Hacker News new | past | comments | ask | show | jobs | submit login
Ask HN: Were you happy moving your API from REST to GraphQL?
301 points by glenscott1 on July 19, 2018 | hide | past | favorite | 175 comments

As a consumer of APIs I vastly prefer REST APIs.

In my opinion, GraphQL moves too much of the burden to the user of the API. It makes most sense if the data is highly dynamic, you have a mobile app and every call is expensive, or (and this seems more common) the backend and frontend teams don't like to talk to each other. As a user, I just want to GET /foo, with a good old API token I pasted from your dev docs, and move on to the next thing. I don't want to spend time figuring out your database schema. Or perhaps I've just yet to see single good public GraphQL API. I just recently had look at the Github GraphQL API and it's non-existent docs (no, I don't want introspect your data models at runtime), noped the hell out of that, and got the REST integration for the bit I needed done in an hour.

This is my experience as well. We briefly worked with GraphQL for one of our smaller projects. One of the biggest shortcomings of the API we created was that there was an incredibly large learning curve for teams/people not interested in using GraphQL. Instead of using the sensible features of GraphQL queries, the other team consuming our API were writing the most verbose queries they possibly could, not utilizing any of the nice convenience operations that GraphQL offers.

They were no doubt frustrated with what they thought was a needless hassle compared to REST, and I myself found a lot of the query building pretty tedious.

That's interesting, because as a frontend engineer, I find most RESTful APIs to introduce more "needless hassle", mostly in the form of making me send multiple requests to get data multiple levels deep and then consolidating it on the frontend, rather than just sending me what I want to begin with.

I agree with query building being tedious, though I'm working on a query builder helper library to make this less painful. Hopefully at some point I can open source it.

Why do you think it makes the most sense if “the backend and frontend teams don't like to talk to each other”, given that your biggest complaint seems to be “I don't want to spend time figuring out your database schema“? Aren’t you the frontend guy in this scenario, and not wanting/able to talk to the backend guys (the people who designed the database schema)?

He (the hypothetical he) doesn't want to talk to them to get them to add the endpoint he needs, so he'd prefer to just get the schema directly and query for whatever.

Note that this means the back end team has to keep the schema backwards compatible forever. Having everyone have direct access to the schema (speaking from SQL experience) is bad. This is the start of a slow moving disaster train that you can't stop.

I disagree; the GraphQL schema should be your point of stability, whereby anything in front or behind can change. That being said, ofc, you'll need to deprecate something at some point and that is fine, it might just take a couple of weeks. See @deprecated decorator examples

> the GraphQL schema should be your point of stability, whereby anything in front or behind can change

This is the bit of abstraction I see most developers fail to understand when evaluating GraphQL.

Not forever, just as long as you support out-of-date clients (either because of cached JS on the web, or mobile apps that haven't been updated recently). Plus, you generally don't want to introduce backwards-incompatible changes to your schema anyway at a certain scale, unless you can afford to have planned downtime, so GraphQL doesn't add anything new here.

I'm not quite sure what does the database schemas have to do with the front-end queries...

In this case OP is talking about the GraphQL schemas, not the DB.

I'm not sure which OP do you mean? The parent mentions "SQL experience", the grandparent is ambiguous, and the GGP explicitly mentions "database schema".

If by OP you mean the original article, I completely agree with you; my comment was directed to its ancestors.

The internal team would have a vested interest in introspecting the API but an end user (developer) would not.

But I actually found their graphql docs to be sensible? Not sure where the introspection comes into play

I <3 this comment.

How I _wish_ I could nope the hell out of our GraphQL dependency, for the reasons listed and for the frighteningly complex client libs we have to use to consume it, rather than straight up ajax/fetch.

You don't have to use anything other than `fetch` to consume a graphql API.

Right, then you basically lose caching, since the everything's-a-POST requests would go to the wire every time. Not that I haven't considered making that trade-off.

You can use GET if you want (since GraphQL is still just over HTTP). The implementation can largely be up to you (like REST) if you want, but you get loads of things already if you opt into some opinions that libraries have established (e.g. Apollo).

See https://graphql.org/learn/serving-over-http/

You can make your queries ride over GET very easily, using just fetch.

You shouldn't be using fetch directly accoriding to it's creator anyway - use a high level HTTP client which handles MIME types, query string encoding, etc.

How does the complexity bite you? Large files, bad error messages, bugs? Something else?

Limited error handling mechanism in Apollo and our backend GraphQL library means I can't use then/catch to reliably handle both network and server errors. Apollo also hijacks state management and somehow decides when to fetch over the network or render a cached response, with no particular logic I can deduce or find documentation for. Then there's the thousand lines of repetitive query code that is a joy to read/write/manage.

Sounds like the problem is with your client/server library choices and not with GraphQL as a concept. Your original comment implied the latter.

I think we handle our GraphQL like we do our SQL -- handle the repetitive queries as templates. I'm curious now and will dig into the repo tonight.

The state management is a problem by default. We wrote our own cache driver to fix it, which is pretty well documented thankfully. "Un-magic it" is a phrase we use a lot.

Recently tried out GraphQL in a new web app - and ripped it out all out shortly after from frontend and backend. Unnecessary complexity over simple rest calls with no benefits.

As a manager/business owner, I like GraphQL because it shifts the burden to the API consumer. That consumer, for us, is a front-end resource who is sometimes less expensive but very often less busy than our backend team.

That's obviously not the same formula for every company. I'm just offering it as a potential counterpoint.

This is a common swing point though:

PM: Make this change to the site

Backend: that would take 2 weeks to make sure it doesn't break anything

Frontend: That's a tiny change! I can do it in 2 hours

Frontend makes the change. Along with the next several. Overtime, the backend team uses the lessened pressure to overhaul/replace/refactor to decrease the complexity. Meanwhile, Frontend is starting to accumulate a collection of unrelated exceptions to the "normal" flow to meet demands. Eventually....

PM: Make this change to the site

Frontend: Um...that would risk all sorts of complications. Consider it a few weeks to make sure it doesn't break anything

Backend: That's a straightforward change, I can do it in 2 hours

And the pendulum starts to swing in the opposite direction...

If a frontend employee isn't earning a comparable salary to a backend employee, you're either not hiring people of equivalent skillset, or you're underpaying your frontend people.

Where the GraphQL burden lies is largely due to who is considered to be its product owner. This varies significantly depending on whether it's an API for first or third parties.

You don't know enough about my products to be so certain about my payment practices.

First of all, our backend is extremely complex and our frontend isn't. Our backend devs need a lot of legal and financial training. Other companies will have an inverse situation.

Second, there's a glut of frontend devs (at least when I look for them) because bootcamps seem to produce people who are better at frontend than backend.

And third, our frontend devs just don't have as much work to do, so they're idle a lot more often.

So you’re not hiring people of equivalent skillset, that’s fair enough.

I just found it a surprising comment in isolation because most people I’ve met who’ve developed both frontend and backend skills to a senior or greater level, would argue that frontend is the more challenging of the two.

The bit about idle time is a bit confusing though, can’t you just have fewer devs if there’s not enough work to go around? Or do you always need a lot of spare resources due to unpredictable workloads?

Interesting! Were you doing a server-to-github integration, or were you using the GraphQL API directly from the frontend?

Absolutely. Before GraphQL we were making a monumental effort to build a REST API. After deliberating on exactly what REST was and how we’d represent a few red haired resources, we were spending a lot of client time fetching deep trees through resource links. When we moved to GraphQL it solved a lot of the administrative and philosophical headaches and considerably reduced the number of connections, wasted data, and made our client code so much simpler through easily grokked queries. Highly recommmend GraphQL to anyone.

