Top Insights from the 2016 Global Agile Operations Summit

Opening Session – Agile Operations Trends

Cheng Yongxin (Xinju Network) described the evolution of enterprise‑grade IT operations toward large‑scale, automated, and intelligent PaaS cloud‑ops. He outlined Xinju’s four core components that together form an enterprise PaaS:

AMP – an automation operations platform that orchestrates provisioning, configuration, and remediation.

APM – application performance management that collects latency, error‑rate, and resource‑usage metrics.

Ivory – a big‑data log analysis platform that stores and indexes terabytes of logs for real‑time querying.

DCOS – a container orchestration layer that provides unified monitoring, business‑level analytics, and intelligent operation capabilities.

The combined stack enables enterprises to shift from manual, process‑driven ops to agile, data‑driven cloud‑ops.

Key Technical Sessions

Open‑Source Adoption in Enterprises

Ren Ming (China UnionPay) presented a maturity‑assessment model for open‑source software, covering:

Evaluation criteria (stability, security, community health).

Current production open‑source stack (e.g., Linux, MySQL, Kafka).

Governance mechanisms such as code‑review policies, version‑locking, and compliance audits.

The model helps large enterprises balance legacy “IOE” stacks with open‑source alternatives.

Network Automation with Forward

Wang Zhe (China Mobile) introduced Forward , a lightweight network‑device automation tool. Its architecture consists of:

Agent (runs on device) → Message Queue → Orchestrator (central logic) → Plugin System

Key features include declarative workflow definitions, idempotent configuration pushes, and extensible plugins for diverse device types. Forward is used to automate firmware upgrades, topology discovery, and fault remediation in complex operational scenarios.

Enterprise Security Checklist

Guan Heng (DEFCON speaker) outlined a security checklist spanning development, operations, and security teams. Highlights:

Secure coding standards and automated static analysis.

Runtime hardening (container isolation, least‑privilege IAM).

Incident‑response playbooks and regular red‑team exercises.

Common pitfalls such as over‑privileged service accounts and unpatched third‑party libraries.

Practical advice includes forming dedicated security engineering squads and integrating self‑developed defense tools into CI/CD pipelines.

Database & Operations Highlights

Performance‑Optimization Fundamentals

Yang Zhihong (Xinju) emphasized a systematic approach: baseline measurement → bottleneck identification → index tuning → query rewrite → hardware scaling. Case studies from finance, e‑commerce, and gaming illustrate how DBAs can translate technical findings into actionable recommendations for developers.

Search Processing Language (SPL)

Tang Wenjun introduced SPL, the query language of real‑time search engines. Compared with traditional SQL, SPL supports:

Streaming operators (filter, aggregate, window).

Full‑text tokenization and relevance scoring.

Time‑series analysis via built‑in functions.

Demo queries showed log‑event correlation and anomaly detection in sub‑second latency.

MySQL Horizontal Sharding (Ele.me)

Guo Guofei described Ele.me’s sharding architecture:

Decision phase – workload profiling and shard key selection (order_id).

Planning – capacity modeling, data‑migration strategy, and consistency guarantees.

Implementation – use of ProxySQL for routing, pt‑online‑schema‑change for schema evolution, and automated data‑split scripts.

Risk mitigation – dual‑write during cut‑over, monitoring of cross‑shard latency.

Post‑implementation – throughput increased by 3×, write latency reduced from 120 ms to 35 ms.

MySQL Performance‑Tuning Cases

Yang Jianrong presented several real‑world interventions:

Graceful MySQL restart using pt‑heartbeat to avoid client disruption.

Semi‑automated Data‑Guard (DG) construction with Ansible playbooks.

System‑level optimisations: kernel TCP buffers, NUMA‑aware memory allocation, and I/O scheduler tuning.

Each case includes before/after metrics and scripts used.

High‑Concurrency MySQL & InnoDB MVCC

Zhang Qinglin explained typical bottlenecks (lock contention, buffer‑pool pressure) and demonstrated parameter tuning:

innodb_buffer_pool_size = 70% of RAM
innodb_flush_log_at_trx_commit = 2
innodb_thread_concurrency = 0  # auto

He also described InnoDB’s multi‑version concurrency control (MVCC) workflow, emphasizing the role of undo logs and consistent read snapshots for reducing lock wait times.

Enterprise‑Wide Space‑Management Platform

Zhao Yongtian (ICBC) critiqued threshold‑based alerts and introduced a centralized space‑management system that aggregates storage metrics from all DB instances, provides predictive capacity forecasting, and automates cleanup policies. Architecture components include:

Agent on each DB host reporting free/used space via REST.

Central analytics engine (Spark) that predicts growth trends.

Policy engine that triggers table‑partitioning or archive jobs.

The platform reduced out‑of‑space incidents by 80 % in pilot deployments.

Cloud & Container Operations

Docker Fundamentals with Alibaba Cloud Container Service

Jiang Jizhong demonstrated how Alibaba Cloud Container Service (ACK) abstracts Kubernetes cluster provisioning, image registry integration, and auto‑scaling groups. Typical operational scenarios covered include rolling updates, blue‑green deployments, and health‑check configuration.

Database x86 Architecture & Active‑Active Design

Luo Chun (Woqiu Technology) discussed migrating legacy database workloads to x86 servers to lower TCO. Key technical points:

CPU‑bound workloads benefit from higher core counts and SIMD extensions.

Resource isolation via cgroups and NUMA pinning.

Active‑active data‑center topology using synchronous replication (Galera) and automatic failover.

Encryption at rest (LUKS) and in‑flight (TLS 1.3) to meet security requirements.

Weibo Scaling with Docker & Hybrid Cloud

Wang Guansheng described Weibo’s transition from static capacity planning to a Docker‑based hybrid cloud model. By deploying a Kubernetes cluster across public and private clouds, they achieved:

Elastic provisioning of thousands of nodes within 10 minutes.

Automated CI/CD pipelines that build, test, and roll out micro‑services.

Resource‑level cost monitoring via Prometheus and Grafana.

Middleware Management Tooling

Zhang Peng highlighted pain points such as fragmented observability and cost leakage in middleware stacks (e.g., Kafka, RabbitMQ, Redis). He proposed a tooling suite that integrates:

Metrics exporters (JMX, StatsD).

Centralized tracing (OpenTelemetry) for request‑level latency.

Policy‑driven scaling rules based on queue depth and CPU usage.

NetEase Hive Cloud – Developer‑Centric Automation

Zhang Xiaolong explained the Hive Cloud’s “developer‑first” philosophy. Docker containers are used to provide self‑service environments where developers can:

Instantiate isolated test clusters with a single CLI command.

Leverage pre‑configured CI pipelines that include static code analysis and security scanning.

Consume shared micro‑service libraries via an internal artifact repository.

Shanda Game Cloud Platform

Xu Feng outlined Shanda’s custom G‑cloud architecture tailored for game workloads. Highlights include:

High‑throughput networking using RDMA over Converged Ethernet.

GPU‑accelerated instances for real‑time physics simulation.

Multi‑resource scheduling that balances CPU, memory, and bandwidth per game session.