大家好，我是小耶，写功课只是为了我踩过的坑，你们别再踩了！
金仓发布会，大多数人的注意力都在新发的 AI 数据库上。我刷新闻的时候，反而被后面一段话勾住了——电科金仓在会上正式发布了全球化战略。
这条消息在朋友圈里没什么水花。一家国产数据库说要全球化，听起来像每年都会说一遍的场面话。但我今年 6 月在香港科技园听过他们一场沙龙，当时现场有些细节，让我觉得这事没那么虚。
一、先说说那场被 AI 抢了风头的发布
发布会主题叫「Hello，AI 数据库」，主角是 KES AI 数据库。全球化这件事，在新闻稿里只占了一小段，信息量倒是不小：
- 全球化战略正式发布
- 之前已经在港澳市场、东南亚等地做过业务实践
- 愿景是「成为世界卓越的数据库产品与服务提供商」
最后这句话，五年前的国产数据库厂商不太敢写。那时候大家的目标很实在——先把国内的信创单子拿下。
总裁杜胜的说法是「持续打磨世界级产品与服务体系」。这话听着像公关稿，但要判断真假，得看支撑它的东西是什么。所以我往前翻了半年。
二、今年 6 月，香港科技园那场沙龙
那场活动叫「KING 大咖面对面」，主题是「不止于兼容，融合多模构筑坚实数智底座」。我当时的身份是听众，坐后排，记了满满两页。
有几个片段我到现在还记得。
LSCM 大湾区科研项目及业务发展总经理孙丽霞讲《数据库在本地化适配测试实践介绍》。她开口第一句就很实在：很多产品在内地跑得挺好，一到香港就各种水土不服。不是数据库本身不行，是软硬件环境、合规要求、系统架构都不一样。她把这套本地化适配拆成了一套可复用的测试方法，讲的全是真项目里踩出来的坑。
还有一位香港本地的交付经理，来自 Automated System Ltd，叫 Matthew Wong。他复盘了几个香港金融、政企、商贸行业的落地项目，从评估、迁移割接、上线运维讲到长期优化。他说了一句话，我原样抄在本子上：
「我们不是拿国产库去跑边缘系统，是拿它跑核心业务。」
这句话比任何性能数据都狠。跑边缘系统和跑核心业务，是两个世界的事。
当时我以为这只是个别项目的说法。现在回头看，那更像是一次提前演练——四个月后，全球化战略正式发布。
三、出海真正的门槛，不在 TPC-C 分数上
国内选型，大家比的是「能不能替代 Oracle」。海外不一样。海外客户不缺数据库，他们缺的是「能塞进我这套跑了二十年的系统里、还不给我添麻烦」的东西。
我梳理下来，门槛大概是三道。
第一道：兼容，是门票，但不是卖点
海外的遗留系统比国内更杂。欧洲的老系统还在 Oracle 上，互联网业务全是 MySQL 生态，一些云原生团队用 PostgreSQL，还有些地方留着 SQL Server。你想进去，只有一条路：先把门进得去。
KINGbaseES V9 的做法是初始化时就能选兼容模式，原生兼容 SQL 标准、Oracle、MySQL、SQL Server 和 PG 生态。金仓技术博客里给过的数据是：典型迁移场景下，Oracle 对象兼容率 95% 以上，PL/SQL 兼容率超过 90%。这个数字意味着大量存储过程可以少改甚至不改。
MySQL 兼容这条线，在出海语境下比在国内重要得多。MySQL 是全球部署量最大的开源数据库，驱动、ORM、监控工具、云厂商托管服务全都围着它转。你兼容了它，等于不用让几千上万个海外开发者重新学一门数据库方言——这是最贵的迁移成本，而且不在报价单上。
配套的工具也得跟上，不然兼容率再高也落不了地。金仓这套是 DTS、DMS 做评估和对象转换，KFS 做全量搬迁加增量追平，走的是「评估→改造→全量迁→增量追平→灰度切流」的标准路径。这个流程国内项目已经跑了很多遍，出海就是换个环境再跑。
但兼容只是进门。金仓解决方案专家付义在沙龙上说得挺清醒：兼容只是基础门槛，真正的竞争力在别处。这话我认同，因为兼容只能让你「不输」，赢不了单子。
第二道：合规和供应链可信，才是硬门槛
这一条国内很多技术人容易忽略，但海外客户特别在意。
他们会问一些在国内很少被追问的问题：你的内核代码是谁写的？用了哪些开源组件？构建过程能不能追溯？出了 Log4j 这种事，你多久能给出补丁和影响面清单？
这不是刁难。全球对软件供应链安全的立法都在收紧，2026 年 3 月国内也施行了《国务院关于产业链供应链安全的规定》。海外关键行业客户在选型时，已经不只看基准测试分数，还要厂商拿得出全生命周期的安全证明。
金仓在这块的底子是：内核 100% 自主研发，核心代码自主可控；KingbaseES V9 通过了中国软件评测中心的软件供应链安全能力三级认证（厂商方面称是国内数据库厂商里首个拿到这个等级的）。这两个东西平时不显眼，但到了国际招投标、到了合规审计那一步，就是能不能进短名单的差别。
数据主权是另一半。不同国家对数据出境的要求天差地别，细粒度权限、透明加密、全链路审计这些能力，在国内是加分项，在海外是入场条件。
供应链信任这个东西没法速成，只能靠时间和一次次交付攒出来。这也是我认为国产数据库出海最慢的一环。
第三道：本地化服务，决定你能不能活下去
产品卖出去只是开始。香港那场沙龙上，圆桌环节聊得最实在的是这一块：本地客户对外部供应商的信任建立周期很长，需要一个个真实案例去支撑；香港对数据安全和合规的要求跟内地不同，需要更细的权限管理和审计能力。
孙丽霞讲的那套本地化适配测试方法，本质上是在解决「交付前一公里」的问题——跨境部署、原有业务迁移里高频出现的适配故障，怎么提前测出来。
服务网络也是硬功夫。跨时区、跨地域的 7×24 响应，集中化的统一监控和运维（金仓这边是 KEMCC），这些听着不性感，但客户半夜出事的时候，只认这个。
四、那金仓手里真正值钱的牌是什么？
如果只打「兼容 Oracle + 价格便宜」，海外市场其实没什么机会。Oracle 的替代品、PG 的发行版，那边一抓一大把，而且比你有本地根基。
