P2-C (parent): Long-tail observability sinks (~30 LiteLLM integrations)

Background

After P0-7 (real OTLP) and P1-8 (Helicone/Langsmith/Datadog/Slack), the rest of LiteLLM's observability surface.

Backlog (each spawns its own child issue when scheduled)

LLM-specific tracing/eval

• Lunary
• Langtrace
• Athina
• Galileo
• MLflow
• Braintrust
• Argilla
• Literal AI
• Humanloop
• Weights & Biases (Weave)
• Opik (Comet)
• Deepeval
• AgentOps
• PromptLayer

APM (in addition to OTLP)

• Honeycomb (direct)
• New Relic (direct)
• Logfire
• Arize AI
• PostHog
• Greenscale
• Traceloop

Log sinks (in addition to S3 once added)

• AWS S3 v1 + v2 — unified object_store sink (#531)
• AWS SQS
• Google Cloud Storage bucket — unified object_store sink (#531)
• GCS Pub/Sub
• Azure Blob — unified object_store sink (#531)
• Azure Sentinel SIEM
• DynamoDB
• Supabase
• GitLab
• BitBucket

Billing exporters

• Lago
• OpenMeter
• Vantage
• CloudZero

Pattern

Each sink = an exporter sub-module under `aisix-obs`. Same shape as P1-8: bounded mpsc → background batcher → HTTP POST → retry. ~1d each.

Track here, schedule by customer ask.

Start with the aisix-obs exporter sub-module pattern described here and review P1-8 for the established sink shape. Select one unchecked integration and confirm its scope in a child issue before changing code. Done means the chosen sink has a bounded mpsc, background batcher, HTTP POST, and retry behavior, with the child issue tracking it separately.