---
title: "Comment by Kartik N V J K on \"Observability on a Budget: Logs, Metrics, and Traces with Open-Source Tools\" | Hashnode"
description: "Keeping every error and slow trace covers most services, but model calls tend to fail quietly. I had an agent trace come back green end to end while the agent h"
url: "https://hashnode.com/posts/observability-on-a-budget-logs-metrics-and-traces-with-open-source-tools-zxgpm/6ae824538caaf19abed0e126/comment/6ae8929ca3de00710290368c"
---

[Discussion](https://hashnode.com/posts/observability-on-a-budget-logs-metrics-and-traces-with-open-source-tools-zxgpm/6ae824538caaf19abed0e126)

Comment [Observability on a Budget: Logs, Metrics, and Traces with Open-Source Tools](https://hashnode.com/posts/observability-on-a-budget-logs-metrics-and-traces-with-open-source-tools-zxgpm/6ae824538caaf19abed0e126)

[Kartik N V J K](https://hashnode.com/@kartik-nvj)

[Nov 2](https://hashnode.com/posts/observability-on-a-budget-logs-metrics-and-traces-with-open-source-tools-zxgpm/6ae824538caaf19abed0e126/comment/6ae8929ca3de00710290368c)

Keeping every error and slow trace covers most services, but model calls tend to fail quietly. I had an agent trace come back green end to end while the agent had done nothing, and a run like that is fast and error-free. So for LLM calls I wouldn't push the routine sample rate anywhere near zero.

[ZyVOP](https://hashnode.com/@zyvop)

[Nov 2](https://hashnode.com/posts/observability-on-a-budget-logs-metrics-and-traces-with-open-source-tools-zxgpm/6ae824538caaf19abed0e126/comment/6ae8929ca3de00710290368c#reply-6ae897150d864278f0f85f0f)

Good point. A successful trace doesn’t necessarily mean the agent achieved anything. LLM calls need some baseline sampling, with outcome-based checks to catch failures that latency and error rates miss.