How to avoid Query Plan re-compilation when using IEnumerable.Contains in Entity Framework LINQ queries? - performance

Question

I have the following LINQ query executed using Entity Framework (v6.1.1):
private IList<Customer> GetFullCustomers(IEnumerable<int> customersIds)
{
IQueryable<Customer> fullCustomerQuery = GetFullQuery();
return fullCustomerQuery.Where(c => customersIds.Contains(c.Id)).ToList();
}
This query is translated into fairly nice SQL:
SELECT
[Extent1].[Id] AS [Id],
[Extent1].[FirstName] AS [FirstName]
-- ...
FROM [dbo].[Customer] AS [Extent1]
WHERE [Extent1].[Id] IN (1, 2, 3, 5)
However, I get a very significant performance hit on a query compilation phase. Calling:
ELinqQueryState.GetExecutionPlan(MergeOption? forMergeOption)
Takes ~50% of the time of each request. Digging deeper, it turned out that query gets re-compiled every time I pass different customersIds.
According to MSDN article, this is an expected behavior because IEnumerable that is used in a query is considered volatile and is part of SQL that is cached. That's why SQL is different for every different combination of customersIds and it always has different hash that is used to get compiled query from cache.
Now the question is: How can I avoid this re-compilation while still querying with multiple customersIds?

Answer 1

This is a great question. First of all, here are a couple of workarounds that come to mind (they all require changes to the query):
First workaround
This one maybe a bit obvious and unfortunately not generally applicable: If the selection of items you would need to pass over to Enumerable.Contains already exists in a table in the database, you can write a query that calls Enumerable.Contains on the corresponding entity set in the predicate instead of bringing the items into memory first. An Enumerable.Contains call over data in the database should result in some kind of JOIN-based query that can be cached. E.g. assuming no navigation properties between Customers and SelectedCustomers, you should be able to write the query like this:
var q = db.Customers.Where(c =>
db.SelectedCustomers.Select(s => s.Id).Contains(c.Id));
The syntax of the query with Any is a bit simpler in this case:
var q = db.Customers.Where(c =>
db.SelectedCustomers.Any(s => s.Id == c.Id));
If you don't already have the necessary selection data stored in the database, you will probably don't want the overhead of having to store it, so you should consider the next workaround.
Second workaround
If you know beforehand that you will have a relatively manageable maximum number of elements in the list you can replace Enumerable.Contains with a tree of OR-ed equality comparisons, e.g.:
var list = new [] {1,2,3};
var q = db.Customers.Where(c =>
list[0] == c.Id ||
list[1] == c.Id ||
list[2] == c.Id );
This should produce a parameterized query that can be cached. If the list varies in size from query to query, this should produce a different cache entry for each list size. Alternatively you could use a list with a fixed size and pass some sentinel value that you know will never match the value argument, e.g. 0, -1, or alternatively just repeat one of the other values. In order to produce such predicate expression programmatically at runtime based on a list, you might want to consider using something like PredicateBuilder.
Potential fixes and their challenges
On one hand, changes necessary to support caching of this kind of query using CompiledQuery explicitly would be pretty complex in the current version of EF. The key reason is that the elements in the IEnumerable<T> passed to the Enumerable.Contains method would have to translate into a structural part of the query for the particular translation we produce, e.g.:
var list = new [] {1,2,3};
var q = db.Customers.Where(c => list.Contains(c.Id)).ToList();
The enumerable “list” looks like a simple variable in C#/LINQ but it needs to be translated to a query like this (simplified for clarity):
SELECT * FROM Customers WHERE Id IN(1,2,3)
If list changes to new [] {5,4,3,2,1}, and we would have to generate the SQL query again!
SELECT * FROM Customers WHERE Id IN(5,4,3,2,1)
As a potential solution, we have talked about leaving generated SQL queries open with some kind of special place holder, e.g. store in the query cache that just says
SELECT * FROM Customers WHERE Id IN(<place holder>)
At execution time, we could pick this SQL from the cache and finish the SQL generation with the actual values. Another option would be to leverage a Table-Valued Parameter for the list if the target database can support it. The first option would probably work ok only with constant values, the latter requires a database that supports a special feature. Both are very complex to implement in EF.
Auto compiled queries
On the other hand, for automatic compiled queries (as opposed to explicit CompiledQuery) the issue becomes somewhat artificial: in this case we compute the query cache key after the initial LINQ translation, hence any IEnumerable<T> argument passed should have already been expanded into DbExpression nodes: a tree of OR-ed equality comparisons in EF5, and usually a single DbInExpression node in EF6. Since the query tree already contains a distinct expression for each distinct combination of elements in the source argument of Enumerable.Contains (and therefore for each distinct output SQL query), it is possible to cache the queries.
However even in EF6 these queries are not cached even in the auto compiled queries case. The key reason for that is that we expect the variability of elements in a list to be high (this has to do with the variable size of the list but is also exacerbated by the fact that we normally don't parameterize values that appear as constants to the query, so a list of constants will be translated into constant literals in SQL), so with enough calls to a query with Enumerable.Contains you could produce considerable cache pollution.
We have considered alternative solutions to this as well, but we haven't implemented any yet. So my conclusion is that you would be better off with the second workaround in most cases if as I said, you know the number of elements in the list will remain small and manageable (otherwise you will face performance issues).
Hope this helps!

