Uber’s Pinot Query Overhaul: Streamlined Layers and Improved Observability

Uber Redesigns Apache Pinot Query Architecture for Simpler Execution and Improved Predictability

Uber has redesigned its Apache Pinot query architecture to:

• Simplify execution
• Support richer SQL capabilities
• Improve predictability for internal analytical workloads

The prior Neutrino system, which layered Presto on top of Pinot, has now been replaced with a lightweight proxy called Cellar, leveraging Pinot’s Multi-Stage Engine Lite Mode. The goal is reduced complexity, stricter query limits, and stronger multi-tenant isolation.

---

Transition: From Neutrino to Cellar

Previous Architecture — Neutrino

• Stateless microservice combining Presto coordinator and worker processes
• PrestoSQL queries partially pushed to Pinot as PinotSQL
• Remaining query logic executed within Neutrino
• Enforced default or user-defined limits to avoid full table scans

Drawbacks:

• Added semantic complexity in query planning
• Limited isolation between tenants sharing the same proxy

Uber’s Neutrino query architecture — source: Uber blog

---

Scaling Challenges

Uber’s Pinot environment:

• Tables up to hundreds of terabytes
• Billions of records
• Query rates from single-digit QPS to thousands QPS

At this scale, multi-stage queries can easily exceed latency or resource budgets.

Pinot 1.4 — Lite Mode Features

• Configurable record limits at leaf stages
• Scatter-gather execution:
• Leaf stages on Pinot servers
• Remaining operators execute on the broker
• Predictable performance for complex queries

---

Cellar: New Query Architecture

Cellar proxy now directly forwards queries to Pinot brokers.

Workload Handling

• Simple workloads → Pinot’s single-stage query engine
• Advanced SQL → Multi-Stage Engine Lite Mode

Core Enhancements

• Configurable leaf-stage limits, exposed in the explain plan
• Retains scatter-gather pattern
• Controlled support for joins and window functions
• Enhanced monitoring and logging for query performance visibility

High-level Cellar query architecture — source: Uber blog

---

Direct Connect Mode & Time-Series Integration

Direct Connect Mode

• Tenants can bypass the proxy and connect directly to Pinot brokers
• Provides full isolation

Time-Series Plugin

• Enables M3QL queries through Cellar
• Supports use cases:
• Tracing
• Log search
• Segmentation

Adoption Status:

• Serves ~20% of former Neutrino volume
• Plans to fully retire Neutrino over time

Cellar direct connect mode — source: Uber blog

---

Developer Experience with Client Libraries

Uber-Provided Clients (Java & Go)

Features:

• Handle Pinot’s response format
• Support partial results with warnings
• Enforce timeouts & retries
• Emit metrics: latency, success rates, warnings

Operational Visibility:

• Out-of-the-box Grafana dashboards for new users

---

Forward-Looking Roadmap

Uber sees this redesign as an evolution of OLAP systems aiming for:

• High QPS
• Sub-second latency
• Strong isolation
• Predictability

Future Plans:

• Broader rollout of MSE Lite Mode later this year
• Further performance & feature enhancements

---

Related Tools for Data Publishing & Monetization

For organizations or creators managing and tracking AI-powered data insights, platforms such as AiToEarn官网 can be relevant:

• Open-source global AI content monetization platform
• Enables generation, cross-platform publishing, and analytics
• Supports channels like Douyin, Kwai, WeChat, Bilibili, Facebook, YouTube, LinkedIn, X (Twitter), etc.
• Aligns with Uber-like goals for visibility and performance tracking in analytical tooling

---

Original Source:

Inside Uber’s Pinot Query Overhaul: Simplifying Layers and Improving Observability