# Serving Vertex inference to end users: what do you actually do on a 429 from dynamic shared quota? — retained record

> Retained reddit record from reddit:googlecloud.

Canonical: https://highsignal.app/data/records/00fe65a53f9d15c0

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:googlecloud)  published 2026-09-14 15:32 ingested 2026-09-15 02:39

[original source ↗](https://www.reddit.com/r/googlecloud/comments/1wg6orj/serving_vertex_inference_to_end_users_what_do_you/)

## retained text



```
I run a desktop app where people can use cloud inference without bringing their own API key. A small proxy relays to Vertex (gemini-2.5-flash-lite), meters the usage per account against a server side price table, and fails closed when an account runs out of budget. It works well at my current volume.

What I am trying to plan for is a spike. As I understand it, the newer Gemini models on Vertex are on dynamic shared quota, so there is no per project quota to raise, and a 429 can land at low usage if global capacity is tight that day. Provisioned Throughput is the only real reservation mechanism and it is far out of scale for where I am.

So for those of you serving user facing inference on Vertex without Provisioned Throughput:

* what do you actually do when you get a 429? Retry with backoff, fall back to a smaller model, something else?
* does the global endpoint meaningfully help compared to a regional one, or does it just move the problem?
* at low volume, how often does this really happen? I have no feel for the base rate and I would rather not discover it on a launch day.

My app can fall back to a local model, so I do have an escape hatch. But I would like to know what the realistic failure rate looks like before I lean on it.
```



document fetched 2026-09-15 00:39 · canonical https://reddit.com/r/googlecloud/comments/1wg6orj/serving_vertex_inference_to_end_users_what_do_you
