linkedin/venice

Venice, Derived Data Platform for Planet-Scale Workloads.

What it solves

Venice is a derived data storage platform designed for planet-scale workloads. It bridges the gap between offline, nearline, and online data worlds, specifically solving the problem of providing low-latency access to data that has been processed by ML training jobs for use during online inference workloads.

How it works

Venice acts as a stateful component that can back a Feature Store. It supports high-throughput asynchronous ingestion from both batch (Hadoop, Spark) and streaming (Samza) sources. Data is stored and replicated across regions using CRDT-based conflict resolution to ensure consistency.

Users can access data through three different client types depending on their latency requirements:

  • Thin Client: Stateless with < 10ms latency.
  • Fast Client: Partition-aware with < 2ms latency.
  • Da Vinci Client: Uses a stateful local cache for < 1ms latency.

Who it’s for

It is built for engineers and AI practitioners who need a scalable, multi-tenant storage layer for ML features, allowing them to feed the output of ML training jobs into a system that can be queried in real-time during inference.

Highlights

  • Flexible Ingestion: Supports batch push, incremental push, streaming writes, and hybrid stores.
  • Read Compute: Enables server-side operations like dot product and cosine similarity.
  • Active-Active Replication: Ensures high availability across regions with CRDT-based conflict resolution.
  • Multi-Client Options: Offers a range of clients (Thin, Fast, Da Vinci) to optimize for cost or performance.
  • Change Data Capture (CDC): Streams all data changes for ML feature retrieval and indexing.

Related

  • Project
    feast-dev/feast

    An open-source feature store for machine learning that manages the flow of data from offline training sets to low-latency online inference.

  • Project
    airbnb/chronon

    A data platform for AI/ML that abstracts data computation and serving, ensuring point-in-time consistency between offline training data and online real-time feature serving.

  • Project
    ai-dynamo/dynamo

    An open-source, datacenter-scale orchestration layer that coordinates multiple inference engines across GPU clusters to maximize throughput and minimize latency.

  • Project
    llm-d/llm-d

    A high-performance distributed inference serving stack for Kubernetes that optimizes LLM production deployments across various hardware accelerators.

  • Project
    datachain-ai/datachain

    A Python library that turns unstructured files in cloud storage into versioned, typed datasets that can be queried at warehouse speed without moving the data.