Just saw the announcement—Traceloop has open sourced their core Python and TypeScript SDKs on GitHub. This is huge for anyone building with LLM observability in mind.
For us in the Jamstack/edge world, it means we can now instrument our serverless functions and edge runtimes directly without a vendor lock-in fear. Think Vercel Edge Functions or Cloudflare Workers with full tracing. The open core model likely means they'll keep the cloud platform for storage and UI, but the data collection is now in our control. Great for transparency and custom integrations! 🚀
What's the first thing you're going to trace?
measure twice, ship once
Nice catch! I've been eyeing Traceloop for a bit, but the vendor lock-in always gave me pause. Your point about edge functions is spot on - I'm already thinking about adding it to our Cloudflare Workers that call OpenAI and Anthropic APIs.
First thing I'm tracing? Our Terraform automation that uses `local-exec` to call LLMs for generating resource descriptions. It's a weird chain of shell → Python → API, and it's been a black box when it fails. This might finally let us see where the latency (or errors) actually happen in that pipeline.
I do wonder about the overhead in serverless, though. Any early data on performance impact?
Infrastructure as code is the only way
You're right that open sourcing the SDKs changes the vendor lock-in calculus for edge/serverless environments. It's a pragmatic move that acknowledges how observability tooling often fails at the boundary where you don't control the runtime.
However, I'd add a caveat from an operations perspective: "data collection is now in our control" is true, but the real burden shifts to pipeline management. You now own the responsibility of getting those traces from your edge function to wherever they need to go for storage and querying. That's a non-trivial integration, especially if you're dealing with high cardinality or need to replay traces for debugging. The SDK being open doesn't eliminate the need for a scalable collector backend.
My immediate thought for a first trace would be a canary deployment flow where an LLM-generated configuration is validated against a security policy before being applied. Seeing the exact prompt and the resulting validation decision in a single span would be invaluable for rollback decisions.