Answer 2

As of now, this is still a problem in Entity Framework Core when using the SQL Server Database Provider.
💡 Still on Entity Framework 6 (non-core)? skip to the next section.
I wrote QueryableValues to solve this problem in a flexible and performant way; with it you can compose the values from an IEnumerable<T> in your query, like if it were another entity in your DbContext.
In contrast to other solutions out there, QueryableValues achieves this level of performance by:
Resolving with a single round-trip to the database.
Preserving the query's execution plan regardless of the provided values.
Usage example:
// Sample values.
IEnumerable<int> values = Enumerable.Range(1, 10);
// Using a Join.
var myQuery1 =
from e in dbContext.MyEntities
join v in dbContext.AsQueryableValues(values) on e.Id equals v
select new
{
e.Id,
e.Name
};
// Using Contains.
var myQuery2 =
from e in dbContext.MyEntities
where dbContext.AsQueryableValues(values).Contains(e.Id)
select new
{
e.Id,
e.Name
};
You can also compose complex types!
It's available as a nuget package and the project can be found here. It's distributed under the MIT license.
The benchmarks speak for themselves.
An Alternative for Entity Framework 6 (non-core)
🎉 NEW! QueryableValues EF6 Edition has arrived!
I'll explain how to manually provide some of the functionality of QueryableValues on this legacy version of Entity Framework, specifically, the ability to compose an IEnumerable<int> with any of your entities in the same way that QueryableValues does on EF Core. You can use this same technique to support collections of other simple types like long, string, etc.
Requirements
Must use the SQL Server provider
Must use the database-first strategy OR you already have a way to map a TVF using the code-first strategy
Instructions Summary
Create a method that takes an IEnumerable<int> and returns XML.
Create a TVF in your database that takes XML and returns a rowset.
Add the TVF to the EDMX using the designer.
Encapsulate the code that glues the functions created on step 1 and 2 and return an IQueryable<int>.
Use the IQueryable<int> in your queries as desired.
Instructions
1. Create a method that takes a IEnumerable<int> and returns XML
This method will serialize the provided values as XML, so later on it can be transmitted as a parameter in your query.
static string GetXml<T>(IEnumerable<T> values)
{
var sb = new StringBuilder();
using (var stringWriter = new System.IO.StringWriter(sb))
{
var settings = new System.Xml.XmlWriterSettings
{
ConformanceLevel = System.Xml.ConformanceLevel.Fragment
};
using (var xmlWriter = System.Xml.XmlWriter.Create(stringWriter, settings))
{
xmlWriter.WriteStartElement("R");
foreach (var value in values)
{
xmlWriter.WriteStartElement("V");
xmlWriter.WriteValue(value);
xmlWriter.WriteEndElement();
}
xmlWriter.WriteEndElement();
}
}
return sb.ToString();
}
If the above method is provided with new[] { 1, 2, 3 }, it will return a XML string with the following structure:
<R><V>1</V><V>2</V><V>3</V></R>
2. Create a TVF in your database that takes XML and returns a rowset
The following table-valued function (TVF) will take the XML created by the previous function and project it as a rowset with a single column (V), that can then be used from SQL Server's side in your query. Must be created in the database associated with your EDMX file, so it can be added to your EDMX model in the next step.
CREATE FUNCTION dbo.udf_GetIntValuesFromXml
(
#Values XML
)
RETURNS TABLE
WITH SCHEMABINDING
AS
RETURN
(
SELECT I.value('. cast as xs:integer?', 'int') AS V
FROM #Values.nodes('/R/V') N(I)
)
The above function when provided with the <R><V>1</V><V>2</V><V>3</V></R> XML, will return the following rowset:
V
1
2
3
3. Add the TVF to the EDMX using the designer
Table-Valued Functions (TVFs) - EF Docs
After adding this function to your EDMX model, ensure to save the changes to the EDMX file so that your DbContext generated code is up to date.
4. Encapsulate the code that glues the functions created on step 1 and 2 and return an IQueryable<int>
The following code encapsulates the XML serializer function explained above and everything else you need on the .NET side to make this work:
