Take the 2-minute tour ×
Programmers Stack Exchange is a question and answer site for professional programmers interested in conceptual questions about software development. It's 100% free, no registration required.

Is Entity Framework 4 a good solution for a public website with potentially 1000 hits/second?

In my understanding EF is a viable solution for mostly smaller or intranet websites, but wouldn't scale easily for something like a popular community website (I know SO is using LINQ to SQL, but.. I'd like more examples/proof...)

Now I am standing at the crossroads of either choosing a pure ADO.NET approach or EF4. Do you think the improved developer productivity with EF is worth the lost performance and granular access of ADO.NET (with stored procedures)? Any serious issues that a high traffic website might face, was it using EF?

Thank you in advance.

share|improve this question
add comment

migrated from stackoverflow.com Nov 1 '11 at 22:29

This question came from our site for professional and enthusiast programmers.

3 Answers

up vote 63 down vote accepted

It depends a bit on how much abstraction you need. Everything is a compromise; for example, EF and NHibernate introduce great flexibility for representing the data in interesting and exotic models - but as a result they do add overhead. Noticeable overhead.

If you don't need to be able to switch between database providers, and different per-client table layouts, and if your data is primarily read, and if you don't need to be able to use the same model in EF, SSRS, ADO.NET Data Services, etc - then if you want absolute performance as your key measure you could do far worse than look at dapper. In our tests based on both LINQ-to-SQL and EF, we find that EF is significantly slower in terms of raw read performance, presumably due to the abstraction layers (between storage model etc) and materialization.

Here at SO, we are obsessive-compulsive about raw performance, and we're happy to take the development hit of losing some abstraction in order to gain speed. As such, our primary tool for querying the database is dapper. This even allows us to use our pre-existing LINQ-to-SQL model, but simply: it is heaps faster. In performance tests, it is essentially exactly the same performance as writing all the ADO.NET code (parameters, data-readers etc) manually, but without the risk of getting a column name wrong. It is, however, SQL based (although it is happy to use SPROCs if that is your chosen poison). The advantage of this is that there is no additional processing involved, but it is a system for people who like SQL. Which I consider: not a bad thing!

A typical query, for example, might be:

int customerId = ...
var orders = connection.Query<Order>(
    "select * from Orders where CustomerId = @customerId ",
    new { customerId }).ToList();

which is convenient, injection-safe, etc - but without tons of data-reader goo. Note that while it can handle both horizontal and vertical partitions to load complex structures, it will not support lazy-loading (but: we're big fans of very explicit loading - fewer surprises).

Note in this answer I am not saying that EF isn't suitable for high-volume work; simply: I know that dapper is up to it.

share|improve this answer
7  
+1 for dapper. Using a complex ORM for read models is just unnecessary. The approach we take now is to use an ORM for our domain model (where the fancy ORM stuff is actually useful) and dapper for our read model. This makes for super fast applications. –  Ben Oct 31 '11 at 20:09
1  
@Marc, thank you for the great answer - I can finally make my decision with confidence! Will definitely look into dapper in more detail later. Really like how it is just one file :) –  niaher Nov 1 '11 at 10:43
2  
I wrote up my own ORM. Its slow. I looked at dapper and liked it. Now i use dapper for all my reads and my own ORM for inserts (which supports FK, transactions and all the good stuff). Its the easiest most readable code i have ever written. –  acidzombie24 Nov 1 '11 at 22:39
1  
@acidzombie24 dapper supports transactions, and the contrib part of dapper (not part of the nuget deploy) is gaining insert etc options. Just mentioning for completeness. I'm glad dapper was handy. –  Marc Gravell Nov 1 '11 at 22:43
    
Really? great then. I am definitely keeping my eyes on that. I'm not really asking this for a feature since my own does it but will there be support for inserts where if a class has a object that belongs to another class it will try to insert that obj/table as well?. class A { long id; B b; int[] data; } class B { long id; ... }; dapper.insert(new A(){b=new B(){...}, data=data}); //insert obj b and obj a in table b and table a respectively. If so that would be great but i do know its a pain to implement as it took 2/3 of my time(with unique keys and checks to see if B has an existing PK etc) –  acidzombie24 Nov 1 '11 at 23:00
show 2 more comments

The question "which ORM should I use" is really targeting the tip of a huge iceberg when it comes to the overall data access strategy and performance optimization in a large scale application.