I should also mention that we finished our migration ahead of schedule. It was super easy to have GraphQL alongside REST, and we quickly iterated on converting each rest call to graphQL.

We’ve also found that on boarding our new hires is much simpler. There’s a lot of misinformation about REST, and we were having to retrain people, and when they wanted to see our schema we would then have to teach them swagger as well. With GraphQL we just send them to the official docs with our schema with is our single source of truth for the API and they come back a day later ready to go. Generally GraphQL being more standardized and centrally managed has been great from a training perspective.

Teach them Swagger? Do you mean to generate Swagger or to consume it? Because as far as clarity I don't think Swagger could be clearer, what with the interactive endpoint GUI and copy/paste curl commands and so on.

What language/framework was this being implemented on? If you dont mind me asking.

We wrote our own server in C++. We also have a few smaller graphql endpoints running the official JS implementation on express. On the client side we are using relay modern.

Out of curiosity, what does "tech them swagger" entail? I was under the impression that most of it was automated and that it's more of a standard than an implementation.

"We moved to GraphQL because things were bad, and now things are good. GraphQL is amazing".

I don't want this to come off as a personal attack (and I apologize if it does), but your comment contains absolutely no information whatsoever regarding a specific situation/use-case, nothing from which the rest of us can formulate our own opinions on the REST/GraphQL discussion.

>we were spending a lot of client time fetching deep trees through resource links

>it solved a lot of the administrative and philosophical headaches

>considerably reduced the number of connections

>considerably reduced wasted data

>made our client code so much simpler through easily grokked queries

I feel that grandparent does contain information which might be valuable for adoption.

My interpretation was: “We did ReST trying to follow dogma and ended up with a badly designed API that din’t fit our needs. Switching to a tech with less dogma on how to do things led us to a better API-design with a better fit” (inferring some context from my own experience here, could be wrong)

In the end I would guess you can end up in a similar place with a standard fetch-json design if you just ignore ReST dogma and focus on getting the API into a shape that fits the need.

Not saying ReST dogma is necessarily wrong, or bad, just that it’s easy to get lost in design when focusing more on learning others design than understanding the actual problem you’re trying to solve.

For reference we’re switching from the non dogmatic get/post requests over http to graphql and it’s been equally good for us too for different reasons.

Also consider that a second implementation of a system is always going to be cleaner than the first because you actually know what you need to do. This is an inherent bias inherent to any GraphQL migration story.

> considerably reduced the number of connections, wasted data, and made our client code so much simpler through easily grokked queries.

I mean, you could have also just done this with what you had.

Not really. With REST, the server dictates the minimum requirements of the client. In GraphQL, the client dictates its own requirements, and the server is able to respond with no more and no less than what the client requires. Also REST necessitates multiple HTTP queries to fetch multiple resources. This is not a requirement of GraphQL.

> the client dictates its own requirements, and the server is able to respond with no more and no less than what the client requires.

The burden of dictating specs is just shifted around, and it seems the workload on the server side is bigger with h GraphQL (wider range of cases to handle).

Am I missing something ?

> REST necessitates multiple queries

This is a self imposed limitation at best

Your graphql library will handle that for you. The server defines data as it knows, the client asks for data that it wants, and the library does the transform work.

What will the library handle ? From my understanding it’s only the query parsing and the data filtering part. Correct me if I’m wrong.

You still have to do the data mapping, fetching everything from DB in a reasonable way, and make sure it all makes sense performance wise.

All of that will be needed for any API, but it seems to me GraphQL adds the uncertainty on how much data will be exchanged (lots of small queries ? a few big queries ?), what can be optimized, how will cache behave etc.

I remember a study on te github API on how some types of queries would make it crawl excessively. It fear it becomes a nightmare to try to cover for all the cases that can go wrong.

Yes, it handles data filtering, field renaming, etc, which turns out to be a big chunk of work if you're working on APIs to support multiple frontend experiences (the BFF pattern).

I don't know what your use case is, but it sounds like you should just try it, I can only say that for us (multiple FE experiences, pushing logic to backend), it's been very good. If you have a different usecase, ymmv.

Thanks for the details.

It was this study: http://olafhartig.de/files/HartigPerez_WWW2018_Preprint.pdf

Overall it acknowledges GraphQL's advantages, while warning that the data retrieval part shouldn't be done naively.

An excerp

> We note that this issue is somehow acknowledged by the Github GraphQL interface and, as a safety measure to avoid queries that might turn out to be too resource-intensive, it introduces a few syntactic restrictions [7]. As one such restriction, Github imposes a maximum level of nesting for queries that it accepts for execution. However, even with this restriction (and other syntactic restric- tions imposed by the Github GraphQL interface [7]), Github fails to avoid all queries that hit some resource limits when executed

> With REST, the server dictates the minimum requirements of the client.

Says who?

> Also REST necessitates multiple HTTP queries to fetch multiple resources.

Again, says who?

With a standard REST interface, you have to name all such combinations in advance. Possibly build custom code for each. With GraphQL, you specify them all at once, and are guided towards implementing code that will handle that. (There's no magic in GraphQL, of course, but the conceptualization alone can be useful. And if you're already in some particular ecosystem, they may have some localized magic you can use.)

You could just take your GraphQL backend and implement those specialized REST calls, sure, but then why not expose it?

> With a standard REST interface

There's no such thing.

> you have to name all such combinations in advance.

No you don't.

> Possibly build custom code for each.

As much as any other API.

There is a "standard" REST interface described in the paper on REST. There are a lot of rules/guidelines for a RESTful API but many API's don't follow them and tend to be a mix of REST and JSON RPC.

Do you though?

... OData ...

It's a thing.

How do you do caching with your GraphQL APIs ? I guess you can't leverage HTTP-level caching.

What is a red haired resource?

A wildly offensive analogy.

I assume he means like a red-headed step-child.

PC police might be on to terms like this nowadays, but by using the term I understood them to mean:

- A less loved resource - A potentially ugly resource - A resource that causes trouble

Think "red-headed stepchild". The weird ones.

When I was a freshman in HS, a friend and I were playing Betrayal at Krondor late one night, just talking randomly about the stupid things that nerdy freshman talk about. At some point he just turns to me with this weird look on his face and says, completely ernest, "I just realized... I'm literally a red-headed step-child."

There was this long pause and then we both just started laughing.

File under: shit that doesn't happen any more because nobody plays single-player games side-by-side, late at night any longer.

Entities that one doesn't normally think of as a resource, such as an algorithm.

>Highly recommmend GraphQL to anyone.

This sounds a little over-the-top. There are certainly cases where REST would be the better recommendation over GraphQL. I have no idea what your specific requirements were but if building a REST API was a 'monumental effort' then GraphQL was probably a good choice for you. That does not mean that in all cases GraphQL > REST.

In my opinion, I think GraphQL offers enough over REST to be a total replacement. Especially now that the vast majority of clients are mobile, IoT, etc, the benefits of GraphQL make outlier cases rare. This is natural, the GraphQL authors had the benefit of hindsight. I'm not disparaging roy's work, it's a stepping stone. But REST as a design pattern was rarely implemented as he envisioned it. When it was done correctly, you end up with much of the same benefits as GraphQL.

Yes. (Mostly.)

REST semantics are a distraction. The best possible outcome of REST is when developers are encouraged to consider the concept of "idempotency" when they stumble upon the technical definition of the "PUT" verb. Everything else is line noise.

