Atlas API

(Public API Platform)

A public GraphQL and REST platform with keys, docs and limits developers actually enjoy.

2B requests / month
42ms p99 latency
600+ developers

2024 · Node.js · GraphQL · PostgreSQL · Kubernetes

atlasapi.app

Problem

(what hurt)

Partners scraped HTML to get their own data, which broke on every redesign. Internal endpoints were undocumented and unauthenticated.

Project goal: Partners were scraping the dashboard because there was no API. The goal was a public GraphQL and REST platform with keys, docs and fair limits.

Architecture

(how it fits together)

A GraphQL gateway sits in front of existing services with per-request DataLoaders. REST endpoints are generated from the same schema. Keys, quotas and usage metering run at the edge before requests reach the cluster.

Tech decisions: A GraphQL gateway with DataLoader batching to kill N+1 queries, cursor pagination, token-bucket rate limiting at the edge, and generated TypeScript SDKs. Runs on Kubernetes.

Node.js
GraphQL
PostgreSQL
Kubernetes

(architecture)

graphql/resolvers/project.ts

export const Project = {
  // batched: one query per request, not one per row
  owner: (project, _args, { loaders }) =>
    loaders.user.load(project.ownerId),

  deployments: async (project, { first = 20, after }, { db }) => {
    const rows = await db.deployment.findMany({
      where: { projectId: project.id },
      take: first + 1,
      cursor: after ? { id: decode(after) } : undefined,
    });
    return toConnection(rows, first);
  },
};export const Project = {
  // batched: one query per request, not one per row
  owner: (project, _args, { loaders }) =>
    loaders.user.load(project.ownerId),

  deployments: async (project, { first = 20, after }, { db }) => {
    const rows = await db.deployment.findMany({
      where: { projectId: project.id },
      take: first + 1,
      cursor: after ? { id: decode(after) } : undefined,
    });
    return toConnection(rows, first);
  },
};export const Project = {
  // batched: one query per request, not one per row
  owner: (project, _args, { loaders }) =>
    loaders.user.load(project.ownerId),

  deployments: async (project, { first = 20, after }, { db }) => {
    const rows = await db.deployment.findMany({
      where: { projectId: project.id },
      take: first + 1,
      cursor: after ? { id: decode(after) } : undefined,
    });
    return toConnection(rows, first);
  },
};

Features

(what it does)

One schema, two APIs
GraphQL and REST are generated from one source, so docs and SDKs never drift from reality.
Fair rate limits
Token buckets per key, with clear headers that tell developers exactly when to retry.
Docs you can run
Every example in the docs is executable against a sandbox key.

(in use)

ship it ✓

atlasapi.app/settings

Challenges

(what was hard)

01N+1 queries
Nested queries fanned out into thousands of reads. Batching brought p99 down to 42ms.
02Breaking changes
Schema checks in CI block any change that would break a query seen in the last 30 days.

Results

2B requests / month
42ms p99 latency
600+ developers

Atlas serves 2 billion requests a month to 600+ developers, and scraping traffic dropped to zero.

(next project)Nimbus AnalyticsSaaS Dashboard ↗Get in touch

design & code by Shreya Pathak

Problem

Architecture

Features

One schema, two APIs

Fair rate limits

Docs you can run

Challenges

01N+1 queries

02Breaking changes

Results