fixes broken AWS rate limit backoff

src/shared/key-management/anthropic/provider.ts

@@ -153,11 +153,6 @@ export class AnthropicKeyProvider implements KeyProvider<AnthropicKey> {
 
     const selectedKey = keysByPriority[0];
     selectedKey.lastUsed = now;
-    selectedKey.rateLimitedAt = now;
-    // Intended to throttle the queue processor as otherwise it will just
-    // flood the API with requests and we want to wait a sec to see if we're
-    // going to get a rate limit error on this key.
-    selectedKey.rateLimitedUntil = now + KEY_REUSE_DELAY;
     return { ...selectedKey };
   }
 
@@ -226,4 +221,11 @@ export class AnthropicKeyProvider implements KeyProvider<AnthropicKey> {
     });
     this.checker?.scheduleNextCheck();
   }
+
+  public throttle(hash: string) {
+    const key = this.keys.find((k) => k.hash === hash)!;
+    const now = Date.now();
+    key.rateLimitedAt = now;
+    key.rateLimitedUntil = now + KEY_REUSE_DELAY;
+  }
 }

src/shared/key-management/aws/provider.ts

@@ -37,13 +37,13 @@ export interface AwsBedrockKey extends Key, AwsBedrockKeyUsage {
  * Upon being rate limited, a key will be locked out for this many milliseconds
  * while we wait for other concurrent requests to finish.
  */
-const RATE_LIMIT_LOCKOUT = 1000;
+const RATE_LIMIT_LOCKOUT = 4000;
 /**
  * Upon assigning a key, we will wait this many milliseconds before allowing it
  * to be used again. This is to prevent the queue from flooding a key with too
  * many requests while we wait to learn whether previous ones succeeded.
  */
-const KEY_REUSE_DELAY = 500;
+const KEY_REUSE_DELAY = 250;
 
 export class AwsBedrockKeyProvider implements KeyProvider<AwsBedrockKey> {
   readonly service = "aws";
@@ -131,11 +131,6 @@ export class AwsBedrockKeyProvider implements KeyProvider<AwsBedrockKey> {
 
     const selectedKey = keysByPriority[0];
     selectedKey.lastUsed = now;
-    selectedKey.rateLimitedAt = now;
-    // Intended to throttle the queue processor as otherwise it will just
-    // flood the API with requests and we want to wait a sec to see if we're
-    // going to get a rate limit error on this key.
-    selectedKey.rateLimitedUntil = now + KEY_REUSE_DELAY;
     return { ...selectedKey };
   }
 
@@ -199,4 +194,11 @@ export class AwsBedrockKeyProvider implements KeyProvider<AwsBedrockKey> {
       this.update(hash, { lastChecked: 0, isDisabled: false })
     );
   }
+
+  public throttle(hash: string) {
+    const key = this.keys.find((k) => k.hash === hash)!;
+    const now = Date.now();
+    key.rateLimitedAt = now;
+    key.rateLimitedUntil = now + KEY_REUSE_DELAY;
+  }
 }

src/shared/key-management/index.ts

@@ -63,6 +63,7 @@ export interface KeyProvider<T extends Key = Key> {
   getLockoutPeriod(model: Model): number;
   markRateLimited(hash: string): void;
   recheck(): void;
+  throttle(hash: string): void;
 }
 
 export const keyPool = new KeyPool();
@@ -80,4 +81,4 @@ export {
 export { AnthropicKey } from "./anthropic/provider";
 export { OpenAIKey } from "./openai/provider";
 export { GooglePalmKey } from "./palm/provider";
-export { AwsBedrockKey } from "./aws/provider";
+export { AwsBedrockKey } from "./aws/provider";

src/shared/key-management/key-pool.ts

@@ -72,6 +72,11 @@ export class KeyPool {
     }, 0);
   }
 
+  public throttle(key: Key) {
+    const provider = this.getKeyProvider(key.service);
+    provider.throttle(key.hash);
+  }
+
   public incrementUsage(key: Key, model: string, tokens: number): void {
     const provider = this.getKeyProvider(key.service);
     provider.incrementUsage(key.hash, model, tokens);

src/shared/key-management/openai/provider.ts

@@ -221,15 +221,6 @@ export class OpenAIKeyProvider implements KeyProvider<OpenAIKey> {
 
     const selectedKey = keysByPriority[0];
     selectedKey.lastUsed = now;
-
-    // When a key is selected, we rate-limit it for a brief period of time to
-    // prevent the queue processor from immediately flooding it with requests
-    // while the initial request is still being processed (which is when we will
-    // get new rate limit headers).
-    // Instead, we will let a request through every second until the key
-    // becomes fully saturated and locked out again.
-    selectedKey.rateLimitedAt = now;
-    selectedKey.rateLimitRequestsReset = KEY_REUSE_DELAY;
     return { ...selectedKey };
   }
 
@@ -383,20 +374,16 @@ export class OpenAIKeyProvider implements KeyProvider<OpenAIKey> {
     this.checker?.scheduleNextCheck();
   }
 
-  /** Writes key status to disk. */
-  // public writeKeyStatus() {
-  //   const keys = this.keys.map((key) => ({
-  //     key: key.key,
-  //     isGpt4: key.isGpt4,
-  //     usage: key.usage,
-  //     hardLimit: key.hardLimit,
-  //     isDisabled: key.isDisabled,
-  //   }));
-  //   fs.writeFileSync(
-  //     path.join(__dirname, "..", "keys.json"),
-  //     JSON.stringify(keys, null, 2)
-  //   );
-  // }
+  /**
+   * Called when a key is selected for a request, briefly disabling it to
+   * avoid spamming the API with requests while we wait to learn whether this
+   * key is already rate limited.
+   */
+  public throttle(hash: string) {
+    const key = this.keys.find((k) => k.hash === hash)!;
+    key.rateLimitedAt = Date.now();
+    key.rateLimitRequestsReset = KEY_REUSE_DELAY;
+  }
 }
 
 /**

src/shared/key-management/palm/provider.ts

@@ -122,11 +122,6 @@ export class GooglePalmKeyProvider implements KeyProvider<GooglePalmKey> {
 
     const selectedKey = keysByPriority[0];
     selectedKey.lastUsed = now;
-    selectedKey.rateLimitedAt = now;
-    // Intended to throttle the queue processor as otherwise it will just
-    // flood the API with requests and we want to wait a sec to see if we're
-    // going to get a rate limit error on this key.
-    selectedKey.rateLimitedUntil = now + KEY_REUSE_DELAY;
     return { ...selectedKey };
   }
 
@@ -186,4 +181,11 @@ export class GooglePalmKeyProvider implements KeyProvider<GooglePalmKey> {
   }
 
   public recheck() {}
+
+  public throttle(hash: string) {
+    const key = this.keys.find((k) => k.hash === hash)!;
+    const now = Date.now();
+    key.rateLimitedAt = now;
+    key.rateLimitedUntil = now + KEY_REUSE_DELAY;
+  }
 }