GraphQL has a schema with types. It makes it very straightforward and approachable for all developers to reason about what an API should deliver up-front, and also easy to reason about what is and is not a breaking change. Automatic validation of queries against the schema also saves massive amounts of time in writing validation logic.

There are still things that could be better with GraphQL. The query language is... interesting. The whole thing is still very client-server asymmetric -- look at a client query syntax versus the schema syntax you'll use on the server side for three seconds and you'll immediately and viscerally know what I mean -- and that strikes me as disappointing in this age. It's still very easy for developers to fall into mental quicksand which causes them to make many individual requests even when GraphQL would let them batch things up into one. And so on.

But overall: yes, working on a GraphQL stack is an awesome experience compared to going it alone with REST and JSON and crossed fingers.

> Everything else is line noise.

It's always weird to see this type of thing on HN. How anyone can call the way the web works "line noise" is baffling to me.

Fundamentals vs. changing tech. "Idempotency" as a concept is fundamental. Which header means what for some cache layer is not. REST unavoidably has a lot of things in it that are changing tech. (Nominally, it has a lot of fundamentals in it too, but per the other REST discussion on the homepage, which fundamentals it encompasses seems to vary on a person-by-person basis.)

Most of the REST understanding is basically "how do distributed caches work?".

Thanks, said it better than I could.

>strikes me as disappointing in this age

A good summary

Mostly, NOTE: I'm using python/django/graphene server side and apollo client side.

I love how flexible it is for client developers and because the great client side libraries it helps to eliminate a ton of boiler plate code on the client side.

My biggest complaint has been "lost" exceptions and caching.

Because it's possible for an exception to be thrown server side on one field while the other ones succeed I've been plagued with hard to monitor/find errors. I ended up writing a shim to parse the response in an attempt to get more insight into #errors / fields (this has also been really helpful for monitoring slow queries in new relic since all requests go to the same endpoint which breaks a ton of APM monitoring).

My other issue has been around caching, in apollo there are ways to say "don't use the cache for this request", but it's not to give an object a cache ttl. My app allows users to search for events that are happening near them right now, and I've run into several issues where apollo decided that an event from yesterday should be added to a result. It happened frequently enough even with queries that included times as an argument that I ended up basically implementing a "middleware" between what apollo gives back and the component, which felt really ugly.

you should take a look at the caching policies (fetch policy) apollo client provides as well as the error policy. neither are easy to find

for the error policy, you basically control the behavior (for each client instance or each request) when a request is considered failed. none, ignore and all. [1]

fetch policy allows you immense control over caching. this all depends on what you're doing but in some instances it can even make sense to never cache any requests, depending on how you application is structured. the docs are hard to google for this, but here's a link for you [2]

[1] https://www.apollographql.com/docs/react/features/error-hand...

[2] https://www.apollographql.com/docs/react/api/react-apollo.ht...

So I understand that I can use the cache, not use the cache, or use the cache and always go over network for the entire request but that's all client side. GraphQL in general lacks a good answer for object level caching (so in my case I just filter things out before they make it to the component). It also lacks a solution for the server to indicate to the client how long something should be cached (like you would with cache headers or Etags)

It's just really annoying that I have a ton of queries that return lists of objects, where each object has an easily known expiration time, and there's no way to say, "cache this result but remove the elements that have expired". I can't even say, "cache this list until X time", which would be very easy in a RESTful environment. I know there are extensions to support caching but none of them are universal yet. I'm hopeful that the community will land on an answer for object level cache expiration.

That's great feedback. I'm working on vastly improving exception reporting for the next version of Graphene :)

I just have to chime in and thank you for all the hard work with Graphene! You have almost become a meme at the office, we're looking at the latests RC's at the office and talking about what Syrus might be up to this time :-). As a side note, I'm happy to read about Quiver and I hope it becomes a sustainable business!

> My biggest complaint has been "lost" exceptions and caching.

This is quite annoying because the server never throws a 500 error so the normal logging doens't kick in.

Regarding monitoring, I would recommend to check out Apollo Engine if you haven't already. Basically middleware on the server side that injects error and performance info into the extensions object of the response which is then sent to them and presented in a nice UI.

I’ve played with it a little, and I may resort to creating a proxy node app that reads through to the python app exclusively so I can get better monitoring.

You can implement cache ttls with an Apollo link. See https://github.com/kamilkisiela/apollo-link-maxage for an example.

Nice, I tried to hand roll something like this using link-state and my own renders but it didn't work well for me. I'll definitely check this out.

I've never looked at graphene before (I've stopped spending time learning about each fad) but this really looks like it solves a real problem I've experienced. Definitely going to try it out.

Protip: ignore the docs on the website, use the docs on the github package / in the examples instead. They're much better!

My company has been flirting with graphql.

As a frontend dev, I had a positive experience with one service because the backend was far more willing to add new query options. The much publicized "only get what you ask for" part was largely irrelevant.

I am, however, unsettled at the prospect is losing all the built in network and browser caching for idempotent calls (mostly I'm unsettled because no one else seems to seriously consider the issue - it may end up too small to matter, but I dont trust that anyone else here has honestly evaluated it).

Another poster mentioned the issue with partial errors, which sounds like something else that will not get the upfront attention it deserves, while not being an immediate dealbreaker. Add in to that how to manage deprecation of particular query statements as they can no longer be distinguished as distinct endpoints.

My other concern is how much magic frontend libraries provide. This magic looks great if your app is nothing more than input/output over CRUD calls, but sounds very brittle if your app has client side logic (and while perhaps a webapp should ideally avoid that, other services can also be clients.)

So far I have concerns but not concrete problems, I just worry that we wont be able to confirm the severity until we've already invested and committed, particularly when our initial adopters are so enthusiastic. At the same time I don't want to be the guy unwilling to change and adopt new things.

We found our sweet spot in terms of enabling the flexibility of front-end devs experimenting and defining their own queries while maintaining cacheability. During development, front-end devs use the Graphiql endpoint to play around with the data and figure out exactly what they want. Once that's settled, we turn it into a persisted query that is stored on the server and keyed by a unique ID that the client apps use in production instead of the raw GraphQL payload. You add a small amount of overhead for the coordination to create the persisted queries, but we're considering even building a self service process in the future.

Is it fair to say that your system relies, more or less, upon one server change per client request change/new request type? Doesn't sound like a dealbreaker necessarily, but I thought this was a big part of the problem GraphQL was supposed to solve.

We decided against it. We’re in a java backend and GraphQL in Java with ORM is considerably problematic when trying to create efficient resolvers. We simply ran into one hurdle after another and we were finding ourselves in diminishing returns.

The concept is great, and if you write custom SQL queries for each resolver (if necessary), properly caching things that can be cached, and use the first class citizen programming language (JavaScript), then it seems GraphQL works wonderfully.

Trying to fit it into an existing ORM paradigm with respect to complex sub-collections, lazy loading, and efficient database querying, it just didn’t work out for us.

It's not impossible in Java but it does require a strong grasp of graphql-java's execution model and the DataFetchingEnvironment and associated classes. We accomplish something similar to what you desire when querying our time series database via GraphQL.

The big difference between Java and Javascript when it comes to GraphQL is the amount of noise, tutorials and examples on the Javascript side far outweigh other options right now.

Javascript may be the "happy" path for now but I do hope that we see continued investment in Java/Scala/Ruby/Python/.NET etc as those implementations have the opportunity to be better architected and more performant that the reference JS implementation. The reference implementation uses a very simplistic execution model which for now has been copied almost verbatim into other languages but there is great potential for a proper query planner and optimisation layer at the GraphQL layer.

