AI 模型和平台
MariaDB 发布企业平台 2027 用于 Agent 工作量
在 Google 上的首选来源中添加 Unite.AI
2026年11月4日，MariaDB 宣布 Entertainment平台2027，使其综合交易，分析，内存和AI平台普遍开放。 发布版新增 agent内存，检索和响应搜索工具，公司报告企业服务器12.3的峰值吞吐量比赞助测试11.8版高56%.
发行软件包 Entertainment Server 12.3 配有 Exa, GridGain, Max Scale，企业管理器和 Entertainment Kubernetes 操作器。 它可以在任何云层的场地上运行，或者作为管理下的MariaDB云层服务。 一个重新设计的标志、彩色调色板和网站与MariaDB将自己定位为一个统一的数据平台，而不仅仅是一个可操作的数据库。
Agent 工具使用相同的活数据
一个发布的帖子将agent的特征分为开发和运行时间功能。 Agent内存在每个步骤后都会将agent的状态存储在内存中，并在agents和会话之间共享。 agent为人类批准而暂停或中断，可以从其保存状态恢复，而不是重启；LangGraph被支持出框。
语义缓存通过识别一个措辞不同的问题具有与先前相同的含义并返回存储的响应来处理重复的模型调用。 MariaDB 称这样可以缩短响应时间，降低用户提出类似问题的应用程序的象征性成本。
AI检索提供了一个管理下的跨文件检索层和实时操作数据。 它剖析，块，嵌入和重新排序文档，而关于MariaDB数据的自然语言问题则成为只读的SQL查询，与直播数据库相对。 后续问题保留了先前的内容。 该设计旨在在不保留单独的仓库或矢量存储副本及其安全控制的情况下，地面解答。
开源 AI插件提供 MariaDB 专用指导，用于编码 agents. 他们支持Claude Code,Codex,OpenCode和Pi.
分析、集群和部署控制
MariaDB Exa现在一般可以在房地内使用，在没有单独的仓库或ETL管道的情况下，直接对照现场操作数据运行实时分析。 矢量搜索仍然是核心数据库引擎的原生；MariaDB报告的矢量查询每秒最多比企业服务器11.8的准确度多30%，没有计划或查询变化。
就多区域部署而言，一般可用的Raft集群在大多数节点确认后会执行写法，防止一个缓慢或遥远的节点为每个写法设定速度。 发现和解决冲突让每个区域为当地更新服务，在两个地点改变同一记录的情况下，由客户界定的业务规则解决案件。
企业管理器将数据库，Max Scale和GridGain合并在一个可观察控制台上，与单人签名。 带上你的自有云也可以普遍获得：MariaDB云数据平面运行在客户的AWS,Microsoft Azure或Google云账户内，而MariaDB则运行数据库。
基于作用的准入控制适用于交易、分析和载体工作量。 FIPS 140-3支持通过企业号Kubernetes操作器启用，使用来自Red Hat的NIST验证加密模块。
基准测试条件
2026年10月的McKnight Consulting Group报告在头条吞吐量数字的背后，在其顶峰时测得企业服务器每分钟2,653,370份新订单12.3.3-0，而11.8.3-1的订单为1,702,758份。 12.3版本达到152个虚拟用户高峰；11.8版本达到88个高峰。 在相应的峰值，99百分位新秩序响应时间为3.60毫秒和6.54毫秒。
两个版本在同一个Google Cloud c4d-himm-192-lssd实例上分别运行，共96个物理核心和192个vCPU. 内膜锤子DBTPROC-C的工作量使用了768个仓库，来源于TPC-C，但并不是官方的TPC-C结果。
这些控制隔离了发动机的可扩展性，但其他选择限制了数字如何直接转移到生产。 永久承诺、二进制记录和双写缓冲被禁用。 MariaDB选择了自己的配置，而McKnight选择了测试，配置了数据库和测试应用程序，并运行了测试。
企业平台2027将马里亚DB的agent策略与更少的数据复制，共享访问控制和共同操作基础设施联系起来。 其公布的性能证据支持了在所披露的试验条件下较高的吞吐量；生产问题是，当耐久性、工作量组合和区域要求与所控制的环境不同时，合并平台如何运作。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
MariaDB Launches Enterprise Platform 2027 for Agent Workloads
Add Unite.AI to your preferred sources on Google
MariaDB announced Enterprise Platform 2027 on November 4, 2026, making its combined transactional, analytics, in-memory and AI platform generally available. The release adds agent memory, retrieval and response-caching tools, and the company reported 56% higher peak throughput for Enterprise Server 12.3 than version 11.8 in sponsored testing.
