According to When is primitive obsession not a code smell?, I should create a ZipCode object to represent a zip code instead of a String object.

However, in my experience, I prefer to see

public class Address{
    public String zipCode;

instead of

public class Address{
    public ZipCode zipCode;

because I think the latter one requires me to move to the ZipCode class to understand the program.

And I believe I need to move between many classes to see the definition if every primitive data fields were replaced by a class, which feels as if suffering from the yo-yo problem (an anti-pattern).

So I would like to move the ZipCode methods into a new class, for example:


public class ZipCode{
    public boolean validate(String zipCode){


public class ZipCodeHelper{
    public static boolean validate(String zipCode){

so that only the one who needs to validate the zip code would depend on the ZipCodeHelper class. And I found another "benefit" of keeping the primitive obsession: it keeps the class looks like its serialized form, if any, for example: an address table with string column zipCode.

My question is, is "avoiding the yo-yo problem" (move between class definitions) a valid reason to allow the "primitive obsession"?

  • 9
    @jpmc26 Then you would be shocked to see how complex our zip code object is -- not saying it's right, but it does exist Jan 25, 2019 at 3:33
  • 9
    @jpmc26, I fail to see how you get from "complex" to "badly-designed." Complex code is often the result of simple code coming into contact with the complexity of the real world rather than the ideal world we might wish existed. "Back to that two page function. Yes, I know, it’s just a simple function to display a window, but it has grown little hairs and stuff on it and nobody knows why. Well, I’ll tell you why: those are bug fixes."
    – Kyralessa
    Jan 25, 2019 at 10:13
  • 21
    @jpmc26 - the point of wrapping objects like ZipCode is type safety. Zip code is not a string, it's a zip code. If a function expects a zip code, you should only be able to pass a zip code, not a string. Jan 25, 2019 at 12:51
  • 4
    This feels highly language specific, different languages do different things here. @DavorŽdralo In the same stretch we should also add a lot of numeric types. "only positive integers", "only even numbers" could all also be types.
    – paul23
    Jan 25, 2019 at 14:07
  • 6
    @paul23 Yes indeed, and the main reason we don't have those is that a lot of languages don't support elegant ways to define them. It's perfectly reasonable to define "age" as a different type from "temperature in degrees celsius", if only so that "userAge == currentTemperature" is detected as nonsense.
    – IMSoP
    Jan 25, 2019 at 18:08

9 Answers 9


The assumption is that you don't need to yo-yo to the ZipCode class to understand the Address class. If ZipCode is well-designed it should be obvious what it does just by reading the Address class.

Programs are not read end-to-end - typically programs are far too complex to make this possible. You cannot keep all the code in a program in your mind at the same time. So we use abstractions and encapsulations to "chunk" the program into meaningful units, so you can look at one part of the program (say the Address class) without having to read all code it depends on.

For example I'm sure you don't yo-yo into reading the source code for String every time you encounter String in code.

Renaming the class from ZipCode to ZipCodeHelper suggest there now is two separate concepts: a zip code and a zip code helper. So twice as complex. And now the type system cannot help you distinguish between an arbitrary string and a valid zip code since they have the same type. This is where "obsession" is appropriate: You are suggesting a more complex and less safe alternative just because you want to avoid a simple wrapper type around a primitive.

Using a primitive is IMHO justified in the cases where there is no validation or other logic depending on this particular type. But as soon as you add any logic, it is much simpler if this logic is encapsulated with the type.

As for serialization I think it sounds like a limitation in the framework you are using. Surely you should be able to serialize a ZipCode to a string or map it to a column in a database.