Disclaimer: I work at MDG on Apollo Engine (the backend of which is Kotlin utilising graphql-java).

Do you have a kotlin graphql library you'd recommend. I just started working on Kotlin with GraphQLJava. I saw graphql-kotlin, but that appears to have been updated like a year ago. I've been using graphql-spqr. Mixing kotlin and java in my projects, to get around the annotation bug.

I do agree on the noise and lack of tutorials. I'm getting ready to present on a GraphQL option. We don't want to use javascript. The tutorials for java are a bit cumbersome.

Btw I'm really liking apollo.

Agreed, there is a ton of potential for query planning an optimization.

These last months, I've been working non-stop in a blazing-fast GraphQL engine that beats the rest by an order of magnitude - https://graphql-quiver.com/

I think we will in time find someone cracking the code and creating a "requests" type library for graphql interaction. Either that or some graphql++ language will make things better. Maybe automatic/declarative DB integration? The concept of graphql is light years ahead of REST, but right now it basically requires server maintainers to implement half-baked query processors and solve problems that have been soloved since the 70s.

Did you try clients like Relay and Apollo?

We migrated most of our backend (written in Clojure) from REST to GraphQL (actually a homemade alternative to GraphQL, but not relevant to this discussion).

It went well, it greatly simplified both our backend and frontend code. The backend code got simpler and more stable because it no longer had to deal with "data packaging". The frontend code became more transparent, because you can now easily read what data gets exchanged, and more decoupled, because different components can independently require the data they care about. One of the biggest benefits has been the degree of independence to evolve both the client and server.

We haven't had the performance issues some people have mentioned (N+1 query etc.), because of the server-side design of our homemade GraphQL engine - in which data resolvers are batching and asynchronous by default, unlike most backend libs which approach this problem more naively. Will open-source that soon.

The biggest limitations I see to GraphQL are:

- it doesn't really have a story for caching. I have some ideas for addressing that, but it hasn't been a problem for us really.

- it repeated the SQL mistake of exposing a query language based on text, not data structures. Now we have to write queries as templates instead of assembling them programmatically. This hurts both application developers and library authors.

- it doesn't really have a story for structured writes.

> it repeated the SQL mistake of exposing a query language based on text, not data structures. Now we have to write queries as templates instead of assembling them programmatically.

Out of curiosity couldn't you just treat queries as a composable AST and "compile" the query text from that?

> Out of curiosity couldn't you just treat queries as a composable AST and "compile" the query text from that?

You could write such a wrapper, but it won't be standard, which means that there will never be a robust library ecosystem and set of practices that arises from that. If GraphQL was to evolve, you also wouldn't be sure your wrapper can express all of GraphQL. Finally, this would add new practical hurdles, like a new dependency, additional runtime cost of compilation, obscured source mapping, etc.

Meanwhile, because GraphQL is just a text-based DSL, most people don't have a clear idea of the _information model_ of GraphQL queries - you have to read the spec to get this idea. If we are to build a robust ecosystem on top of GraphQL, this information model should be evident in the API.

All these problems are present, and very damaging, in SQL.

I tend to write publicly facing APIs so this conversation is colored a little by that. An internal API or Microservice is a different story.

I don't think it is either of. I use both. In the same API. The two are largely compatible.

All REST APIs can be modeled in RPC style APIs and that's no different with GraphQL.

I've done APIs where I have a GraphQL facade in front of REST and REST Facades infront of GraphQL by having all my REST endpoints be two lines

1. Graph QL query

2. Format result as REST

That first like maps to a graphQL query and the second one can be standardized for all your REST endpoints.

I tend to like REST in-front of GraphQL better since it allows for some performance optimizations when you know ahead of time all the data you need to grab.

And since GraphQL can be mapped to classes you could also just skip the GraphQL query compile and use the classes directly from your REST with only a few more lines of code.

Overall I like GraphQL a lot because it allows the frontend to make less round trips to the server and makes it easier to exclude data you don't want (which helps when data is large). Json:API tries to solve some of this with includes, related, and fields but it doesn't quite allow as much expressiveness as GraphQL.

> I tend to write publicly facing APIs so this conversation is colored a little by that

Out of curiosity, how do you secure your public-facing APIs and how do you authenticate/rate-limit users?

For read and search operations I typically rate limit per machine based on probability that machine gets hit using machine signature. And as a secondary (more leniant metric, just IP).

I don't centrally track rates unless a signature comes close to 1/N the limit where N is the number of nodes. At which point I will talk to the other nodes Peer to Peer.

Can still be abused but works pretty well most of the time. It also doesn't work if you have a number of nodes that is approximating your rate limit because if you do, you hit 1/N on request #1.

For that reason I tend to choose pretty lenient rate limits (call it one request a second with bursts in a 5 minute window)

For write I use OAuth2 with bearer tokens being a JWT token with a short expiry. I only need to maintain a blacklist of invalidated tokens for the length of the expiry. Rate limiting would work the same way as reads.

Great info -- thanks!

You're going to get responses from people who have invested a considerable amount of time in something they already had plans for (a "sunk cost") so I'm not sure you will get the sort of information you are after here.

I'll be the voice of someone who is actively implementing GraphQL for the first time.

We were mid-way through our project before realizing GraphQL might be a better fit for our use case so we paused for a week to play around with it and see if we could stand up something inside of our project that made sense.

GraphQL felt very "plug and play" to us. Aside from having to re-work some validation to fit into GraphQL's idea of mutations, we were mostly able to drop our existing models and logic directly in and see it working right away.

Having built very well defined REST APIs (and SOAP before that) for years, the flexibility that GraphQL offers made me feel a bit "uneasy" at first but I have come around to appreciating how much freedom it gives the front-end to only request the data they need.

I'm usually the type to shy away from flashy new doodads and stick with what I know is safe+reliable, especially in an enterprise environment, but as the project continues I'm feeling more and more confident in our choice. I suppose only time will tell though.

Could you share some info about your stack?

What about caching though?

GraphQL has magic record-level caching that I do not understand but somehow appears to sometimes work and sometimes doesn't (but using the default apollo-client, you can't easily turn it off)

Also, many people who have only recently adopted GraphQL, as it is relatively new. These people will not have the long-term experience necessary to assess GraphQL in hindsight.

There is an advantage to making technology decisions behind-the-curve, choosing mature, "battle-tested" technology, even if it means overlooking some known warts, and missing out on what's "hot". We can call this being a "late-adopter". I am most interested in N-years-later perspectives where N is around 3 or more.

Sure, but as always in software that's still better than the speculations of people who have no experience using it but think they can make an informed comparison.

These are taken from some of the other comments in this thread:

    * As a consumer of APIs I vastly prefer REST APIs.
    * Highly recommend GraphQL to anyone.
    * I *love* how flexible it is for client developers.
    * As a manager/business owner, I like GraphQL.
    * How I _wish_ I could nope the hell out of our GraphQL dependency
    * Unnecessary complexity over simple rest calls with no benefits.
    * REST semantics are a distraction
    * my experience is just ok.
What I see here is many different reactions from different people, presumably with different experiences and use cases on their hands. And that's perfectly OK -- REST, GraphQL, XML-RPC, heck even SOAP are just tools, and use whichever works best for your particular situation.

Just because it worked for me doesn't mean it will work great for everyone else; and if it sucks for me doesn't mean it won't be a life-saver for someone else. Those are just tools; use them as you find fit.

