René's URL Explorer Experiment

Title: feat: MLflow distributed tracing for Feast feature server by Vperiodt · Pull Request #6405 · feast-dev/feast · GitHub

Open Graph Title: feat: MLflow distributed tracing for Feast feature server by Vperiodt · Pull Request #6405 · feast-dev/feast

X Title: feat: MLflow distributed tracing for Feast feature server by Vperiodt · Pull Request #6405 · feast-dev/feast

Description: What this PR does / why we need it Adds end-to-end MLflow integration for Feast: distributed tracing, fine-tuning dataset export, and MLflow GenAI Dataset sync — enabling ML teams to trace agent-feature interactions, produce training data from production traces, and serve curated evaluation datasets through Feast's online store. 1. Distributed Tracing (mlflow.enable_distributed_tracing: true) Server-side spans: Every feature retrieval endpoint (get-online-features, retrieve-online-documents, write-to-online-store) creates an MLflow span with feature names, entity count, and project metadata. Cross-process trace linking: W3C traceparent header propagation — Feast spans become children of the caller's trace. MCP server forwards mcp-session-id for session grouping. In-process LLM span tagging: A span processor (feast_span_processor) automatically stamps LLM/CHAT_MODEL spans with feast.context_features when the SDK is used directly. PII redaction: Optional redact_entity_pii scrubs entity values from span inputs. Sampling: trace_sampling_ratio controls what percentage of requests produce spans. Graceful degradation: Missing MLflow, incompatible versions, or span creation failures never block feature serving. 2. Fine-Tuning Export Pipeline (feast finetuning export) Trace extraction: Queries MLflow for CHAT_MODEL spans, extracts prompt/completion pairs with feast.* context attributes. Assessment extraction: Automatically pulls expectations and feedback logged on traces (via MLflow UI or API) into metadata["expectations"] and metadata["feedback"] — enabling resolve_labels_from_mlflow to work end-to-end without manual label population. Label resolution: Joins with a Feast LabelView via get_historical_features (conflict policies: LAST_WRITE_WINS, LABELER_PRIORITY, MAJORITY_VOTE) or promotes MLflow expectations to corrected_response. Export formats: OpenAI-compatible JSONL or Feast-enriched JSONL (includes trace provenance, feature refs, entity values). --sync-feedback: Bridges resolved labels back to MLflow as feedback assessments. --register: Logs exported dataset as a tracked MLflow artifact. --dataset-tags: Attaches key=value metadata to registered datasets. 3. MLflow GenAI Dataset Sync (feast datasets sync) MlflowDatasetSource: A CUSTOM_SOURCE DataSource that stores MLflow dataset connection metadata (dataset name, tracking URI, field mappings) with full proto serialization. Sync engine: Fetches records from MLflow GenAI Datasets, flattens nested columns (inputs, expectations, source, tags), writes to online + offline stores via write_to_online_store / push. Incremental sync: Watermark-based (feast_last_sync_time tag) — only syncs records updated since last run. feast datasets preview: Inspect flattened records before committing a sync. Field mapping: Dot-path overrides (e.g. expectations.expected_response → golden_answer). Retry + batching: Exponential backoff on fetch failures, configurable batch size for writes. 4. Trace Assessment Sync (feast datasets sync-assessments) Assessment ingestion: Scans MLflow traces for expectations and feedback logged via the MLflow UI or mlflow.log_expectation / mlflow.log_feedback, flattens them into rows (trace_id, assessment_name, value, source_id, rationale), and writes to a Feast FeatureView or LabelView. Selective sync: --assessment-names filters which assessments to ingest. Bridges MLflow annotations to Feast: Enables runtime lookup of human/automated feedback and direct use as fine-tuning labels via the LabelView path. Configuration: mlflow: enabled: true tracking_uri: "http://mlflow:5000" enable_distributed_tracing: true trace_sampling_ratio: 1.0 redact_entity_pii: false Which issue(s) this PR fixes: Checks I've made sure the tests are passing. My commits are signed off (git commit -s) My PR title follows conventional commits format Testing Strategy Unit tests Integration tests Manual tests Testing is not required for this change Misc

Open Graph Description: What this PR does / why we need it Adds end-to-end MLflow integration for Feast: distributed tracing, fine-tuning dataset export, and MLflow GenAI Dataset sync — enabling ML teams to trace agent-fe...

X Description: What this PR does / why we need it Adds end-to-end MLflow integration for Feast: distributed tracing, fine-tuning dataset export, and MLflow GenAI Dataset sync — enabling ML teams to trace agent-fe...

Opengraph URL: https://github.com/feast-dev/feast/pull/6405

X: @github

direct link

Domain: github.com

Links:

Viewport: width=device-width