  • 2
    I agree with the "meaningful units" (main-) part, but not so much that a zip code and zip code validation are the same concept. ZipCodeHelper (which I would rather call ZipCodeValidator) might very well establish a connection to a web service to do it's job. That would not be part of the single responsibility "hold the zip code data". Making the type system disallow invalid zip codes can still be achieved by making the ZipCode constructor the equivalent of Java's package-private and calling that with a ZipCodeFactory which always calls the validator.
    – R. Schmitz
    Jan 24, 2019 at 13:00
  • 16
    @R.Schmitz: That is not what "responsibility" means in the sense of the single responsibility principle. But in any case, you should of course use as many classes as you need as long as you encapsulate the zip code and its validation. The OP suggest a helper instead of encapsulating the zip code, which is a bad idea.
    – JacquesB
    Jan 24, 2019 at 14:05
  • 1
    I want to respectfully disagree. SRP means a class should have "one, and only one, reason to be changed" (change in "what a zipcode consists of" vs. "how it is validated"). This specific case here is further elaborated on in the book Clean Code: "Objects hide their data behind abstractions and expose functions that operate on that data. Data structure expose their data and have no meaningful functions." - ZipCode would be a "data structure" and ZipCodeHelper an "object' . In any case, I think we agree that we shouldn't have to pass web connections to the ZipCode constructor.
    – R. Schmitz
    Jan 24, 2019 at 16:31
  • 9
    Using a primitive is IMHO justified in the cases where there is no validation or other logic depending on this particular type. => I disagree. Even if all values are valid, I would still favor conveying the semantics to the language rather than use primitives. If a function can be called on a primitive type which is nonsensical for its current semantic usage, then it should not be a primitive type, it should be a proper type with only the sensible functions defined. (As an example, using int as ID allows multiplying an ID by an ID...) Jan 24, 2019 at 19:47
  • @R.Schmitz I think ZIP codes are a poor example for the distinction you're making. Something that changes often might be a candidate for separate Foo and FooValidator classes. We could have a ZipCode class that validates the format and a ZipCodeValidator that hits some Web service to check that a correctly formatted ZipCode is actually current. We know that ZIP codes change. But practically, we're going to have a list of valid ZIP codes encapsulated in ZipCode, or in some local database. Jan 24, 2019 at 19:53

If can do:

new ZipCode("totally invalid zip code");

And the constructor for ZipCode does:

ZipCodeHelper.validate("totally invalid zip code");

Then you've broken encapsulation, and added a pretty silly dependency to the ZipCode class. If the constructor doesn't call ZipCodeHelper.validate(...) then you have isolated logic in its own island without actually enforcing it. You can create invalid zip codes.

The validate method should be a static method on the ZipCode class. Now the knowledge of a "valid" zip code is bundled together with the ZipCode class. Given that your code examples look like Java, the constructor of ZipCode should throw an exception if an incorrect format is given:

public class ZipCode {
    private String zipCode;

    public ZipCode(string zipCode) {
        if (!validate(zipCode))
            throw new IllegalFormatException("Invalid zip code");

        this.zipCode = zipCode;

    public static bool validate(String zipCode) {
        // logic to check format

    public String toString() {
        return zipCode;

The constructor checks the format and throws an exception, thereby preventing invalid zip codes from being created, and the static validate method is available to other code so the logic of checking the format is encapsulated in the ZipCode class.

There is no "yo-yo" in this variant of the ZipCode class. It's just called proper Object Oriented Programming.

We are also going to ignore internationalization here, which may necessitate another class called ZipCodeFormat or PostalService (e.g. PostalService.isValidPostalCode(...), PostalService.parsePostalCode(...), etc.).

  • 29
    Note: The main advantage with @Greg Burkhardt's approach here is that if someone gives you a ZipCode object, you can trust that it contains a valid string without having to check it again, since its type and the fact that it was successfully constructed gives you that guarantee. If you instead passed strings around, you might feel a need to "assert validate(zipCode)" at various places in your code just to be sure that you had a valid zip code, but with a successfully constructed ZipCode object, you can trust that its contents are valid without having to check them again.
    – Some Guy
    Jan 24, 2019 at 19:42
  • 3
    @R.Schmitz: The ZipCode.validate method is the pre-check that can be performed before invoking a constructor that throws an exception. Jan 25, 2019 at 12:10
  • 10
    @R.Schmitz: If you are concerned about a vexing exception, an alternate approach to construction is to make the ZipCode constructor private, and provide a public static factory function (Zipcode.create?) that performs validation of the passed-in parameters, returns null if unsuccessful, and otherwise constructs a ZipCode object and returns it. The caller will always have to check for a null return value, of course. On the other hand, if you are in the habit, for instance, of always validating (regex? validate? etc.) before constructing a ZipCode, the exception may not be so vexing in practice.
    – Some Guy
    Jan 25, 2019 at 12:45
  • 11
    A factory function that returns an Optional<ZipCode> is also a possibility. Then the caller has no choice but to explicitly handle possible failure of the factory function. Regardless, in either case, the error will be discovered somewhere near where it was created rather than possibly much later, by client code far from the original problem.
    – Some Guy
    Jan 25, 2019 at 12:48
  • 6
    You can't validate ZipCode indepenently, so don't. You really need the Country object to look up the ZipCode/PostCode validation rules.
    – Joshua
    Jan 26, 2019 at 0:19