I'd like to add that correctly designed resolvers allow you: - to control very easily who can fetch what where it's fetched (permissions) - to fetch nested data when you need it without writing serializers - to help your frontend team find what they are looking for without asking the backend team everytime - mutations are a huge plus when it comes to standardization of your API too

FYI we're using it in production over a django backend (which comes with some drawbacks, since subscriptions == pushed updates are not perfectly implemented) with our react/apollo apps (web and native) and in my opinion the overhead lies surprisingly more in the frontend side (writing data connectors is longer, but way more explicit, than using rest queries returning json)than on the backend (where you just declare resolvers, a thing you don't even need to do in nodejs) and handle permissions.

>to control very easily who can fetch what where it's fetched (permissions)

This is a piece of GraphQL I haven't been able to get my head around. Could you elaborate or point me to a good explanation of how this is implemented? Everything I found when I looked into GraphQL previously was something like "you control access to individual resources in your business layer" but never explained how.

In order for GraphQL to return a result when you ask for field “foo”, you need to define a resolver function for that field. Whatever that function returns will be returned to the client.

Inside that function you can write any code you’d like, including permissions code.

as sgdesign said, you can write your own resolver function (which allows you to retrieve a particular field deep in the nested graph of graphql). It's very easy to create decorators to handle the usual permission cased on the field level.

One important thing to note: performance-wise, both in front-end when used with apollo, and in backend (you need to handle nested queries smartly) a good implementation of a REST api would probably be faster / lighter, but it should be largely enough for any web-app / web-service used by human beings ;)

Citation needed; what is your evidence for such an assertion?

Graphql and apollo add some overhead to your queries, both in your client and your API. And since you're still performing a POST query with each graphql query it's kinda logical to see that pure / well designed REST can outperformed graphql (if you know exactly what you want to return you don't have to run through the graphql schema / node / resolver to get the value your looking for)

GraphQL is great for the frontend, but moving to GraphQL involves both people and tech issues. Common mistakes made when using new technologies are made all over again.

* Watch out for bad implementation of the GraphQL API (this will definitely result in bad performance).

* Design the GraphQL schema that you want the user to see/perceive. Not every object or field in your database needs to be exposed via the API the way it is.

My workplace is currently moving a huge monolith into a bunch of manageable components. Each of these components has its own GraphQL endpoint. Using schema-stitching, these are being stitched together into one endpoint for API users.

As a result of our codebase, we've tried GraphQL in:

* Ruby (graphql-ruby) - WATCHOUT Relay arguments for connection fields are not exposed to the library user. So basically you have to implement your own Relay-compliant stuff if you need access to the pagination arguments from Relay. Also, documentation is broken.

* Python (graphene) - We've had no issues so far. We worked around it.

* Node.js (Apollo GraphQL) - OH MY BUTTERFLIES. So far, this is the ONLY library I have come across that is polished and has plenty of documentation.

* Elixir (Absinthe) - My coworker worked on this part. He did not complain. So I'm assuming he had no issues.

The "Learn * in a day" joke applies to GraphQL. As simple as GraphQL looks for the client-side, it is beast of a job to build a GraphQL backend that is optimized for production.

Servers-side implementation of GraphQL is not very well documented apart from hello-worldly examples. Most of the knowledge found online is about client-side usage.

Due to poor documentation/examples provided, ramping up people with GraphQL is hard. Most first iterations I've had to review were slower than our REST APIs because of unoptimized code. Sitting down for a few minutes solves that problem.

To ramp up people at work place, I ended up having to do this:

* Ask people to use the GitHub v4 API to checkout GraphQL.

* Make them build a GraphQL server for a blog app.

* Dive straight into whatever feature/API they would build.

* Review their work a few dozen times and show them optimization tricks.

My most valuable lesson: When in doubt, dig into the source of these libraries.

my experience is just ok. as someone here puts it, great for frontend devs but bad for backend devs.

if you have db schemas on the backend if using orm, get ready to duplicate them again for graphql.

and on the frontend, get prepared to write out every songle fields you need from the backend. i can imagine it may be brutal for those who have a lot of changes in their schemas.

my conclusion is that, since im a fullstack who does both frontend and backend, i feel myself getting a bit more fatigued than when i was doing rest style api. i find myself wanting rest time to time, esp at times i dont feel like writing out all the fields i need back that i cant remember off top of my head.

> and on the frontend, get prepared to write out every songle fields you need from the backend. i can imagine it may be brutal for those who have a lot of changes in their schemas.

Wouldn’t you need to do this in some form anyway (since those fields would be displayed or used in some way)?

well, it won't be just a simple GET /user/1 anymore. it will be like: user { id name role status age sex address group { id name } etc etc }

Right - what I meant was more like when you get some data you probably want to display it in the UI and would already be doing something like ‘Hi ${user.name}, your age is ${user.age}’. This is affected by schema changes irrespective of the way the API is invoked.

> if you have db schemas on the backend if using orm, get ready to duplicate them again for graphql.

I would consider this a weakness of ORMs, not of GraphQL.

The Vietnam of CS.

We've just started to dabble in GraphQL, and like many others we've seen mixed results.

On the upside, we can construct complex queries than eliminate many consecutive RPCs that you'd end up with in a traditional REST API. At scale this should work wonderfully, greatly reducing the client/server latency for our realtime app.

On the downside, the tooling is still far behind. This is somewhat due to GraphQL being a younger technology so you have to give it some time. OTOH, I feel like you can get things off the ground with REST more quickly. Problems with GraphQL tend to be harder to reconcile due to the debug tooling handicap.

Some of our engineers take a little time ramping up to GraphQL due to its complex nature. This is probably a good thing in the long run though, since it stresses the importance of keeping RPCs to a minimum and eliminates having to sync or batch consecutive RPCs.

Overall I still think it's a win. The tooling should improve over time, and hopefully it will be a first-class citizen in IDEs and libraries soon. Until then, you've got to be prepared to muscle though it.

Nick Schrock here, one of the GraphQL co-creators. I agree with a lot of the criticism in terms of the difficulty of implementing GraphQL backends. I think there's a big opportunity for folks to build vertically integrated toolkits that deal with N+1 issues, integrate DataLoader natively and so forth. Good versions of these would deal with a lot of issues described here in greenfield GraphQL backends. I talked about this at the GraphQL Europe keynote last month (https://www.youtube.com/watch?v=zMa8rfXI6MM).

Current greenfield implements are typically stacked on ORMs like Django and RoR, and the impedance mismatch is real. Personally I abide by the dictum that ORMS are "Vietnam of computer programming" and should be avoided at all costs for anything that will grow beyond a small app. GraphQL was not originally implemented on top of an ORM, but instead an object model built on key-value + edge store internal to Facebook.

In terms of other criticisms in this thread:

1) Exceptions: The default behavior in graphql-js (mimicked in other language implementations) of swallowing native exceptions by default was probably a mistake in hindsight. Whenever I've played with GraphQL using different toolsets the first thing I change is to add a wrapper function which checks for errors and then rethrows the initial exception that caused the GraphQL error for use in testing and CI/CD contexts.

2. Caching: Personally I've always been confused about the concern with leveraging HTTP-level caching. While a clever hack, with any real app with any sort of even mildly dynamic behavior you don't want to do this. Staleness will be interpreted, rightly, as bugs by your users. If you want to replicate the behavior the most straightforward way would be to use persisted queries (described here https://blog.apollographql.com/persisted-graphql-queries-wit...) combined with HTTP GETs. With persisted queries you can encode the entire request in the query string, which should get you the HTTP-level caching you want.

