Skip to main content

GraphQL Trusted Documents

GraphQLConf 2023 was an absolute delight! I finally met many of the people that I've worked with at the GraphQL Working Group over the past few years, and they're even nicer in person! And the attendees were delightful; it was really interesting hearing about how they use GraphQL.

My biggest takeaway from the first day of the conference was that almost everyone should be protecting their GraphQL endpoints with an allowlist, but almost no-one is!

Who should be using a GraphQL allowlist?

Anyone who

  1. exposes their GraphQL endpoint to the internet, and
  2. doesn't intend their GraphQL API to be consumed by third parties.

This is anyone who is using GraphQL to power their own websites, mobile apps and desktop apps but isn't deliberately exposing their API for others to use. Those of you who this applies to (and that is the vast majority of GraphQL users!) should be using an allowlist so that only GraphQL operations that your own developers write can be executed against your GraphQL schema.

Adopting a GraphQL allowlist significantly decreases the attack surface of your GraphQL API since only operations that your developers have written can be executed. This technique has been used within Facebook since before GraphQL was open sourced, it's very much a best practice if you meet the criteria above!

Why isn't everyone using GraphQL allowlists already?

Many haven't heard of the technique. Of those who have, many knew that they should be doing it, but failed to find resources on the "how," or expected it to be a lot of work. Another issue was people confusing the allowlist technique known as "persisted queries" (aka "stored operations", "persisted documents", and various other names) for the bandwidth-saving technique "automatic persisted queries (APQ)".

Worst of all, some people felt they were already protecting their endpoints by disabling query introspection, when in reality there are so many ways for an attacker to work around that: extracting hints from error messages, sniffing network traffic, and fuzzing field names to name just a few. At best, disabling introspection gives you security through obscurity.

In my opinion, if you're disabling introspection then you're doing it wrong; you should instead be using an operation allowlist such as "trusted documents" to prevent untrusted operations from running against your API.

What is a "trusted document"?

In GraphQL, an executable document is a text string that consists of one or more query, mutation or subscription operations and their associated fragments using the GraphQL language. People commonly refer to them as "queries", but that term is a little ambiguous — "executable document" is the precise term.

A trusted document is an executable document, identified via a unique identifier (typically a hash), that you've told the server to trust. In most cases, a trusted document would be written by the developers of your web, mobile and/or desktop apps.

Yes, a "trusted document" is an instance of what we've traditionally called a "persisted query" (or persisted document/stored operation/etc); but specifically it is one the server can trust (typically because it was written by your developers) and thus can be used to form an allowlist.

I hope that the entire GraphQL ecosystem can move towards using the term "trusted document" when referring to this concept. It's much more obvious what the term "trusted document" implies, and it clearly differentiates this use from "automatic persisted queries" (a bandwidth optimization), and "registered documents" (an untrusted allowlist, requiring greater run-time scrutiny).

Very much related, I recently wrote up a specification for Persisted Documents which can be used to implement trusted documents (or automatic persisted queries).

How do I add trusted documents to my stack?

If you already use code generation with your GraphQL clients (e.g. for type safety) then it's relatively easy. When you build your application:

  1. Have the code generator write out the document(s) that your client is using,
  2. Generate a hash for each of these documents using SHA256, and
  3. Have your server store into a trusted key-value store the GraphQL document as the value and the SHA256 hash as the key.

When the client issues a request to the GraphQL endpoint, it should replace the query parameter with a documentId parameter which is sha256: followed by the SHA256 hash of your document.

When the server receives a request, it should look for this documentId. If there is no documentId in the request, it should raise an exception* and stop processing the request. Otherwise, it should look up the GraphQL document for this documentId in the key-value store, and continue executing the request as if this were the query the client submitted all along.

* If you're doing this for existing GraphQL APIs then you may wish to capture the hashes of all documents in use for the next month or so, and explicitly allow these through to avoid breaking existing clients.

That's really all there is to it. Choosing what to use as a key-value store is entirely up to you; but here's a couple of ideas:

  • If you have a monorepo for your server and client(s), you could store the operations as .trusted_documents/<hash>.graphql into the git repository; this will even help you know when and why the given document was generated.
  • Otherwise, maybe from CI, your client build process should issue the queries that are needed (and their hashes) to an authenticated endpoint on the server. The server should then store these wherever it finds convenient: a database, a persistent key-value store service (e.g. Redis), or maybe an external service like DynoDB or S3.

Do trusted documents have more benefits?

Besides security, you mean? Well, as it happens, yes!

Trusted documents can help reduce network bandwidth because you don't need to send the (rather long, at times) GraphQL documents from the client to the server each time — just a short hash instead.

If you set your server up such that it accepts GraphQL queries (but NOT mutations!) via GET requests, you can easily make your queries HTTP cacheable: use a dedicated URL for each trusted document/operationName combo (e.g.<hash>/<operationName>) and set the relevant caching headers (don't forget to use Vary if you have your client send variables via headers!) and voila! You could even combine this with a content delivery network to get caching on the edge; though this is quite coarse whole-response caching. (For a more powerful take on GraphQL caching at the edge, check out my sponsor Stellate's partial query caching — it looks fantastic!)

One huge benefit of trusted documents that's not talked about enough is that they give you a great insight into exactly which fields are used, and by which clients. Want to remove a field from your GraphQL API, but you're not sure it's safe to do so? Simply remove it and then validate all of your trusted documents against the new GraphQL API — if the validations pass then you know it's safe to remove.

Are trusted documents a silver bullet?

It might seem at first that with persisted operations there's no need for the server to:

  • disable introspection
  • apply depth limits
  • apply pagination limits
  • perform query cost analysis

And you're right; those needs are significantly diminished! But you still need to be careful about the queries you write. Though an attacker can no longer issue arbitrary queries against your GraphQL API, they can still take the queries you already have and issue them with their own carefully crafted inputs.

Each of the above concerns still exists, but now it applies to just the trusted documents that your developers are writing, rather than runtime checks against arbitrary operations your server is receiving. You should check your documents before you persist them to ensure that they meet your requirements for safety; this is a one-time cost at document persistence time rather than a cost incurred for every request.

You should also train your developers on the writing of "safe" operations. Imagine you trusted a document such as:

query TopUsers($limit: Int! = 10) {
topUsers(first: $limit) {

An attacker could issue this query with a $limit of 2147483647 and now your server is on the hook to return up to 2 billion results. Teaching your developers to hardcode pagination limits into the query itself is one solution to this, another is to maintain limits in the server and throw out requests that clearly exceed sensible bounds.

Similarly if you have large input object trees (for example "filter" objects) then it's best to specify as much as you can into the query itself, and make the variables only for the "leaves" - this way an attacker can't make a punishingly complex filter for your server to execute.

Share the news of trusted documents today!

"Persisted queries" has been an imprecise technique that is widely adopted; on the client side Relay has a specification for their Persisted Queries, and Apollo also has their own. The Guild (another of the companies sponsoring my open source work) also specifies persisted operations for GraphQL Yoga and Valu Digital have a plugin for GraphQL Code Generator to generate the persisted query IDs for you.

With the introduction of a vendor agnostic GraphQL Foundation-hosted specification for persisted documents as part of the GraphQL-over-HTTP project, I aim to work with the maintainers of these projects to maximize compatibility and ease adoption of trusted documents across the entire ecosystem.

I'm a community-funded open source developer; if you would like to support the work I'm doing please consider becoming a sponsor for as little or as much as you can afford each month. I couldn't do what I do without the support of my sponsors. Thank you!