Photo by Natalia Y on Unsplash

Using Go Generics

I was pairing with another engineer the other day and we needed to download some largish mp4 files and process the headers. The service could possibly get multiple requests at once, so we wanted to deduplicate the expensive requests and return the same value to all the concurrent requesters. I’ve often written similar code and with Go’s excellent concurrency primitives it’s usually a pretty simple task. After about an hour we had the code working and it seemed pretty straight-forward. The next day when I was reviewing the PR, I noticed a possible deadlock. I guess this wasn’t as simple as I had first thought. After fixing the deadlock, I decided to extract the core of the code into a reusable library so the next person wouldn’t make the same mistake.

I think you might be guessing where this is going. Once the package was extracted to be reusable, the concrete types we originally used for our specific use case all became the empty interface (interface{}).

I had been casually following the development of generics in Go. I read several blog posts and articles, but I hadn’t dug into the proposals or written any code. I knew the latest proposal was approved and should be landing in a future go release (probably 1.18). This little library (deduper) seemed like a great opportunity to try out generics in Go.

Getting Started

Edit 2021–8–23: Generics landed on go-tip; updated instructions to remove go2go usage

The first thing I realized is that the generics stuff isn’t included in the current mainline version of Go (1.17). To play with generics, you’ll have to grab the tip and build yourself.

git clone goroot
cd goroot
cd src
cd ..
export PATH=$(pwd)/bin:$PATH

That last line will point your path for the go and gofmt tools to the branch in the terminal where you run it, so be careful if you’re doing other go development not to use this pre-release go tool.


Once initialized, using a deduper is simple; just call the Get method:

dd := deduper.NewDeduper(…)
value, err := dd.Get(request)

In the go1 version you’d then have to type assert value to your concrete type. Using Go generics, I’ve replaced that with type T so

func (dd *Deduper) Get(m Request) (interface{}, error)


func (dd *Deduper[T]) Get(m Request) (T, error)

The Request passed into the Get call is an interface:

type Request interface {
Payload() interface{}

The Payload function returns a value that is used by the worker function to make the request. In our code it was a struct representing a database record, but it could be an http.Request or anything else. That gives us our second generic type making Request into:

type Request[U any] interface {
Payload() U

and Get into:

func (dd *Deduper[T, U]) Get(m Request[U]) (T, error)

Now that we have our generic return type T and Payload type U we can define the worker function as:

type WorkerFunc[T any, U any] func(req Request[U]) (T, error)

I decided to pass the entire Request but, I could have also just as easily passed in the only the Payload:

type WorkerFunc[T any, U any] func(payload U) (T, error)

That covers the entire public API for the library. The rest of the work was propagating these generic types throughout the internals of the code. This was mostly copy/paste to add the generic types and type parameters where needed. The compiler is your friend on this step as go compile will helpfully point out what you’re missing.

Using the Generic Library

In this short example, both T and U are type string:

workerFunc := func(req deduper.Request[string]) (string, error) { 

dd := deduper.NewDeduper(3, workerFunc) // 3 workers
val, err := dd.Get(myRequest) // val is type string

I didn’t need to specify types for either T or U when I called NewDeduper because the types could be inferred from workerFunc. The variable val is a concrete string and doesn’t need a type assertion. Neither does the string returned by calling req.Payload(). This is great, now I have compile time type safety! Bonus is that with the type inference, the calling code looks almost the same. I think this will apply to many use cases when generics are added, the calling code will look like regular Go with a few less type assertions and compile time type checking.

Comparing Using Generics vs. Not

After I got everything working on go2 with generics, I decided to check how different the code is without them. It was pretty trivial to delete all the type parameter lists and switch the T and U back to interface{}. I also back ported my example to go1. If we compare a few pieces of code we can see the difference.


val, _ := dd.Get(helloRequest(“world!”))


val, _ := dd.Get(helloRequest(“world!”))
// not strictly necessary for this trivial example to do the
// type assertion
if valStr, ok := val.(string); ok {

I needed to add a type assertion and possibly handle a runtime error case if the type isn’t what I expect. The difference is just a few line of extra code.

Let’s look at the Payload usage as well.


return fmt.Sprintf(“Hello, %v”, req.Payload()), nil


// not strictly necessary for this trivial example to do the 
// type assertion
if payload, ok := req.Payload().(string); ok {
return fmt.Sprintf(“Hello, %v”, payload), nil
return nil, fmt.Errorf(“Value not expected string %T”, req.Payload())

This includes error handling in case the type assertion fails, but the non-generic code is still only a few extra lines of code.

Comparing Writing Generics vs. Not

Looking at the library code, we find that it is significantly more complicated with all the type parameter lists. The internals of the deduper library has multiple internal data structures like channels that use the generic types. Type inference didn’t help much here and I needed to add the type parameter lists to almost every function. Here are a few internal examples:

func (dd *Deduper[T, U]) queueRequestChan(request *requestWrapper[T, U], savedRequest map[interface{}][]*requestWrapper[T, U])func (dd *Deduper[T, U]) processResult(rw *resultWrapper[T, U], savedRequest map[interface{}][]*requestWrapper[T, U])

Unlike Java, where the generic types are defined on the class and used consistently, every type or function in Go requires redefinition of the parameter types. This means you can use different placeholders in every instance. Like using different method receiver variables, the solution is “don’t do that.” I accidentally reversed my types at one point, but when I tried to use the wrong type the compiler caught it. I was very impressed with the clarity of the compiler error messages. The most common one is an alert when a required parameter list is missing. The messages were so clear and accurate I’m assuming this will autocomplete in the future.

State of Tooling

If you’re going to mess around with Generics right now (Aug 2021), you’re going to give up a lot of nice bells and whistles. I use VS Code as my daily driver and I could force Go mode on the go2 files. This worked ok and highlighted keywords, comments and constants. It doesn’t recognize the definitions with parameter lists, so things like autocomplete, automatic error checking and click-through navigation don’t work. None of the other tooling worked either like lint or vet. I also had to run gofmt from the command line like it’s 2015.

When I first tried this I used the experimental go2go tool. While generics have now landed in the tip, it’s still experimental and you may see similar issues. The tooling provided by go2go is also bare bones compared to the current main line. Switches like -cover aren’t supported. It’s hard to figure out what is supported because the normal help and usage don’t work either. There were also some weird functionality bugs. When I was intentionally testing my deadlock bug, go test just hung and never detected it. Though the same deadlock in a main file panicked as expected. I also tried using the “hash/maphash” package and that created a random deadlock 100% of the time. If you’re going to try it out expect a few issues like these though I’m sure all the tools and bugs will get updated as generics get closer to launch.


I found it easy and intuitive to use Go’s new generics. I should caveat this by saying that I am familiar with generics in Java and, once I figured out the syntax, that knowledge was easily transferred. I barely skimmed the first part of the proposal before writing the whole library. I didn’t need any fancy constraints for this project.

I’ve been on the fence watching the community’s battle over adding generics to Go. I can see both side of the argument; generics can add complexity and hurt readability vs. the ability to write more reusable code. I’ve read several other articles that include very basic usage, usually targeting list manipulation utilities where generics provide an obvious benefit. I needed to try them myself to form my own opinion, so taking a real piece of code and porting it would be a good way to test the waters.

Working with Go’s generics was simple and straight-forward with my past experience using generics in Java. It took minimal reading just to get the syntax, then the helpful compiler errors helped with the rest. Day to day editing is rough because of lack of tooling, but that should get fixed as generics approach release. I was impressed by how few issues I ran into using a pre-release feature. I was expecting to spend large amounts of time trying to figure out what I had missed based on vague compilation failures, but at every point it was obvious what I needed to fix.

Unfortunately, I think the type parameter lists significantly reduce readability of the library code. Both examples are in the repo so you can judge for yourself. When I compare the two, the benefit of saving a few lines of type checking to the cost in readability of the library code, generics don’t seem worth it to me in this use case. It also takes a significant amount of time to propagate the generic types throughout the code. I’m also worried that many developers will lose the proper design of their code in generic soup. It’s much harder to see design flaws and even bugs when the types are generic rather than concrete. I got confused a few times in the this code and I was already starting from a working design. This was an issue I saw often with generics in Java.

I should note that we’re not actually using this library in production. The path we took was summarized by:

A little copying is better than a little dependency.

— Rob Pike

Our code copies the functionality and replaces the interface{} with the concrete types that we needed. I’m not sure if Rob would agree that a 100+ lines of copying qualifies as “A little,” but it avoids the type checking issue entirely. It also removes a few levels of abstraction that was needed to make the code reusable. This makes the footprint even smaller and easier to manage.

I know these same points have been well documented throughout the entire generics debate, but I’m glad that I took the time to experiment myself and I encourage others to do so as well. I like Go’s clean and non-intrusive approach as well as the clear benefits we’ll get from generic libraries for container and math operations. My opinion about using generics will likely change over time once generics are in broader use (especially in the standard library). I hope that Go keeps its clean and simple look and feel that makes it such a joy to program.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store