feat: use cloudflare's ratelimiter if possible (#2544)

* feat: use cloudflare's ratelimiter if possible

According to our data, these 5 ratelimit configs cover nearly 75% of our
ratelimit operations.

By using cloudflare for these, I expect our latency to drop and accuracy
to increase

* fix: only use cf in async mode

apps/api/src/pkg/env.ts

@@ -1,6 +1,10 @@
 import { z } from "zod";
 import type { MessageBody } from "./key_migration/message";
 
+export const cloudflareRatelimiter = z.custom<{
+  limit: (opts: { key: string }) => Promise<{ success: boolean }>;
+}>((r) => typeof r.limit === "function");
+
 export const zEnv = z.object({
   VERSION: z.string().default("unknown"),
   DATABASE_HOST: z.string(),
@@ -42,6 +46,11 @@ export const zEnv = z.object({
         return 0;
       }
     }),
+  RL_10_60s: cloudflareRatelimiter,
+  RL_30_60s: cloudflareRatelimiter,
+  RL_200_60s: cloudflareRatelimiter,
+  RL_500_10s: cloudflareRatelimiter,
+  RL_200_10s: cloudflareRatelimiter,
 });
 
 export type Env = z.infer<typeof zEnv>;

apps/api/src/pkg/ratelimit/client.ts

@@ -1,8 +1,8 @@
 import { Err, Ok, type Result } from "@unkey/error";
 import type { Logger } from "@unkey/worker-logging";
-import type { Metrics } from "../metrics";
-
+import { cloudflareRatelimiter } from "../env";
 import type { Context } from "../hono/app";
+import type { Metrics } from "../metrics";
 import { retry } from "../util/retry";
 import { Agent } from "./agent";
 import {
@@ -11,7 +11,6 @@ import {
   type RatelimitRequest,
   type RatelimitResponse,
 } from "./interface";
-
 export class AgentRatelimiter implements RateLimiter {
   private readonly logger: Logger;
   private readonly metrics: Metrics;
@@ -67,7 +66,34 @@ export class AgentRatelimiter implements RateLimiter {
     req: RatelimitRequest,
   ): Promise<Result<RatelimitResponse, RatelimitError>> {
     const start = performance.now();
+    try {
+      if (req.async) {
+        // Construct a binding key that could match a configured ratelimiter
+        const lookup = `RL_${req.limit}_${Math.round(req.interval / 1000)}s` as keyof typeof c.env;
+        const binding = c.env[lookup];
+
+        if (binding) {
+          const res = await cloudflareRatelimiter.parse(binding).limit({ key: req.identifier });
 
+          this.metrics.emit({
+            metric: "metric.ratelimit",
+            workspaceId: req.workspaceId,
+            namespaceId: req.namespaceId,
+            latency: performance.now() - start,
+            identifier: req.identifier,
+            mode: "async",
+            error: false,
+            success: res.success,
+            source: "cloudflare",
+          });
+          return Ok({ pass: res.success, reset: -1, current: -1, remaining: -1, triggered: null });
+        }
+      }
+    } catch (err) {
+      this.logger.error("cfrl failed, falling back to agent", {
+        error: (err as Error).message,
+      });
+    }
     const res = await this._limit(c, req);
     this.metrics.emit({
       metric: "metric.ratelimit",

apps/api/wrangler.toml

@@ -35,6 +35,44 @@ bindings = [
   { name = "DO_USAGELIMIT", class_name = "DurableObjectUsagelimiter" },
 ]
 
+[[unsafe.bindings]]
+# The nameing scheme is important, because we're dynamically constructing
+# these in the api code
+#
+# RL_{LIMIT}_{DURATION}s
+#
+# The namespace_id schema is somewhat made up though.
+# I prefixed everything with 9900 -> 9900{limit}{duration}
+name = "RL_10_60s"
+type = "ratelimit"
+namespace_id = "99001060"
+simple = { limit = 10, period = 60}
+
+[[unsafe.bindings]]
+name = "RL_30_60s"
+type = "ratelimit"
+namespace_id = "99003060"
+simple = { limit = 30, period = 60}
+
+[[unsafe.bindings]]
+name = "RL_200_60s"
+type = "ratelimit"
+namespace_id = "990020060"
+simple = { limit = 200, period = 60}
+
+[[unsafe.bindings]]
+name = "RL_500_10s"
+type = "ratelimit"
+namespace_id = "990050010"
+simple = { limit = 500, period = 10}
+
+[[unsafe.bindings]]
+name = "RL_200_10s"
+type = "ratelimit"
+namespace_id = "990020010"
+simple = { limit = 200, period = 10}
+
+
 [queues]
 consumers = [
   { queue = "key-migrations-development", max_batch_size = 10, max_retries = 10, dead_letter_queue = "key-migrations-development-dlq" },
@@ -62,6 +100,45 @@ consumers = [
   { queue = "key-migrations-preview-dlq", max_batch_size = 10, max_retries = 10 },
 ]
 
