Skip to main content

Numerix — Key Functionalities

Overview​

Numerix evaluates mathematical expressions over feature matrices with a simple, low-latency gRPC surface. Each request references a compute_id; Numerix resolves a postfix expression, maps variables to input columns, and evaluates it over fp32/fp64 vectors with compiler-assisted SIMD.

🚀 Core Capabilities​

Expression Evaluation​

  • Postfix execution: Linear-time, stack-based evaluation over aligned vectors.
  • Vectorized math: Compiler autovectorization (NEON/SVE on ARM, SSE/AVX on x86) — no handwritten intrinsics.
  • Typed compute: Inputs converted to fp32 or fp64 for predictable performance.

Input Formats​

  • Strings: Easy-to-produce feature values (converted internally).
  • Bytes: Efficient wire format for high-throughput paths.

Request Patterns​

  • Single entity or batch: Multiple entity_scores per call.
  • Schema-driven: Column order in schema drives variable mapping in expressions.

🎯 Developer Experience​

  • gRPC API: Simple single RPC — Numerix/Compute.
  • Protobuf schema: Language-agnostic client generation.
  • Deterministic behavior: No parsing at request time; expression resolved from registry.

gRPC Service​

service Numerix {
  rpc Compute(NumerixRequestProto) returns (NumerixResponseProto);
}

Example Call (grpcurl)​

grpcurl -plaintext \
  -import-path ./numerix/src/protos/proto \
  -proto numerix.proto \
  -d '{
    "entityScoreData": {
      "schema": ["feature1", "feature2"],
      "entityScores": [ { "stringData": { "values": ["1.0", "2.0"] } } ],
      "computeId": "1001",
      "dataType": "fp32"
    }
  }' \
  localhost:8080 numerix.Numerix/Compute

📊 Observability​

  • Datadog/DogStatsD: Latency (P50/P95/P99), RPS, error rate via UDP client.
  • Optional /metrics endpoint for local/adhoc debugging.

⚙️ Configuration & Registry​

  • etcd registry: compute_id (int) → postfix expression mapping.
  • Environment-driven config: endpoints, timeouts, sampling rate.

🧪 Example Scenarios​

Batched evaluation​

  • Submit multiple entities in one call to reduce RPC overhead; evaluate the same compute_id across all rows.

Mixed input formats​

  • Start with string inputs for ease; migrate to bytes for performance without changing the expression or API.

🔧 Tuning Knobs​

  • Data type: choose fp32 (speed) vs fp64 (precision).
  • Batch size: tune number of entities per call for your p99 vs throughput goals.
  • Sampling rate: adjust Datadog metric sampling to balance signal vs cost.

Contributing​

We welcome contributions from the community! Please see our Contributing Guide for details on how to get started.

Community & Support​

License​

BharatMLStack is open-source software licensed under the BharatMLStack Business Source License 1.1.

Built with ❤️ for the ML community from Meesho
If you find this useful, ⭐️ the repo — your support means the world to us!