真正能拿出去打的，是这次发布会上的东西——AI 原生加多模融合。这两件事在海外语境下，比在国内更值钱。
发布会上讲的三个能力，都在库内：库内语义构建、库内计算推理、库内原生安全管控。核心一句话叫「数据不出库，AI 也能用」。实测数据是 1 亿条 768 维向量、98% 召回率条件下，峰值内存开销比业界方案降低 95.4%，查询延迟降低 80%。
内存降 95% 这个数字的实际含义，是从「要堆 TB 级内存才跑得动」变成「常规服务器就能跑」。对海外客户来说，这不只是省钱，是能不能在自己机房里落地的区别。
多模这块，香港沙龙上萧少聪现场演示过——一条 SQL 同时操作关系数据、向量数据、JSON 半结构化数据和图数据。传统做法是四套库、四个接口、四次数据搬运，还得处理一致性问题。
为什么这套东西在海外更好卖？两个原因。
一是海外企业的信息化历史长，系统杂，技术栈收敛是刚需。他们不缺单点产品，缺一个能把关系、文档、时序、向量收进同一个内核的底座。
二是 GDPR、数据主权这类要求越来越硬，「数据不出库」从技术卖点变成了合规刚需。你把向量库挂在外面，数据就已经出过一次边界了。
五、冷静一点：这条路还很长
我不想把话说得太满。
这个赛道已经有人在跑，而且跑得不慢。OceanBase 2025 年 10 月就宣布全面提速全球化，港澳、东南亚之外还进了日本市场，当时公开的数字是深度接洽 70 多家海外客户、40 家签约付费。国产数据库出海不是金仓一家的故事，是一场集体补课。
差距也是明摆着的：生态。驱动、社区、云市场、ISV 适配、大学课程，这些东西国外厂商攒了二三十年。金仓自己的技术博客里也写了判断——未来 1 到 3 年，国产数据库出海要从「可用」验证期进入「好用」和「可信」的深水区。这个说法我觉得是诚实的。
还有一个更根本的问题：全球化最难的不是把产品运出去，是让别人愿意把核心系统交给你。这跟跑分无关，跟十年二十年的口碑有关。
六、如果你在做选型，可以问三个问题
不管你是出海企业的技术负责人，还是在评估国产数据库的外企团队，我觉得这三个问题比看参数表有用：
第一，多模是内核融合还是拼装？ 让它演示一条 SQL 里同时做关系过滤和向量计算。如果是调外部接口拼出来的，那和外挂方案没本质区别，只是包装得更整齐。
第二，供应链证据链拿得出来吗？ 组件清单、构建可追溯、安全认证、漏洞响应 SLA。拿不出来，国际项目基本走不到技术评审那一步。
第三，目标市场有没有本地交付和真实案例？ 不是把总部案例翻译成英文，是有本地团队、本地客户、本地运维。这一条最容易被跳过，也最容易在项目中途翻车。
七、小结
把两件事放在一起看就清楚了：slogan 从「数据库平替用金仓」变成「AI 数据库用金仓」，是产品定位的升级；全球化战略发布，是市场定位的升级。
本质上是一件事——不再追别人的标准。
平替是「你有我也有」，AI 原生是「AI 时代数据库该长什么样，我来试」。国内打的是替代战，海外打的是定义战。后者难得多，但也只有后者，才配得上「世界卓越」这种说法。
至于能不能成，我不下结论。技术人的习惯是看证据：看明年海外签了多少单，看本地团队建了几个，看客户敢不敢把核心系统放上去。
这些比发布会上的话可靠。
小耶在手，SQL 不愁
还有什么想了解的，欢迎留言！小耶一定知无不言言无不尽……我们下次见~
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。
Hello, I'm a kid. I only write for the pit I stepped on.
Kimura publish, most of the attention is on the new AI database. When I painted the news, I was caught up in the following passage — the publish strategy of globalization was formalized at the conference by the electronics vault.
There's no water in the circle of friends. A database of national production calls for globalization and sounds like an annual talk. But I heard about them at a salon in Hong Kong Science Park in June of this year, and there were some details on the scene that made me think it wasn't that bad.
First, tell me about the publish that was robbed by AI.
The theme of the publish session is "Hello, AI Database", and the main event is KES AI Database. Globalization is only a small part of the press release, but the amount of information is not small:
- Official publish of the strategy of globalization
- I've done business practice in Hong Kong, Australia, South-East Asia, etc.
- Vision "to be a world-class database product and service provider"