If you wrestle a lot with this question, perhaps the language you use is not the right tool for the job? This kind of "domain-typed primitives" are trivially easy to express in, for example, F#.

There you could, for example, write:

type ZipCode = ZipCode of string
type Town = Town of string

type Adress = {
  zipCode: ZipCode
  town: Town

let adress1 = {
  zipCode = ZipCode "90210"
  town = Town "Beverly Hills"

let faultyAdress = {
  zipCode = "12345"  // <-Compiler error
  town = adress1.zipCode // <- Compiler error

This is really useful for avoiding common mistakes, like comparing id's of different entities. And since these typed primitives are much more lightweight than a C# or Java-class, you'll end up actually use them.

  • Interesting - how would it look like if you wanted to enforce validation of ZipCode?
    – Hulk
    Jan 25, 2019 at 8:04
  • 4
    @Hulk You can write OO-style in F# and make the types into classes. However, I prefer functional style, declaring the type with type ZipCode = private ZipCode of string and adding a ZipCode module with a create function. There are some examples here: gist.github.com/swlaschin/54cfff886669ccab895a
    – Guran
    Jan 25, 2019 at 8:12
  • @Bent-Tranberg Thanks for the edit. You are right, a simple type abbreviation does not give compile time type security.
    – Guran
    Jan 28, 2019 at 8:17
  • If you was mailed my first comment, that I deleted, the reason was that I first misunderstood your source. I didn't read it carefully enough. When I tried to compile it, I finally realized that you were actually trying to demonstrate exactly this, so then I decided to edit to get it right. Jan 28, 2019 at 12:28
  • Yeah. My original source was valid alright, unfortunately including the example that was SUPPOSED to be invalid. Doh! Should hav just linked to Wlaschin instead of typing code myself :) fsharpforfunandprofit.com/posts/…
    – Guran
    Jan 28, 2019 at 12:42

The answer depends entirely on what you actually want to do with the ZIP codes. Here are two extreme possibilities:

(1) All addresses are guaranteed to be in a single country. No exceptions at all. (E.g. no foreign customers, or no employees whose private address is abroad while they are working for a foreign customer.) This country has ZIP codes and they can be expected to never be seriously problematic (i.e. they don't require free-form input such as "currently D4B 6N2, but this changes every 2 weeks"). The ZIP codes are used not just for addressing, but for validation of payment information or similar purposes. - Under these circumstances, a ZIP code class makes a lot of sense.

(2) Addresses can be in almost every country, so dozens or hundreds of addressing schemes with or without ZIP codes (and with thousands of weird exceptions and special cases) are relevant. A "ZIP" code is really only asked for to remind people from countries where ZIP codes are used not to forget to provide theirs. The addresses are only used so that if someone loses access to their account and they can prove their name and address, access will be restored. - Under these circumstances, ZIP code classes for all relevant countries would be an enormous effort. Fortunately they are not needed at all.


The other answers have talked about OO domain modelling and using a richer type to represent your value.

I don't disagree, especially given the example code you posted.

But I also wonder if that actually answers the title of your question.

Consider the following scenario (pulled from an actual project I'm working on):

You have a remote application on a field device that talks to your central server. One of the DB fields for the device entry is a zip code for the address that the field device is at. You don't care about the zip code (or any of the rest of the address for that matter). All of the people who care about it are on the other side of an HTTP boundary: you just happen to be the single source of truth for the data. It has no place in your domain modeling. You just record it, validate it, store it, and on request shuffle it off in a JSON blob to points elsewhere.