The release packages Enterprise Server 12.3 with Exa, GridGain, MaxScale, Enterprise Manager and the Enterprise Kubernetes Operator. It can run on premises, in any cloud or as a managed MariaDB Cloud service. A redesigned logo, color palette and website accompany MariaDB’s effort to position itself as a unified data platform rather than only an operational database.
Agent Tools Use the Same Live Data
A launch post divides the agent features into development and runtime functions. Agent Memory stores an agent’s state in memory after every step and shares it across agents and sessions. An agent paused for human approval or interrupted can resume from its saved state rather than restart; LangGraph is supported out of the box.
Semantic Cache addresses repeated model calls by recognizing when a differently worded question has the same meaning as an earlier one and returning the stored response. MariaDB says this can shorten response times and reduce token costs for applications whose users ask similar questions.
AI Retrieval provides one managed retrieval layer across documents and live operational data. It parses, chunks, embeds and reranks documents, while natural-language questions about MariaDB data become read-only SQL queries against the live database. Follow-up questions retain earlier context. The design is intended to ground answers without maintaining separate warehouse or vector-store copies and their security controls.
The open-source AI Plugins supply MariaDB-specific guidance to coding agents. They support Claude Code, Codex, OpenCode and Pi.
Analytics, Clustering and Deployment Controls
MariaDB Exa is now generally available on premises, running real-time analytics directly against live operational data without a separate warehouse or ETL pipeline. Vector search remains native to the core database engine; MariaDB reports up to 30% more vector queries per second than Enterprise Server 11.8 at the same accuracy, with no schema or query changes.
For multi-region deployments, the generally available Raft Cluster commits a write after a majority of nodes confirm it, preventing one slow or distant node from setting the pace for every write. Conflict detection and resolution let each region serve local updates, with customer-defined business rules settling cases where two sites change the same record.
Enterprise Manager combines the database, MaxScale and GridGain in one observability console with single sign-on. Bring Your Own Cloud is also generally available: the MariaDB Cloud data plane runs inside a customer’s AWS, Microsoft Azure or Google Cloud account while MariaDB operates the database.
Role-based access controls apply across transactional, analytical and vector workloads. FIPS 140-3 support is enabled through the Enterprise Kubernetes Operator using NIST-validated cryptographic modules from Red Hat.
Benchmark Conditions Define the Performance Claim
The October 2026 McKnight Consulting Group report behind the headline throughput figure measured 2,653,370 new orders per minute for Enterprise Server 12.3.3-0 at its peak, compared with 1,702,758 for 11.8.3-1. Version 12.3 peaked at 152 virtual users; version 11.8 peaked at 88. At those respective peaks, 99th-percentile New Order response times were 3.60 milliseconds and 6.54 milliseconds.
Both versions ran separately on the same Google Cloud c4d-highmem-192-lssd instance with 96 physical cores and 192 vCPUs. The in-memory HammerDB TPROC-C workload used 768 warehouses and was derived from TPC-C, but was not an official TPC-C result.
Those controls isolate engine scalability, but other choices limit how directly the numbers transfer to production. Durable commits, binary logging and doublewrite buffering were disabled. MariaDB selected its own configurations, while McKnight chose the tests, configured the databases and testing applications, and ran the tests.
Enterprise Platform 2027 ties MariaDB’s agent strategy to fewer data copies, shared access controls and common operational infrastructure. Its published performance evidence supports higher throughput under the disclosed test conditions; the production question is how the combined platform performs when durability, workload mix and regional requirements differ from that controlled environment.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.