All of the following things (roughly in order of importance) are going to affect throughput, and all of them are handled (sometimes in different ways) by most of the major ORM frameworks out there:

  1. Database Design and Maintenance

    This is, by a wide margin, the single most important determinant of the throughput of a data-driven application or web site, and often totally ignored by programmers.

    If you don't use proper normalization techniques, your site is doomed. If you don't have primary keys, almost every query will be dog-slow. If you use well-known anti-patterns such as using tables for Key-Value Pairs (AKA Entity-Attribute-Value) for no good reason, you'll explode the number of physical reads and writes.

    If you don't take advantage of the features the database gives you, such as page compression, FILESTREAM storage (for binary data), SPARSE columns, hierarchyid for hierarchies, and so on (all SQL Server examples), then you will not see anywhere near the performance that you could be seeing.

    You should start worrying about your data access strategy after you've designed your database and convinced yourself that it's as good as it possibly can be, at least for the time being.

  2. Eager vs. Lazy Loading

    Most ORMs used a technique called lazy loading for relationships, which means that by default it will load one entity (table row) at a time, and make a round-trip to the database every time it needs to load one or many related (foreign key) rows.

    This isn't a good or bad thing, it rather depends on what's actually going to be done with the data, and how much you know up-front. Sometimes lazy-loading is absolutely the right thing to do. NHibernate, for example, may decide not to query for anything at all and simply generate a proxy for a particular ID. If all you ever need is the ID itself, why should it ask for more? On the other hand, if you are trying to print a tree of every single element in a 3-level hierarchy, lazy-loading becomes an O(N²) operation, which is extremely bad for performance.

    One interesting benefit to using "pure SQL" (i.e. raw ADO.NET queries/stored procedures) is that it basically forces you to think about exactly what data is necessary to display any given screen or page. ORMs and lazy-loading features don't prevent you from doing this, but they do give you the opportunity to be... well, lazy, and accidentally explode the number of queries you execute. So you need to understand your ORMs eager-loading features and be ever vigilant about the number of queries you're sending to the server for any given page request.

  3. Caching

    All major ORMs maintain a first-level cache, AKA "identity cache", which means that if you request the same entity twice by its ID, it doesn't require a second round-trip, and also (if you designed your database correctly) gives you the ability to use optimistic concurrency.

    The L1 cache is pretty opaque in L2S and EF, you kind of have to trust that it's working. NHibernate is more explicit about it (Get/Load vs. Query/QueryOver). Still, as long as you try to query by ID as much as possible, you should be fine here. A lot of people forget about the L1 cache and repeatedly look up the same entity over and over again by something other than its ID (i.e. a lookup field). If you need to do this then you should save the ID or even the entire entity for future lookups.

    There's also a level 2 cache ("query cache"). NHibernate has this built-in. Linq to SQL and Entity Framework have compiled queries, which can help reduce app server loads quite a bit by compiling the query expression itself, but it doesn't cache the data. Microsoft seems to consider this an application concern rather than a data-access concern, and this is a major weak point of both L2S and EF. Needless to say it's also a weak point of "raw" SQL. In order to get really good performance with basically any ORM other than NHibernate, you need to implement your own caching façade.

    There's also an L2 cache "extension" for EF4 which is okay, but not really a wholesale replacement for an application-level cache.

  4. Number of Queries

    Relational databases are based on sets of data. They're really good at producing large amounts of data in a short amount of time, but they're nowhere near as good in terms of query latency because there's a certain amount of overhead involved in every command. A well-designed app should play to the strengths of this DBMS and try to minimize the number of queries and maximize the amount of data in each.

    Now I'm not saying to query the entire database when you only need one row. What I'm saying is, if you need the Customer, Address, Phone, CreditCard, and Order rows all at the same time in order to serve a single page, then you should ask for them all at the same time, don't execute each query separately. Sometimes it's worse than that, you'll see code that queries the same Customer record 5 times in a row, first to get the Id, then the Name, then the EmailAddress, then... it's ridiculously inefficient.

    Even if you need to execute several queries that all operate on completely different sets of data, it's usually still more efficient to send it all to the database as a single "script" and have it return multiple result sets. It's the overhead you're concerned with, not the total amount of data.