3. Docs: Quite confused about this one. While particular implementations of GraphQL can be problematic the documentation of the core language (which I am not responsible for) is superb. See http://graphql.org/ and https://www.howtographql.com/.

My largest criticism of GraphQL is that it doesnt go far enough. That is to say that I want my components to declare their data dependencies, but not all data comes from a server. Some data is computed, other data comes from different servers, and some other data is locally generated. For example, facebook has some user information that I want to add to my User object; in my case, theres no reason for the component to know that its data is actually coming from multiple sources but I have no simple way of coding that.

Another issue: there currently are a lot of servers that are REST and it would be really nice if there were ways of incorporating them in the GraphQL schema. One of the most useful features to add here would be ways of linking separate REST objects to each other in the GraphQL schema. This would help adoption quite a bit.

I'd check out Apollo for these use cases. They allow a client GraphQL schema to be sourced from a bunch of different places, including (I believe) local-only state, REST APIs, and others. See https://www.apollographql.com/docs/link/links/rest.html

Not at all (we rolled our own back end). The added complexity of building queries to join our API ended up creating such a quagmire of SQL that any dev coming in will basically have to learn our own custom ORM.

Sounds like you had a few problems.

It makes many things easier and many things harder. The lack of really good backend libraries/frameworks outside of NodeJS is the most concerning thing.

Also; debugging and monitoring GraphQL APIs sucks. Considerations:

- Any subfield of a query can throw an error, but the rest of the fields can succeed, because GraphQL frameworks are allowed to run each field resolver asynchronously.

- Because of this, any GraphQL query is capable of returning multiple errors.

- Rate limiting is exceedingly difficult due to nested resolvers. I've seen solutions which involve annotating your schemas with "cost" numbers, and only allowing each query to run up to a maximum "cost" before failing by dynamically adding the costs of each field they request. Traditional rate limiting doesn't work.

- Traditional APM platforms also don't work. Prepare to adopt Apollo Engine and pay them $600/month on top of the money you're already paying New Relic or Datadog.

I’ve only heard positives about Absinthe for Elixir fwiw.

The three biggest frameworks are definitely Apollo (JS), Graphene (Python), and Absinthe (Elixir).

The common thread between those is that they're all languages with weak type systems, which is a theme in GraphQL land. Strongly typed languages tend to have usability and safety issues with GraphQL; you end up needing codegen or bypassing type safety to make it work.

Our front end engineers were extremely happy. Me personally, as a backend engineer, not as much, but it isn't too bad.

There are some quirks (error handling), performance issues (e.g. fixing n+1 queries) and DOS concerns, but again, it isn't all that bad.

(we're using rails/graphql-ruby on backend | react/relay on frontend)

I'd agree with this sentiment, on the backend, it isn't really easier and it's not really harder either. You end up dealing with different problems than with REST (as noted, different types of performance problems, more issues with error handling have been my biggest two).

The frontend developers see the biggest improvements, they have one endpoint to worry about and using a library like Apollo makes things really nice.

Overall, I think the improvements are worth it, especially when working on internal APIs.

I really wish everyone would just move to GRPC and be done with it.

GraphQL just feels too tied to the datastore on the back end to be generally useful. REST/Swagger is hugely overcomplicated for the basic REST premise of moving objects back and forth.

GRPC is what REST should have been. Ship objects back and forth between multiple languages with minimum fuss.

100% agree, RPC was good enough for our forefathers ;) REST and graphQL are both terrible (there, it had to be said).

There's no reason someone who wants an expressive query language like graphQL provides couldn't send that query over gRPC, in which case all the benefits are moot, you can do both. But let the rest of us just make regular RPC calls.

You don't have to "move" to reap GraphQL's benefits – you can just add a GraphQL layer.

I'm backend Systems Architect at a big publishing company, and my current primary project is an aggregating caching GraphQL proxy for our REST microservices.

Our front ends were making too many calls to the REST APIs, so we went overboard embedding related resources – and now they're getting too much unneeded data back, and cache invalidation is a nightmare. Sounds familiar, probably!

So we're building a GraphQL service that stitches those REST APIs together to let the caller request exactly the fields they need, from any API's resource. By caching individual resources, rather than nested multi-resource serializations, we can invalidate easily by UUID on change events – so cache TTLs can be long – and the GraphQL API's field resolvers can assemble complex responses with a few fast Redis MGETs, which are batched by DataLoaders.

This also gives us a place to centralize business logic, rather than having each front end service reimplement field formatting, resource transforms, &c. Since the REST APIs remain available as the source of truth, existing services can migrate to the GraphQL proxy at their own pace, which we hope will be an easy sell since it's so much faster.

I'm really curious to hear from people dealing with massive JSON (or XML) responses via REST. This is where REST really falls apart with complexity that is much worse than trying to learn the complexity of GraphQL.

See: https://brandur.org/graphql

"In GraphQL, every contract is explicit and observable. Providers can use something like a canonical log line to get perfect insight into the fields that are in use for every request, and use that information to make decisions around product development, API changes, and retirement."

You simply cannot know with a REST API what to remove if the decisions on what to return for a request are largely made by the server side team. So, you end up with massive, clunky and confusing responses. GraphQL forces the client side users to understand how to use it, and because of that you get so much more.


I had the chance to participate in moving the client-server architecture from REST to GraphQL when I worked for two companies as a contractor. They didn't regret it so far. There a various advantages (and disadvantages) [0] using GraphQL.

But these are only the direct impacts of using GraphQL. In the case of these two applications, they had a React client application managed by a GraphQL client library. All the state management done with Redux/MobX was reduced to a minimum, because most of the managed state was remote data from an API. Now the GraphQL client was able to take over, leaving only the local UI state for Redux/MobX. Often it is even possible to remove these state management libraries altogether.

- [0] https://www.robinwieruch.de/why-graphql-advantages-disadvant...

Yep, best decision. Both from a back-end experience (caching, performance, type safety) and front-end experience (flexible querying, great documentation, tracing, etc.)

Yes, it's generally quite a bit nicer. It's not as nice as I hoped, but it's better. The old APIs are JSONAPI-style.

The main disappointment is that input types are not nearly as expressive as output types.

I'm only familiar with REST and not GraphQL, but this sounds interesting. Just from the name GraphQL sounds like a query language -- how can one do operations that change state (i.e., analogous to PUT/POST) rather than just query it? Thanks.

It’s a QL in the sense that it accepts a client specified schema, that looks how the client wants the data to be returned.

How does it mutate then? SQL can mutate, I’m not sure why this would be any different. Basically, like SQL’s UPDATE, and INSERT, GQL can have “methods” associated with parts of the schema.

The name is a little misleading in that regard. GraphQL has mutations that tell the server to change data rather than just return a query.

Thanks. That helps!

GraphQL really shines for querying data, but if you do a lot of writing of data, the API starts to feel like a clumsy version of RPC. I ran into this trying to apply GraphQL to a business use case that required lots of user input. Designing mutations on both the client and the backend is pretty tedious, especially when it comes to handling the response of the mutation. For example, there's no straight forward way to handle errors like what's commonly done with REST using HTTP error codes.

Who says GraphQL is not RESTful, though?

You have resources represented as a graph and you're asking for the state (in some representation -- probably JSON, but maybe protobuf, etc), just in a more expressive and deeper way than a simple HTTP Get with a single URL can express.

Nothing about REST ever said that a resource locator had to be a URL, or that a single resource can't represent a collection of resources.

