# GCP Vertex AI Vector Search Endpoint Idle

> Deployed Vector Search indexes keep replica nodes billing per node-hour whether queries arrive or not. ZopNight anchors on the continuous request-utilization CPU gauge staying under 1% and vetoes on any real query traffic. The query counter itself emits no datapoints while idle, so requiring it would blind the rule.

Source: https://zop.dev/integrations/gcp/recommendations/gcp-vertex-ai-vector-search-endpoint-idle
Updated: 2026-08-19

---

## Replicas bill while deployed

Deploying a Vector Search index to an IndexEndpoint allocates replica nodes that bill
per node-hour continuously. Retrieval-augmented prototypes are the classic source, built
in a sprint and left serving an application that was retired. The charge belongs to the
deployment, not the queries, so a zero-query endpoint costs the same as a busy one.

## The idle-blind counter problem

`VertexIndexEndpointQueryCount` is a delta counter: Cloud Monitoring writes datapoints
for it only when queries actually happen. An idle endpoint therefore has no query series
at all, and a rule that required the counter to be present-and-zero could never fire;
that defect existed here once and is why the current design anchors elsewhere. The anchor
is `VertexIndexEndpointCPU`, the request-utilization gauge that is continuously present
for any deployed endpoint: the rule fires when its average and peak both stay under 1%.
A query series that does exist and carries traffic vetoes the recommendation outright.

## Undeploy is the lever

There is no pause for a deployed index. Recovering the node-hours means undeploying the
index from the endpoint, which keeps the built index intact and redeployable; delete the
IndexEndpoint as well once nothing will be redeployed. Rebuilding later costs per-GiB
build fees, so undeploy-and-keep is usually the right first step.

## See which indexes are deployed

```bash
gcloud ai index-endpoints describe INDEX_ENDPOINT_ID --region us-central1 \
  --format="table(deployedIndexes.id,deployedIndexes.createTime)"
```

Metric reads behind this rule use `monitoring.timeSeries.list` from
`roles/monitoring.viewer`.

## Cost attribution and abstains

Vertex serving carries no rack rate in the calculated cost path (ZopNight declines to
invent a per-replica rate), so the rule fires only where GCP billing export ties real
spend to the endpoint's UID, and the recoverable figure is that attributed cost in full.
Silence is guaranteed when the CPU gauge is absent, when utilization reaches 1% on
average or peak, when a present query series shows any traffic, or when no positive cost
is known.