In this scenario, doing much of anything beyond validating the insert with an SQL regex constraint (or its ORM equivalent) is probably overkill of the YAGNI variety.

  • 6
    Your SQL regex constraint could be viewed as a qualified type - within your database, the Zip code is not stored as "VarChar" but "VarChar constrained by this rule". In some DBMSes, you could easily give that type+constraint a name as a reusable "domain type", and we are back in the recommended place of giving the data a meaningful type. I agree with your answer in principle, but don't think that example matches; a better example would be if your data is "raw sensor data", and the most meaningful type is "byte array" because you have no idea what the data means.
    – IMSoP
    Jan 24, 2019 at 16:11
  • @IMSoP interesting point. Not sure I agree though: you could validate a string zip code in Java (or any other language) with a regex but still be dealing with it as a string instead of a richer type. Depending on the domain logic, further manipulation might be required (for instance ensuring that the zip code matches the state, something that would be difficult/impossible to validate with regex). Jan 24, 2019 at 16:58
  • You could, but as soon as you do so, you are ascribing it some domain-specific behaviour, and that is exactly what the quoted articles are saying should lead to the creation of a custom type. The question is not whether you can do it in one style or the other, but whether you should, assuming your programming language gives you the choice.
    – IMSoP
    Jan 24, 2019 at 17:20
  • You could model such a thing as a RegexValidatedString, containing the string itself and the regex used to validate it. But unless every instance has a unique regex (which is possible but unlikely) this seems a bit silly and wasteful of memory (and possibly regex compilation time). So you either put the regex into a separate table and leave behind a lookup key in each instance to find it (which is arguably worse due to indirection) or you find some way to store it once for each common type of value sharing that rule -- eg. a static field on a domain type, or equivalent method, as IMSoP said.
    – Miral
    Jan 25, 2019 at 2:59

The ZipCode abstraction could only make sense if your Address class did not also have a TownName property. Otherwise, you have half an abstraction: the zip code designates the town, but these two related bits of information are found in different classes. It doesn't quite make sense.

However, even then, it's still not a correct application (or rather solution to) primitive obsession; which, as I understand it, mainly focuses on two things:

  1. Using primitives as the input (or even output) values of a method, especially when a collection of primitives is needed.
  2. Classes that grow extra properties over time without ever reconsidering whether some of these should be grouped into a subclass of their own.

Your case is neither. An address is a well-defined concept with clearly necessary properties (street, number, zip, town, state, country, ...). There is little to no reason to break up this data as it has a single responsibility: designate a location on Earth. An address requires all of these fields in order to be meaningful. Half an address is pointless.

This is how you know that you don't need to subdivide any further: breaking it down any further would detract from the functional intention of the Address class. Similarly, you don't need a Name subclass to be used in the Personclass, unless Name (without a person attached) is a meaningful concept in your domain. Which it (usually) isn't. Names are used for identifying people, they usually have no value on their own.

  • 1
    @RikD: From the answer: "you don't need a Name subclass to be used in the Person class, unless Name (without a person attached) is a meaningful concept in your domain." When you have custom validation for the names, the name has then become a meaningful concept in your domain; which I explicitly mentioned as a valid use case for using a subtype. Secondly, for the zipcode validation, you're introducing additional assumptions, such as zip codes needing to follow a given country's format. You're broaching a topic that's much broader than the intent of OP's question.
    – Flater
    Jan 24, 2019 at 9:59
  • 5
    "An address is a well-defined concept with clearly necessary properties (street, number, zip, town, state, country)." - Well that's just plain wrong. For a good way to deal with this, look at Amazon's address form.
    – R. Schmitz
    Jan 24, 2019 at 13:13
  • 4
    @Flater Well I won't blame you for not reading the full list of falsehoods, because it's quite long, but it literally contains "Addresses will have a street", "An address require both a city and a country", "An address will have a postcode" etc., which is contrary to what the quoted sentence says.
    – R. Schmitz
    Jan 24, 2019 at 13:27
  • 8
    @GregBurghardt "Zip code assumes United States Postal Service, and you can derive the name of the town from the zip code. Cities can have multiple zip codes, but each zip code is only tied to 1 city." This is not correct in general. I have a zipcode that is used mainly for a neighboring city but my residence is not located there. Zipcodes do not always align with governmental boundaries. For example 42223 contains counties from both TN and KY.
    – JimmyJames
    Jan 24, 2019 at 14:45
  • 2
    In Austria there exists a valley which is only accessible from Germany (en.wikipedia.org/wiki/Kleinwalsertal). There exists a special treaty for this region which, among other things, also includes that adresses in this area have both Austrian and German postal codes. So in general, you cannot even assume that an address has only a single valid postal code ;)
    – Hulk
    Jan 25, 2019 at 8:21