using System.Collections.Generic;
using System.Linq;
public static class QueryableValuesClassicDbContextExtensions
{
private static string GetXml<T>(IEnumerable<T> values)
{
var sb = new StringBuilder();
using (var stringWriter = new System.IO.StringWriter(sb))
{
var settings = new System.Xml.XmlWriterSettings
{
ConformanceLevel = System.Xml.ConformanceLevel.Fragment
};
using (var xmlWriter = System.Xml.XmlWriter.Create(stringWriter, settings))
{
xmlWriter.WriteStartElement("R");
foreach (var value in values)
{
xmlWriter.WriteStartElement("V");
xmlWriter.WriteValue(value);
xmlWriter.WriteEndElement();
}
xmlWriter.WriteEndElement();
}
}
return sb.ToString();
}
public static IQueryable<int> AsQueryableValues(this IQueryableValuesClassicDbContext dbContext, IEnumerable<int> values)
{
return dbContext.GetIntValuesFromXml(GetXml(values));
}
}
public interface IQueryableValuesClassicDbContext
{
IQueryable<int> GetIntValuesFromXml(string xml);
}
The IQueryableValuesClassicDbContext interface is intended to be explicitly implemented on your DbContext class to provide access to the TVF that was added to the EDMX model.
You can do this by creating a partial class for your DbContext. For example, if your DbContext name is TestDbContext:
using System.Linq;
partial class TestDbContext : IQueryableValuesClassicDbContext
{
IQueryable<int> IQueryableValuesClassicDbContext.GetIntValuesFromXml(string xml)
{
return udf_GetIntValuesFromXml(xml).Select(i => i.Value);
}
}
5. Use the IQueryable<int> in your queries as desired (via AsQueryableValues)
using (var db = new TestDbContext())
{
var valuesQuery = db.AsQueryableValues(new[] { 1, 2, 3, 4, 5 });
var resultsUsingContains = db.MyEntity
.Where(i => valuesQuery.Contains(i.MyEntityID))
.Select(i => new { i.MyEntityID, i.PropA })
.ToList();
var resultsUsingJoin = (
from i in db.MyEntity
join v in valuesQuery on i.MyEntityID equals v
select new { i.MyEntityID, i.PropA }
)
.ToList();
}
Below is the T-SQL generated behind the scenes for the above EF queries. As you can see, it's completely parameterized.
exec sp_executesql N'SELECT
[Extent1].[MyEntityID] AS [MyEntityID],
[Extent1].[PropA] AS [PropA]
FROM [dbo].[MyEntity] AS [Extent1]
WHERE EXISTS (SELECT
1 AS [C1]
FROM [dbo].[udf_GetIntValuesFromXml](#Values) AS [Extent2]
WHERE ([Extent2].[V] = [Extent1].[MyEntityID]) AND ([Extent2].[V] IS NOT NULL)
)',N'#Values nvarchar(4000)',#Values=N'<R><V>1</V><V>2</V><V>3</V><V>4</V><V>5</V></R>'
exec sp_executesql N'SELECT
[Extent1].[MyEntityID] AS [MyEntityID],
[Extent1].[PropA] AS [PropA]
FROM [dbo].[MyEntity] AS [Extent1]
INNER JOIN [dbo].[udf_GetIntValuesFromXml](#Values) AS [Extent2] ON [Extent1].[MyEntityID] = [Extent2].[V]',N'#Values nvarchar(4000)',#Values=N'<R><V>1</V><V>2</V><V>3</V><V>4</V><V>5</V></R>'
Limitations
The provided IEnumerable<int> is enumerated at query build time, not at execution time.
The final query cannot reference more than one IQueryable<T> returned by the AsQueryableValues extension method. This is another limitation around composing the same TVF more than once. EF will create two parameters with the same name, which is illegal and you will get the following error:
A parameter named 'Values' already exists in the parameter collection. Parameter names must be unique in the parameter collection.
Incorrect type used for the XML type parameter of the TVF (notice the use of nvarchar instead of xml in the T-SQL above). This is a deficiency in the EF infrastructure (ObjectParameter) that's used to compose the TVF. Not using the correct parameter type has a detrimental effect in performance due to the implicit casting that must be done by SQL Server.
Conclusion
Despite the limitations, this is still a robust solution when compared to not using parameterized T-SQL queries. To understand the underlying issue that this mitigates you can continue reading here.
Legal Stuff
Feel free to use the code and examples above as you wish. I'm releasing it under the MIT license:
MIT License
Copyright (c) Carlos Villegas (yv989c)
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

