New Feature Submissions:

1. Does your submission pass tests?
   • ✅ cargo test -p common
   • ✅ cargo test -p qdrant
2. Have you formatted your code locally using cargo +nightly fmt --all command prior to submission?
   • ✅ cargo +nightly fmt --all
3. Have you checked your code using cargo clippy --workspace --all-features command?
   • Not fully on my macOS environment due to local Python / pyo3 mismatch (abi3-py310 implies Python >= 3.10; my local interpreter is 3.9).
   • I did run:
     • ✅ cargo clippy -p qdrant --all-targets

Changes to Core Features:

• Have you added an explanation of what your changes do and why you’d like us to include them?
  • See “Problem Statement”, “Solution”, and “Implementation Details” sections above.
• Have you written new tests for your core changes, as applicable?
  • Not yet; happy to add tests if maintainers request specific coverage.
• Have you successfully ran tests with your changes locally?
  • Yes — see commands above.

🎯 Summary (Fixes #3322)

This PR introduces per-collection metrics for the REST API, addressing the request in #3322 to provide more granular visibility into Qdrant’s performance. It enables monitoring of latencies and request counts on a per-collection basis, which is essential for deployments with multiple collections.

/claim #3322

🔧 Problem Statement

Currently, REST metrics are aggregated globally. This makes it difficult for users to:

1. Identify which specific collection is experiencing high latency or traffic.
2. Debug performance bottlenecks that might be isolated to a single collection.
3. Track usage patterns across a multi-tenant environment.

🚀 Solution: REST Per-Collection Granularity

This implementation adds a collection label to relevant REST metrics.

Key Features:

• Automatic Extraction: The Actix middleware now extracts collection names from request paths.
• Path Normalization: To prevent high-cardinality issues in Prometheus, specific paths (e.g., /collections/my-coll/points/search) are normalized to a generic template (/collections/{name}/points/search) for the endpoint label, while still preserving the specific collection in the collection label.
• Telemetry Anonymization: Maintains system privacy by masking collection names in shared telemetry data while preserving the per-collection structure.
• Backward Compatibility: Existing global metrics remain unchanged; per-collection metrics are added as a new dimension.

📝 Implementation Details

Middleware Improvements

Modified src/actix/actix_telemetry.rs to:

• Implement extract_collection_from_path.
• Handle path normalization to ensure metrics labels remain manageable.
• Capture per-collection statistics during the request lifecycle.

Metrics & Telemetry

• WebApiTelemetry: Extended to store responses_by_collection.
• MetricsProvider: Updated to output rest_per_collection_... metrics when whitelisted endpoints are hit.
• Anonymization: Ensured that Anonymize for WebApiTelemetry properly masks collection names to collection_N during export.

Running test cases locally

https://github.com/user-attachments/assets/c07d14e8-d9ee-40c7-b059-a67f8d6f3bb6

Testing feature implementation locally

Calling endpoint without per_collection parameter

sets per_collection=false by default

Calling endpoint with per_collection=true parameter

@generall