That's all GraphQL does, IMO. It lets you more flexibly and expressively ask for stat of a collection of resources.

Not someone who's moved to GraphQL, but still curious how people do load testing and such with GraphQL APIs. Do you log which queries are made, and then plan around that?

Is it common yet for public-facing APIs to be implemented with GraphQL? Is it feasible if you're trying to make it accessible and easy to use for casual developers?

Another django/graphene shop here. I'm really enjoying the transition. My one complaint is with deep queries. If the query asks for a very nested objects, the amount of DB fetches can become large and then you get a performance hit. With REST you generally know the performance of your queries since the schema is static. How do people solve the n + 1 issue here?

Currently evaluating using GraphQL. I have used it for a smaller endpoint [0]. I like the server side as I feel it lets me express the purpose of the API really well.

I have been using straight HTTP requests to query the GraphQL endpoint. Now, wanting to do something more complex, I am disappointed in the state of GraphQL clients. Everyone says use Apollo but it has some serious issues (e.g. caching [1]). Relay seems over-kill/restrictive and none of the other alternatives seem complete. I feel like maybe my life will be a bit easier if I just stick with REST especially since I am proxying an exiting API.

[0]: https://github.com/monostable/kitspace-partinfo

[1]: https://github.com/apollographql/apollo-client/issues/3452

I'm holding off until I don't have to define types in 3 places:

1. TypeScript

2. Document database

3. GraphQL

I know there are things that do two of these. I want all three.

There are stacks that can give you delicious sharing here. Clojure/clojurescript with Lacinia[0] for example could lead to something pretty terrific, as Clojure DB access is all immutable structs instead of activerecord-style stuff.

[0]: https://github.com/walmartlabs/lacinia

graphql-tag [1] is very close to automatic typing for TypeScript. It lets you import a `x.gql` file (or whatever suffix) that contains query text directly, and translates the contained queries into named input/output types and parsed query documents based on your schema.

It's great but it's not quite perfect for me: I'd prefer if it produced a function that accepts your query executor and inputs, and returns an observable of the result, and it doesn't work very smoothly with angular (especially 5+).

[1]: https://github.com/apollographql/graphql-tag

which two?

Usually 2 and 3.

Sometimes 1 and 3 (see http://avant.engineering/graphql-and-typescript/)

We have, so far, only made a GraphQL endpoint for internal use. It has been awesome for quickly finding things in a medium-complexity database schema (30 or so tables, bunch of different relationships). It has become my preferred way to quickly find things in the database.

That being said, there are cases where a join between ad hoc subqueries is the best way, and GraphQL doesn't really offer a way to do that (though I don't see why it wouldn't be possible). E.g. arbitrarily combining two GraphQL queries that return lists where some field in one is equal to some field in another.

But in terms of replacing REST, where you have to do all of that anyway, it is far and away the better option (for ad hoc querying, at least).

> a join between ad hoc subqueries is the best way, and GraphQL doesn't really offer a way to do that

I think you're supposed to create a "virtual" field on the left-hand object that represents a collection of the right-hand type of objects. The field can be parameterized if your join needs extra information. If you want pagination, the virtual field returns an intermediary object describing the cursor (sort order, offset).

No. It cold have been that Relay the JS graphql implementation we used was too heavy handed, but our team never fully grokked what was going on. It wound up introducing a bunch of friction into our workflow, and we ultimately wound up removing it.

I just wanted to say thanks to everyone that commented! We've been exploring if GraphQL was something worth trying out and it seems like it's worthy of a trial on some of our lower impact applications. Great thread here.

Yes, our team has migrated to GraphQL and it has been great, both on the front and backend.

We have a Django app on the backend and have used Django REST Framework for our REST API. That was a great experience, but had some limitations. We've written our new GraphQL API with the help of Graphene and it has been awesome. Everything is very declarative and exploring the API schema couldn't be easier using GraphiQL.

Implementing it on the frontend has been great too. Yes, you do have to specify what fields you want, but having complete control is worth the extra boilerplate. I'd recommend it heartily.

I really wanted to like GraphQL but ultimately decided against it. It's one of those things that seems great in theory but when you actually start working with it you see all the warts.

The client libraries are too complex, there's too much magic going on, debugging is a pain, there's not a lot of native support in browsers and mobile devices. I don't want to step into the code of a client library I didn't write to figure out some simple query issue. It's no longer a simple matter of just an HTTP request anymore. It's not a pragmatic choice.

The problem with REST is largely the people who look at it blindly as "thing that makes the CRUD go" instead of as a methodology of how to effectively use HTTP verbs and designing URIs that make sense.

So I guess my question is, why would I use GraphQL over, say, the Swagger tool suite? Swagger and the OpenAPI spec defines a way of doing REST that best fits both what Roy Fielding meant for REST but also fits IDE and tool automation systems.

There are HUGE benefits for our client developers with GraphQL. Like being able to select as much or as little as a particular view/component/whatever requires and no more. GraphQL from a front end or mobile POV makes your api more like a data store that it can interact with and query for it's needs, which makes app UI work much nicer.

There are also a ton of maintenance benefits, like for example if you add a field in GraphQL that's perfectly fine to not change the version because that break any existing calls, which is not always true with REST.

> There are HUGE benefits for our client developers with GraphQL. Like being able to select as much or as little as a particular view/component/whatever requires and no more.

That's not a unique benefit of GraphQL. Any HTTP based API can do this, regardless of how it's designed.

Some specifications even have it baked in (See: jsonapi.org).

> GraphQL from a front end or mobile POV makes your api more like a data store that it can interact with and query for it's needs, which makes app UI work much nicer.

Again, not unique to GraphQL.

> There are also a ton of maintenance benefits, like for example if you add a field in GraphQL that's perfectly fine to not change the version because that break any existing calls, which is not always true with REST.


> There are also a ton of maintenance benefits, like for example if you add a field in GraphQL that's perfectly fine to not change the version because that break any existing calls, which is not always true with REST.

Assuming you are using some from of Semantic Versioning, adding a field to a REST API should never break existing clients. Did you mean "removing" a field?

Yes exception for File uploads. There isn't a spec and not all apollo clients yet support it.

Hopefully, it'll be standardized soon.

For writing React Apps GraphQL is wonderful!

For a mature product, no. We closely looked at engineering time and realized we'd never make our investment back.

For a new product, yes (so far).

Absolutely. REST APIs are passive and GraphQL APIs are the vice versa, very active and progressive in nature, GraphQL also comes with a simpler model to process real time data (subscriptions). However, it is relatively hard to scale GraphQL (especially if you had Websocket based subscriptions) while REST can be backed by traditional HTTP proxy.

I had hoped that GraphQL would come along with a solution for 'live queries' - alas that seems to still be on the drawing board. Really looking forward to when 'send me a list, and any changes to that list, and merge them together' is a problem I don't have to put any thought into solving.

Could you all tell about your stack, too ? Do we have to use NodeJS in order to get GraphQL working properly ? I'm aware of the libraries available for other languages, but NodeJS seems to be the only platform with proper support. I wonder if there is any Go developers building servers with GraphQL.

Python (Django) support is good (appart from subscription), but not as out of the box as NodeJS indeed

Much prefer REST assuming nested documents are aggregated by default into the parent. If it's just an HTTP RDB client like some can be then not so much. Less data on wire doesn't matter for server to server communication. I can see it for APIs whose sole purpose is low bandwidth client apps.


The productivity gains from having a client-server interface that is dead simple to reason about more than made up for the initial investment cost. We can refactor API schemas without worrying about breaking existing client code. We also saw performance gains from saving on network round trips.

