Skip to main content

Numerix - Release Notes

Version 1.0.0 🚀​

Release Date: September 2025
Status: General Availability (GA)

The first stable release of Numerix — a Rust-based compute service for evaluating mathematical expressions over feature matrices with very low latency. Numerix executes postfix expressions from an etcd-backed registry using a stack-based evaluator and compiler-assisted SIMD.

🎯 What's New​

Core Engine​

  • Postfix Expression Execution: compute_id → postfix mapping in etcd; parser-free request path.
  • Stack-Based Evaluator: Linear-time execution over aligned vectors for predictable latency.
  • Compiler-Assisted SIMD: Relies on LLVM autovectorization (NEON/SVE on ARM; SSE/AVX on x86); no handwritten intrinsics.
  • Typed Evaluation: Internal conversion to fp32/fp64 for consistent performance/precision.

API Surface​

  • gRPC: Single RPC — numerix.Numerix/Compute.
  • Input Formats: Strings for ease, bytes for performance; both map to vectorized math internally.

Observability​

  • Datadog/DogStatsD metrics: Latency (P50/P95/P99), RPS, error rate.
  • Minimal HTTP diagnostics: /health (and optional /metrics).

🚀 Performance & Optimization​

  • Autovectorized Loops: Tight loops over contiguous memory enable the compiler to emit SIMD instructions automatically.
  • ARM Focus Option: Excellent results with AArch64; builds can enable NEON/SVE/SVE2:
RUSTFLAGS="-C target-feature=+neon,+sve,+sve2" \
cargo build --release --target aarch64-unknown-linux-gnu
  • Deterministic Runtime: No dynamic parsing in hot path; O(n) across tokens with vectorized inner ops.

🛠️ APIs​

gRPC​

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

Example call:

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

🏗️ Deployment & Configuration​

Environment​

APPLICATION_PORT=8083
APP_ENV=prd
APP_LOG_LEVEL=ERROR
APP_NAME=numerix

# Performance
CHANNEL_BUFFER_SIZE=10000

# etcd
ETCD_SERVERS=127.0.0.1:2379

# Metrics
METRIC_SAMPLING_RATE=1
TELEGRAF_UDP_HOST=127.0.0.1
TELEGRAF_UDP_PORT=8125

Containers​

  • Multi-arch images: linux/amd64, linux/arm64.
  • Build targets example: x86_64-unknown-linux-gnu, aarch64-unknown-linux-gnu.

🔄 Compatibility​

  • Clients: Any language with gRPC + generated stubs.
  • Architectures: amd64 and arm64; ARM builds can enable NEON/SVE/SVE2.

🐛 Known Issues​

  1. Introduce a configurable log sampling rate to reduce pod memory usage during computation errors.

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!