Unexpected 429 Errors on Vertex AI (Gemini 2.0/2.5 Flash Lite) via Firebase Functions despite <0.1% quota usage

Gemini API
, , , ,
1

Hello everyone,

I’m seeking insights into recurring 429 (Too Many Requests) errors I’ve been encountering with Gemini models via Vertex AI, even though my Google Cloud Quota dashboard shows negligible usage (<0.1%). While the errors have become rare after my initial mitigations, they still appear occasionally, and I haven’t been able to clarify if this is an instantaneous quota limit or a specific rate-limiting behavior.

Technical Environment

  • Platform: React Native Expo (Client)

  • Backend: Firebase Functions v2 (Node.js 22)

  • API: Vertex AI (Direct integration)

  • Models: Gemini 2.5 Flash Lite and Gemini 2.0 Flash Lite

  • Region: Switched from europe-west1 to Global Endpoint after initial errors.

Use Case & Latency Details

  1. File Processing (Gemini 2.5 Flash Lite): Two functions send file URIs from Cloud Storage to Gemini for analysis. This process typically takes 10-15 seconds per request.

  2. Chatbot (Gemini 2.0 Flash Lite): A single function where the first request includes a file URI (latency ~5s), and subsequent turns are text-only (latency 1-3s, sometimes near-instant).

Implemented Solutions & The “Testing” Anomaly

After encountering initial 429 errors, I moved all models to the Global Endpoint and implemented a Jittered Exponential Backoff (1, 2, 3, 4s intervals). This seemed to resolve the issue initially.

However, during a migration to Node.js 22, I deployed the same 3 functions as new instances with a “-testing” suffix for verification. Surprisingly, I started receiving 429 errors on these “testing” functions even with fewer than 10 concurrent users. After refining the jitter mechanism (randomized increases between 1-2s) and updating my original production functions with this new logic, the 429s disappeared entirely.

My Questions

  1. Since my total quota usage is under 0.1%, could the 429s on newly named functions be caused by “Cold Start” related concurrency bursts or instantaneous rate limits?

  2. Is there a known “Warm-up” period for newly created function names or endpoints on the Vertex AI side where rate limits are more restrictive?

  3. Beyond the standard Quota dashboard, is there a specific metric area in Google Cloud Console to monitor instantaneous throttling (RPM/TPM) specifically for Vertex AI?

I would greatly appreciate any technical insights or experiences from anyone who has encountered similar “new deploy/new name” anomalies.

3 Likes
2

Same here, we’re getting 429s very frequently for Europe endpoints despite being far below quota.

1 Like