+[[env.preview.unsafe.bindings]]
+# The nameing scheme is important, because we're dynamically constructing
+# these in the api code
+#
+# RL_{LIMIT}_{DURATION}s
+#
+# The namespace_id schema is somewhat made up though.
+# I prefixed everything with 9900 -> 9900{limit}{duration}
+name = "RL_10_60s"
+type = "ratelimit"
+namespace_id = "99001060"
+simple = { limit = 10, period = 60}
+
+[[env.preview.unsafe.bindings]]
+name = "RL_30_60s"
+type = "ratelimit"
+namespace_id = "99003060"
+simple = { limit = 30, period = 60}
+
+[[env.preview.unsafe.bindings]]
+name = "RL_200_60s"
+type = "ratelimit"
+namespace_id = "990020060"
+simple = { limit = 200, period = 60}
+
+[[env.preview.unsafe.bindings]]
+name = "RL_500_10s"
+type = "ratelimit"
+namespace_id = "990050010"
+simple = { limit = 500, period = 10}
+
+[[env.preview.unsafe.bindings]]
+name = "RL_200_10s"
+type = "ratelimit"
+namespace_id = "990020010"
+simple = { limit = 200, period = 10}
+
+
+
 # canary is a special environment that is used to test new code by a small percentage of users before it is rolled out to the rest of the world.
 # all settings must be the same as production, except for the route pattern
 [env.canary]
@@ -81,6 +158,45 @@ consumers = [
   { queue = "key-migrations-canary", max_batch_size = 10, max_retries = 10, dead_letter_queue = "key-migrations-canary-dlq" },
   { queue = "key-migrations-canary-dlq", max_batch_size = 10, max_retries = 10 },
 ]
+
+[[env.canary.unsafe.bindings]]
+# The nameing scheme is important, because we're dynamically constructing
+# these in the api code
+#
+# RL_{LIMIT}_{DURATION}s
+#
+# The namespace_id schema is somewhat made up though.
+# I prefixed everything with 9900 -> 9900{limit}{duration}
+name = "RL_10_60s"
+type = "ratelimit"
+namespace_id = "99001060"
+simple = { limit = 10, period = 60}
+
+[[env.canary.unsafe.bindings]]
+name = "RL_30_60s"
+type = "ratelimit"
+namespace_id = "99003060"
+simple = { limit = 30, period = 60}
+
+[[env.canary.unsafe.bindings]]
+name = "RL_200_60s"
+type = "ratelimit"
+namespace_id = "990020060"
+simple = { limit = 200, period = 60}
+
+[[env.canary.unsafe.bindings]]
+name = "RL_500_10s"
+type = "ratelimit"
+namespace_id = "990050010"
+simple = { limit = 500, period = 10}
+
+[[env.canary.unsafe.bindings]]
+name = "RL_200_10s"
+type = "ratelimit"
+namespace_id = "990020010"
+simple = { limit = 200, period = 10}
+
+
 [env.production]
 vars = { ENVIRONMENT = "production", SYNC_RATELIMIT_ON_NO_DATA = "1" }
 route = { pattern = "api.unkey.dev", custom_domain = true }
@@ -104,3 +220,42 @@ consumers = [
 
 [env.production.observability]
 enabled = true
+
+[[env.production.unsafe.bindings]]
+# The nameing scheme is important, because we're dynamically constructing
+# these in the api code
+#
+# RL_{LIMIT}_{DURATION}s
+#
+# The namespace_id schema is somewhat made up though.
+# I prefixed everything with 9900 -> 9900{limit}{duration}
+name = "RL_10_60s"
+type = "ratelimit"
+namespace_id = "99001060"
+simple = { limit = 10, period = 60}
+
+[[env.production.unsafe.bindings]]
+name = "RL_30_60s"
+type = "ratelimit"
+namespace_id = "99003060"
+simple = { limit = 30, period = 60}
+
+[[env.production.unsafe.bindings]]
+name = "RL_200_60s"
+type = "ratelimit"
+namespace_id = "990020060"
+simple = { limit = 200, period = 60}
+
+[[env.production.unsafe.bindings]]
+name = "RL_500_10s"
+type = "ratelimit"
+namespace_id = "990050010"
+simple = { limit = 500, period = 10}
+
+[[env.production.unsafe.bindings]]
+name = "RL_200_10s"
+type = "ratelimit"
+namespace_id = "990020010"
+simple = { limit = 200, period = 10}
+
+

internal/metrics/src/index.ts

@@ -82,7 +82,7 @@ export const metricSchema = z.discriminatedUnion("metric", [
     mode: z.enum(["sync", "async"]),
     success: z.boolean().optional(),
     error: z.boolean().optional(),
-    source: z.enum(["agent", "durable_object"]),
+    source: z.enum(["agent", "durable_object", "cloudflare"]),
   }),
   z.object({
     metric: z.literal("metric.usagelimit"),