Answer 3

I had this exact challenge. Here is how I tackled this problem for either strings or longs in an extension method for IQueryables.
To limit the caching pollution we create the same query with a multitude n of m (configurable) parameters, so 1 * m, 2 * m etc. So if the setting is 15; The queryplans would have either 15, 30, 45 etc parameters, depending on the number of elements in the contains (we don't know in advance, but probably less than 100) limiting the number of query plans to 3 if the biggest contains is less than or equal to 45.
The remaining parameters are filled with a placeholdervalue that (we know) doesn't exists in the database. In this case '-1'
Resulting query part;
... WHERE [Filter1].[SomeProperty] IN (#p__linq__0,#p__linq__1, (...) ,#p__linq__19)
... #p__linq__0='SomeSearchText1',#p__linq__1='SomeSearchText2',#p__linq__2='-1',
(...) ,#p__linq__19='-1'
Usage:
ICollection<string> searchtexts = .....ToList();
//or
//ICollection<long> searchIds = .....ToList();
//this is the setting that is relevant for the resulting multitude of possible queryplans
int itemsPerSet = 15;
IQueryable<MyEntity> myEntities = (from c in dbContext.MyEntities
select c)
.WhereContains(d => d.SomeProperty, searchtexts, "-1", itemsPerSet);
The extension method:
using System;
using System.Collections;
using System.Collections.Generic;
using System.Linq;
using System.Linq.Expressions;
namespace MyCompany.Something.Extensions
{
public static class IQueryableExtensions
{
public static IQueryable<T> WhereContains<T, U>(this IQueryable<T> source, Expression<Func<T,U>> propertySelector, ICollection<U> identifiers, U placeholderThatDoesNotExistsAsValue, int cacheLevel)
{
if(!(propertySelector.Body is MemberExpression))
{
throw new ArgumentException("propertySelector must be a MemberExpression", nameof(propertySelector));
}
var propertyExpression = propertySelector.Body as MemberExpression;
var propertyName = propertyExpression.Member.Name;
return WhereContains(source, propertyName, identifiers, placeholderThatDoesNotExistsAsValue, cacheLevel);
}
public static IQueryable<T> WhereContains<T, U>(this IQueryable<T> source, string propertyName, ICollection<U> identifiers, U placeholderThatDoesNotExistsAsValue, int cacheLevel)
{
return source.Where(ContainsPredicateBuilder<T, U>(identifiers, propertyName, placeholderThatDoesNotExistsAsValue, cacheLevel));
}
public static Expression<Func<T, bool>> ContainsPredicateBuilder<T,U>(ICollection<U> ids, string propertyName, U placeholderValue, int cacheLevel = 20)
{
if(cacheLevel < 1)
{
throw new ArgumentException("cacheLevel must be greater than or equal to 1", nameof(cacheLevel));
}
Expression<Func<T, bool>> predicate;
var propertyIsNullable = Nullable.GetUnderlyingType(typeof(T).GetProperty(propertyName).PropertyType) != null;
// fill a list of cachableLevel number of parameters for the property, equal the selected items and padded with the placeholder value to fill the list.
Expression finalExpression = Expression.Constant(false);
var parameter = Expression.Parameter(typeof(T), "x");
/* factor makes sure that this query part contains a multitude of m parameters (i.e. 20, 40, 60, ...),
* so the number of query plans is limited even if lots of users have more than m items selected */
int factor = Math.Max(1, (int)Math.Ceiling((double)ids.Count / cacheLevel));
for (var i = 0; i < factor * cacheLevel; i++)
{
U id = placeholderValue;
if (i < ids.Count)
{
id = ids.ElementAt(i);
}
var temp = new { id };
var constant = Expression.Constant(temp);
var field = Expression.Property(constant, "id");
var member = Expression.Property(parameter, propertyName);
if (propertyIsNullable)
{
member = Expression.Property(member, "Value");
}
var expression = Expression.Equal(member, field);
finalExpression = Expression.OrElse(finalExpression, expression);
}
predicate = Expression.Lambda<Func<T, bool>>(finalExpression, parameter);
return predicate;
}
}
}

Answer 4

This is really a huge problem, and there's no one-size-fits-all answer. However, when most lists are relatively small, diverga's "Second Workaround" works well. I've built a library distributed as a NuGet package to perform this transformation with as little modification to the query as possible:
https://github.com/bchurchill/EFCacheContains
It's been tested out in one project, but feedback and user experiences would be appreciated! If any issues come up please report on github so that I can follow-up.