bits by luke

GraphQL Tradeoffs

Last year I had the opportunity to work with GraphQL for about 5 months on a client project. There has finally been enough distance and hammock time to put together some thoughts on the matter. Plenty of resources exist to explain the specification, the tooling, and the proposed benefits of a GraphQL API. There are a lot fewer resources which talk about tradeoffs and pain points, so I’m going to provide some more pointed critiques.

What you get


The spec requires everything to have a schema. This is great. It enables powerful tooling such as playground and graphiql. If you’re appropriately evolving your schema (and not breaking things), you’re really going to appreciate the wins of GraphQL schemas.

Bytes over the wire

It’s very easy to optimize payload size. This is one of the major selling points of GraphQL. A client requests only what it needs, and no excess data is transferred over the wire.

Number of requests

The number of HTTP requests can be dramatically reduced. Retrieving resources and their relationships (graphs!) is easy, and enables clients to, in general, make far fewer round trips to the server than traditional HTTP APIs.

What you give up


There are fundamental requirements to any web-based application such as safe/unsafe/idempotent methods, authentication schemes, response codes, header controls, and error semantics. While there are definitely warts in HTTP, there is enormous value in using a specification which has been carefully designed and hardened (by someone else) for over 20 years. Currently the spec for GraphQL doesn’t cover any of the aforementioned topics, which places the design burden on each individual application. It is extremely difficult and costly to develop resilient specifications, and that time is probably better spent solving the actual business problem at hand.


While technically part of the HTTP specification (above), caching is a fundamental pillar of the web and deserves a separate mention. GraphQL pushes caching out of the network layer and into the application layer. We lose the long lever of intermediaries (CDNs, reverse proxies, gateways, etc.) and features like Etag, Is-Modified-Since, Last-Modified, or Cache-Control. Each application has to develop a custom caching strategy, which is no small feat. You then have to find intermediaries that speak this new caching language, or write (and maintain, and deploy) some yourself.


A query language over a datastore is not an interface. The whole point of an API is to empower clients to achieve outcomes by having a sensible conversation about a particular domain. GraphQL is akin to RPC-over-HTTP, with an unbounded set of ad-hoc operations acting on domain entities. Simply allowing clients to modify payloads without making changes to the server is not a meaningful decoupling mechanism. When entities evolve, or business processes change, how does the coupling of N operations to Y entities hold up over time? Are those conversations still sensible?


By exposing any node or edge to a client, you have unleashed unlimited-complexity queries. There are some boxing strategies for limiting depth and width, but this is another thing which must be maintained and communicated to clients. Optimizing hot-path queries just became a lot more daunting. It’s no longer a known trade-off of one set of queries instead of another set because you have an unbounded set of unknown queries.

Another approach

If you are considering making the above tradeoffs and adopting GraphQL, I would first urge you to take a closer look at some hypermedia implementations. Hypermedia in general provides excellent specification, grants meaningful client-server decoupling (a closed set of operations on an open set of resources), offers optimization levers (a known set of resources with known query semantics), and utilizes fundamental HTTP semantics (including network-level caching).

In particular, json:api provides a powerful specification that covers both serialization formats and expected HTTP behaviors. Compound documents are a mechanism for reducing the number of HTTP requests. Likewise, sparse fieldsets provide comparable flexibility to GraphQL’s payload optimizations.