Almost all developers/team who actually tried out and used GraphQL, will say the same thing: You never want to go back.

Most of the arguments against GraphQL that I'm reading here, seem to actually be misconceptions (or the GraphQL ecosystem not being far enough yet).

I disagree. Some points are really valid. It adds complexity to your system, especially at the server. It is a costs/benefits thing, not only benefits. The benefits are minimal for simple CRUD-like applications.

I've been building a GraphQL only product for 1.5 years now. It was also the best scenario for graphQL because we were also rewriting an entire app from scratch.

For some background info this app has a huge feature set, low traffic, B2B, and at most 2 developers working on this app for about 1 year.

For the technical side I chose typescript from the back end and front end (Angular). This mean I could use the same typings from our ORM mongoose.

Very early on I was able to pick tooling that ending up being cutting edge compared to any other tool in any language, namely the tools where graphql-compose, grapqhl-compose-mongoose, and graphql-code-generator. In addition GraphQL lent itself pretty well to dependency injection on the server side.

It took about 6 months to get fully proficient, but it was completely worth it being able to use the same types for the front end and back end. GraphQL even made it possible to bootstrap types from mongoose into the server. A lot of time when dealing with 3rd party services I could copy the docs and turn it into types for the resolvers.

The biggest single benefit of it all was side stepping the middleware of Express. I didn't realize how awful express middleware are, because you can't really control the middleware chain very well. But the library with graphql-compose made chaining resolvers extremely precise, which was important.

Overall I love working with GraphQL but I realize that I probably could have done just as well with rest if the backend has decent middleware chaining i.e not express. For performance there is lots of options that are often built in like client caching, batched queries, whitelist queries, and persisted queries. It worked pretty well in a PWA too.



- types everywhere!

- following the logic of any resolver is usually pretty straightforward.

- true middleware

- learned how to make good rest frameworks in the future.


- hugely steep learning curve

my API uses a combination of REST and GraphQL, new api calls are written mostly in GraphQL unless its something like a batch job, or querying, modifying the database, Rest works better for those type of tasks in my experience.

So far so good, especially on the front-end. We love the queries, the caching downside don't bother us too much, but mutations make me a little nervous (conceptually, ecumenically, grammatically...).

On the server side, it is a mixed bag. Defining a new type system on top of our existing one, and on top of the Avro schemas we have for inter-service communication, it feels... a bit much. Each layer has a purpose so I can't advocate for removal, but it does feel silly at times. Also the documentation on pagination feels incomplete to the point of confusion, and depending on how you search for answers, you might find blog posts that add confusion rather than clarity.

Some additional friction came from the specific version of the Ruby server library we jumped in at. They went a little overboard in the amount of metaprogramming used by the DSL, and I say that as someone who enjoys metaprogramming shenanigans for fun on a Friday night! But it took a long time to understand why adding some custom behavior, in a way that seemed utterly necessary, was not possible (e.g. defining methods inside the DSL's type block did not class_exec them into existence on the resulting class in any reasonable way). They have since released a new version whose main API is class-based, which is much more clear, enables everything I desire, and I am eagerly awaiting next month when we get to upgrade to that and refactor. Thanks to the maintainers for recognizing this was needed, and for doing the heavy lifting to pull it off!

What I enjoy about GraphQL conceptually is that it dodges a certain square-peg-round-hole issue I occasionally saw with REST, but in the same way that leaves me nervous that we'll end up back at a land of inconsistent RPC spaghetti. In an API where an iron-fisted "RESTful Actions Only!" rule is enforced, many domains find just those one or two actions that feel like they are a better fit for their own unique verb, a custom-named action. The rhetorical-and-ontological escape hatch for REST's limited verb set is to have an abstract "resource" that only has, say, a POST option. You can only "create" an "/recalculate_contacts" resource, but never GET, DELETE and so forth. Or you PUT to "/contacts/recalculate", or some other silliness that just feels too clever by half.

Those of us who were scarred by inconsistent, poorly-thought-out RPC spaghetti over the years cling to REST because it brought peace to our nightmares via simple rules that almost fit everything, and as a bonus, really gel with HTTP semantics. Because GraphQL's Mutations really don't specify much of anything, I immediately assumed it would descend into a sort of API moral decadence, but that has not borne out so far.

If anything, I am heartened to see that newer (though not necessarily younger!) developers, who grew up in a world where REST was all they knew, do not have this temptation to go off the deep end with action names, or to craft RPC-esque systems with ill-considered boundaries of behavior. For the most part they want to name things ala REST, because that's what they are used to, except once in a while. So that is what our company goes with in terms of mutations: Wherever possible, name it for the CRUD action it represents. Only use a more unique/one-off verb if it is truly justified.

Eh... I'm torn on it and could write at length about this, but here's my off-the-cuff thoughts. For context, my co-founder and I are quite active in the GraphQL community. If you've ever used apollo-cache-persist[1] or graphql-crunch[2], we authored those libraries. Our startup, if you're curious, is a social podcasting app written in react-native (https://banter.fm). We learned a lot of GraphQL lessons building it.

The tl;dr; is that graphql gives you a lot of flexibility and typed schemas are nice, but it doesn't come for free and I miss the tooling around http.


- Typed schema

- Custom queries retrieve all the data the client needs in one request

- Really easy to implement, both server-side and client-side

- Server-side, it's trivial to have any field resolved in any way you want (redis, memcache, postgres, some random service)

- Easy and arbitrary mutations. There's no pontificating over what verbs are the proper ones to us.

- If you use react, the community and ecosystem often assumes you're using GraphQL, so it may make sense to use graphql just so you don't swim against the current.


- The payloads can quickly become huge because there is often a ton of duplication in a responses (depending on your query patterns). See this example on the SWAPI demo: https://bit.ly/2uOFZBP. The result is 1MB of JSON, ~97% of which is data that exists somewhere else in the response already.

- Refactoring types is often impossible to do in a backwards-compatible way, even if the shape of the data is the same.

- You don't know what data you'll need in advance, so you're basically doing all of your joins by going back and forth between the api resolver and your data sources (this can be alleviated with persistent queries, but those come with their own set of issues). A typical query to hydrate a response for a user's feed in our app requests ~1,100 objects. After caching and consolidating queries into multi-gets, it translates to about 50 distinct DB queries.

- Tooling: Working at the HTTP level simply has better tooling and tons of infrastructure around caching and serving content (varnish, nginx, etc...)

We found that graphql payloads were so large that older mobile phones were spending significant time parsing them. We created graphql-crunch to de-duplicate responses before sending them over the wire. This led to nice perf improvements on mobile platforms. It also gave us referential equality when persisting the results to cache, allowing us to reduce a lot of work client side.

If you're going to use GraphQL, embrace the javascript ecosystem. Also use Apollo[3]. Also use DataLoader[4]. Roughly 40% of our queries get resolved for "free" by data loader.

If I were to do it again, I'd at least prototype a REST-api with resources designed specifically for http cache-ability (that is, break out session-specific resources/attributes vs shared resources) and see if HTTP/2 multiplexing + nginx caching + etags results in a good client experience. But I also mostly work on the backend while my co-founder mostly works on the frontend, so we have different desires and constraints. Ideally, as few requests as possible would make it to code that I wrote. With GraphQL that's nearly impossible.

[1] https://blog.apollographql.com/announcing-apollo-cache-persi... [2] https://github.com/banterfm/graphql-crunch [3] https://www.apollographql.com/server [4] https://github.com/facebook/dataloader

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact