Estimation of availability and reliability in CurveBS CurveBS uses the RAFT protocol to maintain consistency of stored data. It generally takes the form of 3 replicas of data. If one replica fails required to handle the failure according to the actual situation of the system. Estimation of availability and reliability in the three- replicas case Assume that the total number of disks in Curve

Than Ever? Yes, It Is • AI User + Usage + CapEx Growth = Unprecedented • AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising • 60% 10% 21% 15% 0% Details on Page 293 USA – LLM #1 China USA – LLM #2 AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising Than Ever? Yes, It Is • AI User + Usage + CapEx Growth = Unprecedented • AI Model Compute Costs High / Rising + Inference Costs Per Token Falling = Performance Converging + Developer Usage Rising •

CURVE: a High-performance Software-Defined storage for Cloud Native applications Xiang Dong Architect, Netease Hangzhou Research InstituteAgenda � Why develop storage Design objectives Achievements file systemopen-source storage ● Requirements ○ Cloud Native ○ Easy operation and maintenance ○ High performance ● CEPH ○ Complex and Large amount of codes ○ system maintenance are difficult in corner block Storage for KVM / Kubernetes / iSCSI Design objectives ● Thin provisioning storage pools ● High performance ● Easy maintenance ● Cloud NativeAgenda Why develop storage Design objectives �

operations, technology, and budgets  Actively involving users throughout development to ensure high operational value Agile practices integrate planning, design, development, and testing into an iterative operations, technology, and budgets  Actively involving users throughout development to ensure high operational value Agile is built around a culture of small, dynamic, empowered teams actively collaborating requirements or features up front, as the program backlog will evolve over time. Subsequent strategic, high-level planning sessions focus on the release and sprint levels. They outline the intent of a release

search, and AI require seamless integration of structured, semi-structured, and unstructured data with high-quality indexing. Yet, fragmented systems and inconsistent formats create silos, limiting accessibility intelligent search. However, many organizations lack a scalable infrastructure, leading to slow processing, high costs, and missed opportunities. Security, Compliance & Governance – AI, search, and analytics must standards (e.g., GDPR, HIPAA). This requires encryption at rest, role-based access control, and high availability to ensure data integrity, privacy, and governance. To overcome these challenges, data engineering

organization.” They are self-healing ii. Netflix – April 21, 2011; entire Amazon AWS US-EAST availability zone went down; Netflix was unaffected. Since 2008 they were building and testing with the Chaos each week with expectation that everyone is to teach or learn something in that timeframe. Place high value on mentoring ii. Encourage life-long learning. Our field is only progressing at faster rates to simultaneously achieve Information Security goals and create high degree of assurance for confidentiality, integrity, and availability b. Don’t inspect security in at the end, it’s integrated as part

accidental, not the essential, difficulties.”Solutions to Accidental Difficulties • High-level languages • The most a high-level language can do is to furnish all the constructs the programmer imagines essence, or are they remainders of our accidental difficulties?”Silver bullets? • Ada and other high-level language advances • Object-oriented programming • Artificial intelligence • Expert systems • servers, but I need more servers so I need… • an Infrastructure as a Service platform, but for high availability I need… • another rack of servers, but to pay for it I need…Always ask yourself • Am I solving

time-sucking configuration changes or complicated processes.” iv. 2015 State of DevOps Report – high performers had MTTR 168x faster than low performers b. CREATE OUR CENTRALIZED TELEMETRY INFRASTRUCTURE threats 6. Resources (RAM, disk, CPU, bandwidth, and others with hard or soft limits) 7. Health & availability 8. Startups & shutdowns 9. Faults & errors 10. Circuit breaker trips 11. Delays 12. Backup “people closest to a problem typically know the most about it.” iii. 2014 State of DevOps Report – high performing organizations relied more on peer review and less on external approval of change. The

OUR TEST SUITE i. Incorporate automate tests to validate the “-ilities” that are important (availability, capacity, security, etc.) ii. Incorporate security hardening testing and evaluation m. PULL due to potential rework and merge problems 1. Cross-cutting dependency improvements often provide high-payoffs d. ADOPT TRUNK-BASED DEVELOPMENT PRACTICES i. CI and Trunk-based development (TBD) are countermeasures smoke test to validate system behavior vi. Provides fast feedback vii. 2014 State of DevOps Report – high performers had deployment lead times measured in minutes or hours; low performers had deployment

'medium-priority item')) q.put((0, 'high-priority item')) q.put((10, 'low-priority item')) print(await q.get()) print(await q.get()) print(await q.get()) asyncio.run(main()) (0, 'high-priority item') (1, 'medium-priority sys.exit(0) after all child processes have exited nor- mally. max_restarts defaults to 100. Availability: Unix tornado.process.task_id() → int | None Returns the current task id, if any. Returns None can be used as an alternative to an exit callback if the signal handler is causing a problem. Availability: Unix wait_for_exit(raise_error: bool = True) → Future[int] Returns a Future which resolves