From the article:

More generally, the yo-yo problem can also refer to any situation where a person must keep flipping between different sources of information in order to understand a concept.

Source code is read far more often than it is written. Thus, the yo-yo problem, of having to switch between many files is a concern.

However, no, the yo-yo problem feels much more relevant when dealing with deeply interdependent modules or classes (which call back and forth between each other). Those are a special kind of nightmare to read, and is likely what the coiner of the yo-yo problem had in mind.

However - yes, avoiding too many layers of abstraction is important!

All non-trivial abstractions, to some degree, are leaky. - the Law of Leaky Abstractions.

For example, I disagree with the assumption made in mmmaaa's answer that "you don't need to yo-yo to [(visit)] the ZipCode class to understand the Address class". My experience has been that you do - at least the first few times you read the code. However, as others have noted, there are times when a ZipCode class is appropriate.

YAGNI (Ya Ain't Gonna Need It) is a better pattern to follow to avoid Lasagna code (code with too many layers) - abstractions, such as types and classes are there to aid the programmer, and should not be used unless they are an aid.

I personally aim to "save lines of code" (and of course the related "save files/modules/classes", etc). I'm confident there are some who would apply to me the epithet of "primitive obsessed" - I find it more important to have code which is easy to reason about than to worry about labels, patterns, and anti-patterns. The correct choice of when to create a function, a module/file/class, or put a function in a common location is very situational. I aim roughly for 3-100 line functions, 80-500 line files, and "1, 2, n" for reusable library code (SLOC - not including comments or boilerplate; I typically want at least 1 additional SLOC minimum per line of mandatory boilerplate). The important point is to keep in mind and respect the limits of human cognition when writing code.

Most positive patterns have arisen from developers doing exactly that, when they needed them. It is much more important to learn how to write readable code than to try to apply patterns without the same problem to solve. Any good developer can implement the factory pattern without having seen it before in the uncommon case where it is the right fit for their problem. I have used the factory pattern, the observer pattern, and probably hundreds besides, without knowing their name (ie, is there a "variable assignment pattern"?). For a fun experiment - see how many GoF patterns are built into the JS language - I stopped counting after about 12-15 back in 2009. The Factory pattern is as simple as returning an object from a JS constructor, for example - no need for a WidgetFactory.

So - yes, sometimes ZipCode is a good class. However, no, the yo-yo problem is not strictly relevant.


The yo-yo problem is only relevant if you have to flip back and forth. That is caused by one or two things (sometimes both):

  1. Bad naming. ZipCode seem fine, ZoneImprovementPlanCode is going to require a look by most people (and the few that don’t won’t be impressed).
  2. Inappropriate coupling. Say you ZipCode class has an area code lookup. You might think it makes sense because it’s handy, but it’s not really related to the ZipCode, and shoeing it into it means that now people don’t know where to go for things.

If you can look at the name and have a reasonable idea of what it does, and the methods and properties do reasonably obvious things, you don’t need to go look at the code, you can just use it. That is the whole point of classes in the first place—-they are modular pieces of code that can be used and developed in isolation. If you have to look at anything other than the API for the class to see what it does, it is at best a partial failure.


Remember, there is no silver bullet. If you are writing an extremely simple app which needs to be crawled through fast, then a simple string may do the job. However in 98% of the times, a Value Object as described by Eric Evans in DDD, would be the perfect fit. You can easily see all the benefits Value objects provide by reading around.


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