So we bit the bullet and moved from a self-hosted OpenTelemetry collector to Traceloop Cloud. The sales pitch was less ops overhead, better UI. Reality was a bit more... interesting.
The main pain wasn't the SDK integration, that's straightforward. It was the semantic conventions. Our old, messy spans didn't magically become "AI traces." Traceloop expects certain attributes (llm.*, etc.) to do its thing. Had to go back and instrument properly. Example of the old vs new for an OpenAI call:
```python
# Old 'works enough' span
with tracer.start_as_current_span("openai_call") as span:
span.set_attribute("model", "gpt-4")
response = client.chat.completions.create(**args)
# New, Traceloop-useful span
from opentelemetry.semconv.ai import SpanAttributes
# ... need to set LLM spans, tokens, etc.
```
Without that, you're just paying for a fancy dashboard showing the same garbage traces. The vendor lock-in is real too. Exporting those "enhanced" traces back out isn't as simple as a collector config file anymore.
SQL is enough