Alright folks, been running Braintrust for a full quarter now across three different EKS clusters handling our main web app, APIs, and data processing. I jumped in because the promise of a single pane for both infra *and* application traces was too good to pass up, especially coming from a cobbled-together Grafana/Loki/Tempo stack that was getting expensive and fragile.
The hype is real on the unified view. Correlating a spike in Lambda duration directly to a specific database query pattern happening in a pod was literally a two-click drill-down. The data freshness is insane compared to what I was used to. Setting up a simple alert for anomalous spend was also way easier than I anticipated. Here's the basic cost alert rule I set up that actually saved our bacon last week:
```yaml
# braintrust/alert-rules/cost-anomaly.yaml
alert: CostSpikeDetector
metric: aws.billing.estimated_charges
condition: current > (7d_avg * 1.5)
window: 1h
tags:
severity: high
team: platform-finops
```
The catch, and it's a significant one, comes when you start scaling. The pricing model is *aggressive*. Ingest is clean, but retention and certain high-cardinality custom attributes (we tag everything with `service`, `env`, `version`, `customer_tier`) can make your bill do unpredictable things. My month two bill was 40% higher than month one, purely from extending trace retention and adding two custom attributes for a new feature launch.
The other thing is agent stability. The DaemonSet for infra metrics has been rock solid. The application side, where you're using the SDKs, we've had a few hairy moments during deployments where spans dropped for a minute or two. Their support is responsive, but it's something to be aware of if you need 100% fidelity.
So, is it worth it? If you're serious about FinOps and need to connect cost directly to service performance, absolutely. The visibility is transformative. But go in with your eyes open: model your costs carefully, especially around retention, and pressure-test the tracing agents before you fully commit. I'm not going back, but I'm watching my custom metric tags like a hawk now.
cost first, then scale