> dotnet ef dbcontext optimize --output-dir MyCompiledModels --namespace MyCompiledModels
Build started...
Build succeeded.
Successfully generated a compiled model, to use it call 'options.UseModel(MyCompiledModels.BlogsContextModel.Instance)'. Run this command again when the model is modified.
PS C:\dotnet\efdocs\samples\core\Miscellaneous\CompiledModels>

The output from running this command includes a piece of code to copy-and-paste into your DbContext configuration to cause EF Core to use the compiled model. For example:

protected override void OnConfiguring(DbContextOptionsBuilder optionsBuilder)
    => optionsBuilder
        .UseModel(MyCompiledModels.BlogsContextModel.Instance)
        .UseSqlite(@"Data Source=test.db");

Compiled model bootstrapping

It is typically not necessary to look at the generated bootstrapping code. However, sometimes it can be useful to customize the model or its loading. The bootstrapping code looks something like this:

[DbContext(typeof(BlogsContext))]
partial class BlogsContextModel : RuntimeModel
{
    private static BlogsContextModel _instance;
    public static IModel Instance
    {
        get
        {
            if (_instance == null)
            {
                _instance = new BlogsContextModel();
                _instance.Initialize();
                _instance.Customize();
            }

            return _instance;
        }
    }

    partial void Initialize();

    partial void Customize();
}

This is a partial class with partial methods that can be implemented to customize the model as needed.

In addition, multiple compiled models can be generated for DbContext types that may use different models depending on some runtime configuration. These should be placed into different folders and namespaces, as shown above. Runtime information, such as the connection string, can then be examined and the correct model returned as needed. For example:

public static class RuntimeModelCache
{
    private static readonly ConcurrentDictionary<string, IModel> _runtimeModels
        = new();

    public static IModel GetOrCreateModel(string connectionString)
        => _runtimeModels.GetOrAdd(
            connectionString, cs =>
            {
                if (cs.Contains("X"))
                {
                    return BlogsContextModel1.Instance;
                }

                if (cs.Contains("Y"))
                {
                    return BlogsContextModel2.Instance;
                }

                throw new InvalidOperationException("No appropriate compiled model found.");
            });
}

Limitations

Compiled models have some limitations:

Because of these limitations, you should only use compiled models if your EF Core startup time is too slow. Compiling small models is typically not worth it.

If supporting any of these features is critical to your success, then please vote for the appropriate issues linked above.

Reducing runtime overhead

As with any layer, EF Core adds a bit of runtime overhead compared to coding directly against lower-level database APIs. This runtime overhead is unlikely to impact most real-world applications in a significant way; the other topics in this performance guide, such as query efficiency, index usage and minimizing roundtrips, are far more important. In addition, even for highly-optimized applications, network latency and database I/O will usually dominate any time spent inside EF Core itself. However, for high-performance, low-latency applications where every bit of perf is important, the following recommendations can be used to reduce EF Core overhead to a minimum:

  • Turn on DbContext pooling; our benchmarks show that this feature can have a decisive impact on high-perf, low-latency applications.
    • Make sure that the maxPoolSize corresponds to your usage scenario; if it is too low, DbContext instances will be constantly created and disposed, degrading performance. Setting it too high may needlessly consume memory as unused DbContext instances are maintained in the pool.
    • For an extra tiny perf boost, consider using PooledDbContextFactory instead of having DI inject context instances directly. DI management of DbContext pooling incurs a slight overhead.
  • Use precompiled queries for hot queries.
    • The more complex the LINQ query - the more operators it contains and the bigger the resulting expression tree - the more gains can be expected from using compiled queries.
  • Consider disabling thread safety checks by setting EnableThreadSafetyChecks to false in your context configuration.
    • Using the same DbContext instance concurrently from different threads isn't supported. EF Core has a safety feature which detects this programming bug in many cases (but not all), and immediately throws an informative exception. However, this safety feature adds some runtime overhead.
    • WARNING: Only disable thread safety checks after thoroughly testing that your application doesn't contain such concurrency bugs.