Error code pattern for API

Question

What are the good choice for API error code response pattern?

Instead of using different codes indicating different type of error

100001 // username not provided
100002 // password not provided
100003 // password too short
...

I see some other use patterns like the following (non-sequential) ...

20000
20001
20004
20015

Are there any other recommendations?

Answer 1

In my experience developing and using web services, I have found that a strategy of using a combination of top-level HTTP status codes and lower level API error codes work reasonably well. Note that the lower level API error codes don't need to be integers, but can be any enumeration. For a well-known public example, AWS Simple Email Service (SES) uses this strategy of using both HTTP status codes and API level error codes. You can see a sample error code response for SES here. Note that although SES uses XML response error payloads, this strategy works equally well for JSON response payloads.

In my experience, there are a few things that you need to keep in mind when using this strategy:

1. Strive to return the correct HTTP response code: HTTP is a ubiquitous protocol and is no doubt understood by your web container. Its response codes fit naturally into REST web services. As such, leverage it! If your web service encounters an error condition, you should do your best to return the correct HTTP status code in whose context, the API error code has meaning. One my biggest headaches in debugging issues with web services occur when developers just unconditionally throw arbitrary (usually runtime) exceptions back up the stack. The result is that everything gets returned back to the caller as an HTTP 500 (Internal Server Error) status code even when that's not the case (e.g. the client sends garbage data and the server just can't process it. Some common HTTP status codes you might want to design for include:
   • 400 Bad Request: There is an issue with the client's request. Note this error isn't just used for things like broken JSON syntax in a POST request, but it is also a legitimate response code for semantic issues as well (i.e. the JSON request payload conformed to the prescribed schema, but there was an issue with the data in the payload, such as a number being negative when it is supposed to be only positive).
   • 401 Unauthorized: The caller's credentials were invalid (i.e. authorization error).
   • 403 Forbidden: The caller's credentials were valid, but their access level isn't sufficient to access the resource (i.e. authentication error).
   • 404 Not Found: The resource of the URL doesn't exist.
   • 500 Internal Server Error: Something bad happened inside the server itself, this error could be anything.
   • 502 Bad Gateway: An error occurred when calling downstream service.
   • 503 Service Unavailable: A useful response code for when you get hammered with a ton of "happy" customers who are inadvertently DDOS'ing your service.
   • 504 Gateway Timeout: Like the 502 status code, but indicates a timeout instead of an actual error with the downstream service, per se.
2. HTTP response codes are the top-level codes, and API error codes only have meaning within that context: By this, I mean that your API error codes are only meaningful for certain HTTP response codes. For example, in the table of SES error codes, each error code is only tied to a single HTTP(S) response code. The error codes ConfigurationSetDoesNotExist and InvalidParameterValue only make sense when a 400 Bad Request is returned by SES - it wouldn't make sense to return these status codes when a 500 Internal Server Error is returned. Similarly, if you were writing a web service that called downstream services and databases, you might have a FooDownstreamServiceTimedOut error code that you would return with a 504 Gateway Timeout HTTP status code when a downstream web service call timed out to the "Foo" web service. You might also have a MyDatabaseError error code that you would return with a 500 Internal Server Error HTTP status code when your query to the internal DB fails.
3. Have a uniform error code schema irrespective of status codes: Your clients need to be able to process your error content programmatically. As such, it needs to conform to a certain schema. Ideally, your API error code schema should include the error code (i.e. name or ID, etc.). You also probably want to include a natural language description of the error code and the ID/GUID of the request that you are responding to. For an example of an error schema, see this sample AWS SES response and schema. Additionally, you might also want to consider returning a client ID in the response. This is as much for your own benefit as the client's since it can help you drill down into the data to see if one particular client is getting a glut of particular errors vs. your other clients.
4. Consider returning natural language descriptions of the error codes in the response: To make things easier on your clients, you might want to consider not just returning the error code in the error payload, but a natural language description as well. This kind of behavior can immediately help confused and busy engineers who really don't care that much about your service quickly diagnose what's happening so that they can resolve the issue ASAP. btw, enabling engineers to quickly diagnose issues with your service increases the all-important "uptime" metric that your customers and managers will no doubt care about.
5. Don't feel obliged to use integers, use enumerations instead: The notion of "error codes" conjures up images of outdated technologies and codebooks where you had to look up what an error meant. It arose from the programming dark ages when engineers needed to fit all possible errors into a byte of space, or a nibble or whatever. Those days are gone, and your error code can be a string, likely without any meaningful impact on performance. You might as well take advantage and make the error code meaningful, as a means of keeping things simple.
6. Return info to clients that they might need to debug, but be mindful of security: If possible, return whatever debug info your clients may need. However, if your service potentially deals with sensitive information such as credit card numbers and the like, you probably don't want to pass that info around for obvious reasons.

Hope that helps.

Answer 2

A recommendation by the IETF (internet standards body) is using the application/problem+json mediatype.

Notable is that they don't use random numbers, they use strings (specifically uris) to identify errors.

This is a subjective question, but even if you don't use their format, I'd argue that username-not-provided is better in almost every way to 100001.

Answer 3

I would say this heavily depends on what kind of API you're providing.

I were to always include a field called ack or something similar in every response that has three states: failure, warning, success. Success obviously being everything went well. On warning, the request went through and the JSON will contain the expected output, but it will also include a warning string, or even better in case multiple warnings could occur an array called errors which consists of multiple objects containg code, string and type. This array will also be returned in case of failure, and nothing else but this array.

The array contains one object per error or warning, having a code (I would suggest going with your initial idea of 10001, 10002, ...) and a string explaining the error in a very short phrase (e.g. Username contains invalid characters). The type is either error or warning, which is useful in case of a failure ack that contains not only errors but also warnings.

This makes it easy to look up errors by their code (I would provide a page, also with an API, that contains all the error codes in a table along with their short and long description plus common causes/fixes/etc. - All this information should also be available via an API where they can be accessed by providing the error code) while still having a quick short text response so the user can tell what's wrong in most cases without having to look up the error.

This also allows for easy output of warnings and errors to the end user, not just the developers. Using my idea with the API call to get informations about an error, developers using your API could easily provide full information about errors to end-users when needed (including causes/fixes/whatever you see fit).

Answer 4

Instead of writing your own API standard from scratch adopt one of the already available, for example the JSON API standard:

If you’ve ever argued with your team about the way your JSON responses should be formatted, JSON API can be your anti-bikeshedding tool.

By following shared conventions, you can increase productivity, take advantage of generalized tooling, and focus on what matters: your application.

Clients built around JSON API are able to take advantage of its features around efficiently caching responses, sometimes eliminating network requests entirely.

If you decide to go with JSON API it has a section dedicated to errors and a few error examples.

Answer 5

For many years, many developent companies have created things like bitmask for errors, so they can encode multiple variables inside the error:

000 - all ok
001 - something failed with X
010 - something failed with Y
011 - something failed with X and Y
100 - something failed with Z
101 - something failed with X and Z

The limitation is that that limits the error space into however many bytes you decide on the encoding, like 16 or 32 possible combinations, it may be enough for you, or not.

You see this being common in COM+ https://learn.microsoft.com/en-us/windows/desktop/com/com-error-codes-1

I hope this helps.