    This might sound like common sense but it's often really easy to lose track of all the queries that are being executed in various parts of the application; your Membership Provider queries the user/role tables, your Header action queries the shopping cart, your Menu action queries the site map table, your Sidebar action queries the featured product list, and then maybe your page is divided into a few separate autonomous areas which query the Order History, Recently Viewed, Category, and Inventory tables separately, and before you know it, you're executing 20 queries before you can even start to serve the page. It just utterly destroys performance.

    Some frameworks - and I'm thinking mainly of NHibernate here - are incredibly clever about this and allow you to use something called futures which batch up entire queries and try to execute them all at once, at the last possible minute. AFAIK, you're on your own if you want to do this with any of the Microsoft technologies; you have to build it into your application logic.

  5. Indexing, Predicates, and Projections

    At least 50% of devs I speak to and even some DBAs seem to have trouble with the concept of covering indexes. They think, "well, the Customer.Name column is indexed, so every lookup I do on the name should be fast." Except it doesn't work that way unless the Name index covers the specific column you're looking up. In SQL Server, that's done with INCLUDE in the CREATE INDEX statement.

    If you naïvely use SELECT * everywhere - and that is more or less what every ORM will do unless you explicitly specify otherwise using a projection - then the DBMS may very well choose to completely ignore your indexes because they contain non-covered columns. A projection means that, for example, instead of doing this:

    from c in db.Customers where c.Name == "John Doe" select c
    

    You do this instead:

    from c in db.Customers where c.Name == "John Doe"
    select new { c.Id, c.Name }
    

    And this will, for most modern ORMs, instruct it to only go and query the Id and Name columns which are presumably covered by the index (but not the Email, LastActivityDate, or whatever other columns you happened to stick in there).

    It's also very easy to completely blow away any indexing benefits by using inappropriate predicates. For example:

    from c in db.Customers where c.Name.Contains("Doe")
    

    ...looks almost identical to our previous query but in fact will result in a full table or index scan because it translates to LIKE '%Doe%'. Similarly, another query which looks suspiciously simple is:

    from c in db.Customers where (maxDate == null) || (c.BirthDate >= maxDate)
    

    Assuming you have an index on BirthDate, this predicate has a good chance to render it completely useless. Our hypothetical programmer here has obviously attempted to create a kind of dynamic query ("only filter the birth date if that parameter was specified"), but this isn't the right way to do it. Written like this instead:

    from c in db.Customers where c.BirthDate >= (maxDate ?? DateTime.MinValue)
    

    ...now the DB engine knows how to parameterize this and do an index seek. One minor, seemingly insignificant change to the query expression can drastically affect performance.

    Unfortunately LINQ in general makes it all too easy to write bad queries like this because sometimes the providers are able to guess what you were trying to do and optimize the query, and sometimes they aren't. So you end up with frustratingly inconsistent results which would have been blindingly obvious (to an experienced DBA, anyway) had you just written plain old SQL.

    Basically it all comes down to the fact that you really have to keep a close eye on both the generated SQL and the execution plans they lead to, and if you're not getting the results you expect, don't be afraid to bypass the ORM layer once in a while and hand-code the SQL. This goes for any ORM, not just EF.

  6. Transactions and Locking

    Do you need to display data that's current up to the millisecond? Maybe - it depends - but probably not. Sadly, Entity Framework doesn't give you nolock, you can only use READ UNCOMMITTED at the transaction level (not table level). In fact none of the ORMs are particularly reliable about this; if you want to do dirty reads, you have to drop down to the SQL level and write ad-hoc queries or stored procedures. So what it boils down to, again, is how easy it is for you to do that within the framework.

    Entity Framework has come a long way in this regard - version 1 of EF (in .NET 3.5) was god-awful, made it incredibly difficult to break through the "entities" abstraction, but now you have ExecuteStoreQuery and Translate, so it's really not too bad. Make friends with these guys because you'll be using them a lot.

    There's also the issue of write locking and deadlocks and the general practice of holding locks in the database for as little time as possible. In this regard, most ORMs (including Entity Framework) actually tend to be better than raw SQL because they encapsulate the unit of Work pattern, which in EF is SaveChanges. In other words, you can "insert" or "update" or "delete" entities to your heart's content, whenever you want, secure in the knowledge that no changes will actually get pushed to the database until you commit the unit of work.

