{
  "id": 11140974,
  "title": "Self-hosted Langfuse: tracing 7% of my AI agents, and ClickHouse logging itself",
  "url": "https://urgent.news/2026/10/01/self-hosted-langfuse-tracing-7-of-my-ai-agents-and-clickhouse-logging",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-01T08:00:15.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/c1-anderson/self-hosted-langfuse-tracing-7-of-my-ai-agents-and-clickhouse-logging-itself-3e80"
  },
  "original_language": "en",
  "account": "On Discord, the author reported that their AI agent, following a comment ID, failed to post a reply despite drafting it. The agent took 25 minutes to respond with a confident, detailed summary that was entirely fabricated. The author self-hosted Langfuse to trace their agent's activities and discovered that the agent had been monitoring approximately 7% of their agents' traffic. The 25-minute turn was traced, revealing that the local model spent 25 minutes thinking and made three short trips to the web. The author initially assumed the delay was due to network issues, but the trace showed the network was the fastest component. The fabricated response was the result of the agent improvising when it discovered it lacked the 'post' skill. A fresh session was needed to correct the issue. The author then examined the traces from Langfuse to understand more about their models' performance. They found that a local model on a 6 GB GPU was significantly slower compared to cheaper cloud models, taking about 30 times longer. The author also noticed that only 7% of the traffic was being traced, with the rest running without tracing. This was due to the tracing plugin being enabled per profile and reading its Langfuse API keys through that profile's secret scope. The author realized the oversight after noticing the low call count for the local model and enabling tracing on all profiles. Additionally, the author discovered that ClickHouse, used for storing traces, was consuming an excessive amount of disk space, with 6.03 GiB being utilized while only 2.2 MiB contained actual trace data. The rest was ClickHouse's own diagnostic logging. The author's fix involved enabling tracing on every profile, copying the Langfuse keys into individual env files, and enabling the plugin for each profile. This would allow for more detailed analysis of cost and latency per agent role within the Langfuse UI.",
  "summary": "On Discord I told my agent post followed by a comment id. That's the approval step for a dev.to reply it had drafted: read the draft, say \"post\", done. Nothing got posted. The turn took 25 minutes and came back with a confident, detailed summary of the article. It was made up. I self-host Langfuse to trace my agents, so I went and looked at the trace. Then I checked what else it had seen. It had…",
  "key_points": [
    "Self-hosted Langfuse used to trace AI agent activities",
    "Agent spent 7% of traffic on tracing, not 25 minutes",
    "ClickHouse logging itself consumed excessive disk space"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}