Using an ORM or plain SQL? [closed]

Speaking as someone who spent quite a bit of time working with JPA (Java Persistence API, basically the standardized ORM API for Java/J2EE/EJB), which includes Hibernate, EclipseLink, Toplink, OpenJPA and others, I'll share some of my observations.

1. ORMs are not fast. They can be adequate and most of the time adequate is OK but in a high-volume low-latency environment they're a no-no;
2. In general purpose programming languages like Java and C# you need an awful lot of magic to make them work (eg load-time weaving in Java, instrumentation, etc);
3. When using an ORM, rather than getting further from SQL (which seems to be the intent), you'll be amazed how much time you spend tweaking XML and/or annotations/attributes to get your ORM to generate performant SQL;
4. For complex queries, there really is no substitute. Like in JPA there are some queries that simply aren't possible that are in raw SQL and when you have to use raw SQL in JPA it's not pretty (C#/.Net at least has dynamic types--var--which is a lot nicer than an Object array);
5. There are an awful lot of "gotchas" when using ORMs. This includes unintended or unexpected behavior, the fact that you have to build in the capability to do SQL updates to your database (by using refresh() in JPA or similar methods because JPA by default caches everything so it won't catch a direct database update--running direct SQL updates is a common production support activity);
6. The object-relational mismatch is always going to cause problems. With any such problem there is a tradeoff between complexity and completeness of the abstraction. At times I felt JPA went too far and hit a real law of diminishing returns where the complexity hit wasn't justified by the abstraction.

There's another problem which takes a bit more explanation.

The traditional model for a Web application is to have a persistence layer and a presentation layer (possibly with a services or other layers in between but these are the important two for this discussion). ORMs force a rigid view from your persistence layer up to the presentation layer (ie your entities).

One of the criticisms of more raw SQL methods is that you end up with all these VOs (value objects) or DTOs (data transfer objects) that are used by simply one query. This is touted as an advantage of ORMs because you get rid of that.

Thing is those problems don't go away with ORMs, they simply move up to the presentation layer. Instead of creating VOs/DTOs for queries, you create custom presentation objects, typically one for every view. How is this better? IMHO it isn't.

I've written about this in ORM or SQL: Are we there yet?.

My persistence technology of choice (in Java) these days is ibatis. It's a pretty thin wrapper around SQL that does 90%+ of what JPA can do (it can even do lazy-loading of relationships although its not well-documented) but with far less overhead (in terms of complexity and actual code).

This came up last year in a GWT application I was writing. Lots of translation from EclipseLink to presentation objects in the service implementation. If we were using ibatis it would've been far simpler to create the appropriate objects with ibatis and then pass them all the way up and down the stack. Some purists might argue this is Bad™. Maybe so (in theory) but I tell you what: it would've led to simpler code, a simpler stack and more productivity.

ORMs have some nice features. They can handle much of the dog-work of copying database columns to object fields. They usually handle converting the language's date and time types to the appropriate database type. They generally handle one-to-many relationships pretty elegantly as well by instantiating nested objects. I've found if you design your database with the strengths and weaknesses of the ORM in mind, it saves a lot of work in getting data in and out of the database. (You'll want to know how it handles polymorphism and many-to-many relationships if you need to map those. It's these two domains that provide most of the 'impedance mismatch' that makes some call ORM the 'vietnam of computer science'.)

For applications that are transactional, i.e. you make a request, get some objects, traverse them to get some data and render it on a Web page, the performance tax is small, and in many cases ORM can be faster because it will cache objects it's seen before, that otherwise would have queried the database multiple times.

For applications that are reporting-heavy, or deal with a large number of database rows per request, the ORM tax is much heavier, and the caching that they do turns into a big, useless memory-hogging burden. In that case, simple SQL mapping (LinQ or iBatis) or hand-coded SQL queries in a thin DAL is the way to go.

I've found for any large-scale application you'll find yourself using both approaches. (ORM for straightforward CRUD and SQL/thin DAL for reporting).

I say plain SQL for Reads, ORM for CUD.

Performance is something I'm always concerned about, specially in web applications, but also code maintainability and readability. To address these issues I wrote SqlBuilder.

ORM is not just portability (which is kinda hard to achieve even with ORMs, for that matter). What it gives you is basically a layer of abstraction over a persistent store, when a ORM tool frees you from writing boilerplate SQL queries (selects by PK or by predicates, inserts, updates and deletes) and lets you concentrate on the problem domain.

Any respectable design will need some abstraction for the database, just to handle the impedance mismatch. But the simplest first step (and adequate for most cases) I would expect would be a DAL, not a heavyweight ORM. Your only options aren't those at the ends of the spectrum.

EDIT in response to a comment requesting me to describe how I distinguish DAL from ORM:

A DAL is what you write yourself, maybe starting from a class that simply encapsulates a table and maps its fields to properties. An ORM is code you don't write for abstraction mechanisms inferred from other properties of your dbms schema, mostly PKs and FKs. (This is where you find out if the automatic abstractions start getting leaky or not. I prefer to inform them intentionally, but that may just be my personal preference).