    Note that a UOW is not analogous to a long-running transaction. The UOW still uses the optimistic concurrency features of the ORM and tracks all changes in memory. Not a single DML statement is emitted until the final commit. This keeps transaction times as low as possible. If you build your application using raw SQL, it's quite difficult to achieve this deferred behaviour.

    What this means for EF specifically: Make your units of work as coarse as possible and don't commit them until you absolutely need to. Do this and you'll end up with much lower lock contention than you would using individual ADO.NET commands at random times.

In Conclusion:

EF is completely fine for high-traffic/high-performance applications, just like every other framework is fine for high-traffic/high-performance applications. What matters is how you use it. Here's a quick comparison of the most popular frameworks and what features they offer in terms of performance (legend: N = Not supported, P = Partial, Y = yes/supported):

                                | L2S | EF1 | EF4 | NH3 | ADO
                                +-----+-----+-----+-----+-----
Lazy Loading (entities)         |  N  |  N  |  N  |  Y  |  N
Lazy Loading (relationships)    |  Y  |  Y  |  Y  |  Y  |  N
Eager Loading (global)          |  N  |  N  |  N  |  Y  |  N
Eager Loading (per-session)     |  Y  |  N  |  N  |  Y  |  N
Eager Loading (per-query)       |  N  |  Y  |  Y  |  Y  |  Y
Level 1 (Identity) Cache        |  Y  |  Y  |  Y  |  Y  |  N
Level 2 (Query) Cache           |  N  |  N  |  P  |  Y  |  N
Compiled Queries                |  Y  |  P  |  Y  |  N  | N/A
Multi-Queries                   |  N  |  N  |  N  |  Y  |  Y
Multiple Result Sets            |  Y  |  N  |  P  |  Y  |  Y
Futures                         |  N  |  N  |  N  |  Y  |  N
Explicit Locking (per-table)    |  N  |  N  |  N  |  P  |  Y
Transaction Isolation Level     |  Y  |  Y  |  Y  |  Y  |  Y
Ad-Hoc Queries                  |  Y  |  P  |  Y  |  Y  |  Y
Stored Procedures               |  Y  |  P  |  Y  |  Y  |  Y
Unit of Work                    |  Y  |  Y  |  Y  |  Y  |  N

As you can see, EF4 (the current version) doesn't fare too badly, but it's probably not the best if performance is your primary concern. NHibernate is much more mature in this area and even Linq to SQL provides some performance-enhancing features that EF still doesn't. Raw ADO.NET is often going to be faster for very specific data-access scenarios, but, when you put all the pieces together, it really doesn't offer a lot of important benefits that you get from the various frameworks.

And, just to make completely sure that I sound like a broken record, none of this matters in the slightest if you don't design your database, application, and data access strategies properly. All of the items in the chart above are for improving performance beyond the baseline; most of the time, the baseline itself is what needs the most improvement.

share|improve this answer
14  
What an awesome and comprehensive answer! –  Slauma Oct 30 '11 at 18:10
    
+1 (more if I could) - one of the best answers I've seen in a while here and I learned a thing or two - thanks for sharing this! –  BrokenGlass Oct 30 '11 at 19:36
    
This is great answer even I don't agree with everything mentioned. The table comparing ORMs is not always correct. What is entities lazy loading? Do you mean lazy loaded columns? That is supported in L2S. Why do you think NH doesn't support compiled queries? I think named HQL queries can be pre-compiled. EF4 has no support for multiple result sets. –  Ladislav Mrnka Oct 30 '11 at 20:50
5  
I have to strongly disagree with the unqualified "EF is completely fine for high-traffic/high-performance applications" statement, we've seen repeatedly that this is not the case. Granted, maybe we disagree on what "high performance" means, but for example optimizing web pages down to 500ms and having 400ms+ of that spent inexplicably inside the framework (and only 10ms actually hitting SQL) isn't "fine" for some situations, it's downright unacceptable for our dev team. –  Nick Craver Oct 31 '11 at 10:34
1  
Simple note about futures in EF. They are not officially provided by the MS EF team but can be achieved through third party projects that define Future<> extensions to IQueryable<>. For example EntityFramework.Extended by LoreSoft, available in NuGet. My personal tests in production applications show performance gain up to 10x when packing dozens of non-dependant queries (all queries can be executed in parallel, no one requires the result of a previous one) in a single batch using Future. Also AsNoTracking() improves performance a lot when just reading lots of records, no later update. –  David Oliván Ubieto Mar 26 at 13:00
show 6 more comments

Edit: Based on @Aaronaught great answer I'm adding few points targeting performance with EF. Those new points are prefixed by Edit.


