Law Of Leaky Abstractions (LLA)

• Architecture
• API Design

All non-trivial abstractions, to some degree, are leaky.¹⁾

A solution is bad if

• the leakiness of abstractions is ignored (bad usage of an abstraction) or
• the benefits of the abstraction cannot justify the disadvantages created by its leakiness (bad abstraction) or
• the abstraction is more leaky than necessary (suboptimal abstraction)

Abstractions are typically not perfect. Especially performance aspects are very hard to abstract away from. So there are cases when using the abstraction properly is not possible without knowing the basics underneath the abstraction. So developing and, even more, fixing defects in a system with leaky abstractions makes it necessary to know about all the details the abstraction is supposed to protect the developers from. Often abstractions reduce the effort to develop a feature while they increase the effort for fixing bugs.

A typical problem with leaky abstractions is that the leakiness is ignored. This does not directly mean that the abstraction itself is bad. Often the situation without the abstraction would be worse so it's good to have it. Nevertheless its unavoidable leaks have to be kept in mind. See example 1.

There are also situations where the abstraction is not crafted well. Sometimes there is a way to make the abstraction better (see example 2) but sometimes the whole abstraction is wrong and not having it would be better.

• Bad usages of abstractions are plain wrong. There is no doubt that they should be avoided.
• Suboptimal abstractions can be made better. There are unnecessary leaks and each leak is a possibility for a future usage fault (see ML). As the leaks are unnecessary, there are ways to improve the abstraction so this one s bad.
• The benefits of an abstraction are only reached at the cost of some liabilities so an abstraction is only good when the liabilities are small enough compared to the benefits. Bad abstractions are worse than no abstraction at all, so they should not be employed.

• Make the abstraction less leaky. There might be a way to reduce the leakiness of the abstraction.
• Find a better abstraction. There might be a better abstraction which is less leaky.
• Remove the abstraction. If the abstraction is so leaky that the leakiness creates more problems than the abstraction solves, it is better not to have it.
  • Especially refrain from stacking Frameworks on top each other (e.g. Spring in an EJB Container, etc.)

The law of leaky abstractions by no means says that abstractions are generally a bad thing. Abstractions are extremely helpful and software development without abstractions is virtually impossible. So abstractions are good and developers should strive for creating good abstractions. But abstractions are never perfect and this has to be kept in mind.

See also section contrary principles.

Joel Spolsky: The Law of Leaky Abstractions

The blog article has another focus as it rather explains LLA as an effect rather than an engineering advice. Because abstractions are leaky, developers should and have to know the details abstractions try to protect them from. The principle aspect is only a side aspect there but the main focus here.

• Proposed

• Fallacies of Distributed Computing: The eight fallacies of distributed computing are abstraction leaks which are typically present in distributed systems. Normally it is not possibly to prevent them (so there might not be a better abstraction) but it is important not to ignore these abstraction leaks.

• Keep It Simple Stupid (KISS): Creating good abstractions is sometimes complicated and ignoring leaks is simple.

• Murphy's Law (ML): According to ML, any abstraction leak will eventually result in a mistake.
• Model Principle (MP): ML might help finding the right abstraction, i.e. the one which is the least leaky.
• Rule of Explicitness (RoE): RoE is another way to look at abstractions. Often abstractions create a level of implicitness. Abstraction leaks are one reason why explicit solutions can be considered preferable.
• Easy to Use and Hard to Misuse (EUHM): The more an abstraction leaks, the less it can be considered hard to misuse.

There is plenty of middleware which centers around the notion of distributed objects: RMI, CORBA, DCOM, … These technologies abstract away from the fact that the objects are not local but distributed over the network. They create the illusion that calling all objects are local. But all these technologies are leaky abstractions. There is no way to abstract from the fact that calling a remote object may fail. The network connection may break down, the remote machine may not be available, etc. Furthermore there are completely different performance characteristics of remote calls. There is some unavoidable latency and no abstraction what so ever can change this.

This does not mean that these technologies are generally bad. There is a value in these abstractions but the leaks have to be kept in mind. Ignoring the fact that remote calls may fail will result in fragile systems. If the distributed system to develop should be robust, there has to be code handling failing remote calls. And for performance reasons, remote interfaces have to be crafted in a way that remote calls are minimized. So for example Data Transfer Objects are employed in order to transfer larger chunks of data instead of making a remote call for every access to a getter method.

In C++ string literals such as “test” are of type char* which means pointer to a character. There is some range in the memory where a sequence of characters is stored and there is a pointer pointing to the first character. These strings are unhandy to use so there are string classes in C++ which abstract from these low-level strings which are simply a heritage of the C programming language C++ is built upon. These string classes are much more convenient to use and create the illusion of a built-in string type.

#include <iostream>
 
int main()
{
	std::string s = "Hello";
	s = s + " World";
	std::cout << s << std::endl;
}

In this code it looks like there was a native string type in C++. This is the abstraction. But the abstraction is leaky. The following code does not work:

#include <iostream>
 
int main()
{
	std::string s;
	s = "Hello " + "World";
	std::cout << s << std::endl;
}

So it would be better if C++ had a real native string type.

(this example is taken from ²⁾)

• The virtual address space is leaky. Paging leads to a loss of performance. And page faults show up for example while iterating over a two-dimensional array.
• Platform details are leaky. The Java Virtual Machine does a pretty good job abstracting from the underlying operating system. But at least when it comes to accessing files based on file paths, the underlying file system structure leaks through. On Windows you have “C:\” whereas on Unixoid systems you have “/usr”, “/etc” and stuff.
• SQL is leaky. Some queries have a considerably lower performance than others despite being semantically equivalent.
• Network file systems (SMB, NFS, etc) are leaky because they are remote. The connection may break down and the access is much slower than local file access.
• The finiteness of real-world machines leaks. Programming languages are said to be Turing complete. They theoretically are as powerful as a Turing machine. In fact they are not. Turing machines are conceptual automata with infinite memory. But in the real world memory is finite. There can be OutOfMemoryErrors: RAM and hard disks and any other kind of storage can and do get full.

Complete

• Joel Spolsky: The Law of Leaky Abstractions
• Joel Spolsky: Lord Palmerston on Programming About some of the consequences of the Law of Leaky Abstractions.
• Thomas Drakengren: The Law of Leaky Abstractions Refined
• Two Kinds of Leaky Abstractions!

Discuss this wiki article and the principle on the corresponding talk page.