How Baidu’s Jarvis2.0 Redefined Cloud‑Native Microservice Governance

Business Background

With the rapid growth of cloud computing and microservice architectures, Baidu launched Jarvis2.0 to support its extensive advertising technology stack. The commercial platform evolved from a few dozen modules to more than 1,000 microservices, creating a complex service mesh that needed robust governance.

Challenges of Scaling Microservices

The explosion of microservices introduced problems such as slow full‑stack releases (average 30 minutes, 95th‑percentile over 100 minutes), poor container isolation, lengthy configuration updates, and inefficient monitoring and debugging processes.

Jarvis Platform Overview

Jarvis provides a self‑service platform covering the entire microservice lifecycle—development, testing, deployment, and operations. It integrates more than ten governance components (JVMTI probes, Launcher, Honeybee, traffic recording, log collection, application diagnostics, etc.) and starter kits (ConfigStarter, StarlightStarter, JdbcStarter, AcutatorStarter, RedisStarter).

Multi‑Runtime Architecture

Jarvis2.0 introduces a Multi‑Runtime design that moves various distributed capabilities into independent runtimes, which are then co‑located with the application runtime. The three core parts are:

Moonlight Runtime (Data Plane) : Deployed as a sidecar, built with GraalVM native image to start in under 5 seconds and consume only 128 MiB. It bundles over ten components for startup management, monitoring, diagnostics, dynamic tuning, and security.

Deployment Plane : Uses layered Docker images (Jib), OpenKruise CloneSet for in‑place upgrades, and KubeVela OAM for multi‑cluster orchestration, enabling zero‑code migration, gray‑scale releases, and automatic sidecar injection.

Gravity Control Plane : Implements an xDS‑based control protocol for service registration, configuration, routing, traffic weighting, and instant parameter adjustments. It also extends xDS to support rate limiting, circuit breaking, log level changes, and diagnostics.

Key Governance Paths

Probe Insertion Path : Supports dynamic probe injection, removal, and hot upgrades. Native compilation required custom adaptations for JVMTI capabilities.

Dynamic Governance Path : Provides traffic coloring, rate limiting, circuit breaking, parameter tuning, hot configuration updates, and application diagnostics via sidecar‑based agents.

Static Governance Path : Uses the Launcher component for automatic JAR replacement during application restarts.

Monitoring & Alerting Path : Offers Metrics, Tracing, and Log pipelines; Tracing processes over 7 billion calls per day with sub‑10‑second query latency.

Deployment Innovations

Jarvis automates Docker image layering for SpringBoot, static web, Node, and Go applications using Jib, reducing image pull time by 75 %. Multi‑cluster gray‑scale releases are achieved with the OAM model and KubeVela, allowing one‑click deployments across clusters.

Project Benefits

Jarvis2.0 now serves 60+ product lines, 3 k+ backend services, and over 40 k instances (200 k CPU cores). It saved 2.14 PD (person‑days) per day in migration effort and reduced core governance operation times dramatically (e.g., full‑stack release 95th‑percentile from 126 minutes to 36 minutes, configuration hot‑update from 30 minutes to 1 minute, online traffic switch from >30 minutes to 3‑5 seconds).

Stability improvements include reducing full‑scale governance component rollout from 1‑2 months to 1 day and cutting abnormal node removal time from 30 minutes to a few seconds. The platform now supports multi‑language stateless web services (Java, Node, Go) with a unified governance stack.