The biggest improvement in performance in high traffic websites is achieved by caching (= first of all avoiding any web server processing or database querying) followed by asynchronous processing to avoid thread blocking while database queries are performed.

There is no bullet proof answer to your question because it always depends on requirements for application and on complexity of queries. The truth is that developer productivity with EF hides complexity behind which in many cases leads to incorrect usage of EF and terrible performance. The idea that you can expose high level abstracted interface for data access and it will smoothly works in all cases doesn't work. Even with ORM you must know what is happening behind the abstraction and how to correctly use it.

If you don't have previous experience with EF you will meet a lot of challenges when dealing with performance. You can make much more mistakes when working with EF comparing to ADO.NET. Also there is a lot of additional processing done in EF, so EF will always be significantly slower than native ADO.NET - that is something you can measure by simple proof of concept application.

If you want to get best performance from EF you will most probably have to:

  • Very carefully revise your data access with SQL profiler and review your LINQ queries if they correctly use Linq-to-entities instead of Linq-to-objects
  • Very carefully use advanced EF optimization features like MergeOption.NoTracking
  • Use ESQL in some cases
  • Pre-compile queries which are executed often
  • Think about taking advantage of EF Caching wrapper to get "second level cache" like feature for some queries
  • Use SQL views or custom mapped SQL queries (requires manual maintaining of EDMX file) in some scenarios for often used projections or aggregations which needs performance improvements
  • Use native SQL and stored procedures for some queries which don't provide sufficient performance when defined in Linq or ESQL
  • Edit: Carefully use queries - every query makes separate roundtrip to the database. EFv4 has no query batching because it is not able to use multiple result sets per executed database command. EFv4.5 will support multiple result sets for mapped stored procedures.
  • Edit: Carefully work with data modifications. Again EF completely lacks command batching. So in ADO.NET you can use single SqlCommand containing multiple insert, updates or deletes but with EF every such command will be executed in separate roundtrip to database.
  • Edit: Carefully work with identity map / identity cache. EF has special method (GetByKey in ObjectContext API or Find in DbContext API) to query the cache first. If you use Linq-to-entities or ESQL it will create roundtrip to the database and after that it will return existing instance from the cache.
  • Edit: Carefully use eager loading. It is not always win-win solution because it produces one huge dataset. As you can see it is a lot of additional complexity and that is the whole point. ORM makes mapping and materialization simpler but when dealing with performance it will make it much more complex and you will have to make trade-offs.

I'm not sure if SO is still using L2S. They developed new open source ORM called Dapper and I think the main point behind this development was increasing performance.

share|improve this answer
    
Ladislav, that's a really helpful answer. This is the first time I hear about Dapper (and consequently discovered PetaPoco, Massive) - and it looks like an interesting idea. –  niaher Oct 30 '11 at 17:07
1  
SO seems to use a mix of LINQ to SQL and Dapper now: samsaffron.com/archive/2011/03/30/… Quote: "We are using our new ORM [Dapper] for a specific problem: mapping parameterized SQL to business objects. We are not using it as a full blown ORM. It does not do relationships and other bells and whistles. This allows us to continue using LINQ-2-SQL where performance does not matter and port all our inline SQL to use our mapper, since it is faster and more flexible." –  Slauma Oct 30 '11 at 17:11
5  
@Slauma well, that is a statement from months ago, in general all new work on SO is done in Dapper, for example a new table I added today is not even in the dbml file. –  Sam Saffron Oct 31 '11 at 10:27
1  
@Sam: Is there a new blog post about current data access strategy on SO? Would be very interesting! Has Dapper been extended in the meantime? My understanding was that Dapper isn't a complete ORM, no support of relationships - and what about updates, inserts, deletes, transactions, change tracking, etc. –  Slauma Oct 31 '11 at 11:16
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.