Finally, five years ago, the national production database manufacturer was less daring to write. At the time, the goal was real — first, to take down the domestic letter list.
President Do-sung says "continue to grind world-class product and service systems." That sounds like a PR script, but it depends on what sustains it. So I went on for six months.
In June of this year, the salon at Hong Kong Science Park
The event was called "KING Face-to-Face" and the theme was "More than Compatibility, Combining Multimodels to Build a Solid Digital Background." I was a listener, sitting in the back row and remembering two pages.
There's a few pieces I still remember.
LSCM General Manager of Scientific Research Projects and Business Development in the Great Bay Region, Sun Lixia, gave a presentation on the Database ' s Practice in Localizing Adaptation Testing. Her first sentence is true: many of the products run well in the interior, and every kind of land and water comes to Hong Kong. It is not the database itself that does not work; it is the soft and hardware environment, the compliance requirements, and the architecture of the system. She made this local adaptation into a reusable test, all talking about pits from real projects.
There is also a local delivery manager from Automated System Ltd. called Matthew Wong. He re-opened several projects in Hong Kong's finance, business and trade sectors, ranging from assessment, migration harvesting and offline processing to long-term optimization. He said one thing. I copied it as it was:
"We're not taking the treasury to run the peripheral system, we're running the core business."
That's worse than any performance data. Running edge systems and running core business are two worlds.
At the time, I thought it was just a statement of individual projects. Now look back, it's more like an early exercise — four months later, the strategy of globalization is officially publish.
III. Real threshold for sailing, not on TPC-C scores
The choice is "can we replace Oracle?" Overseas is different. Overseas clients have no shortage of databases, and what they lack is something they can put into my 20-year-old system without causing me any trouble.
I combed it down, and the threshold was probably three.
First: Compatibility, tickets, but not points of sale.
Legacy systems abroad are more diverse than domestic ones. Europe's old system is still on Oracle, and the Internet is full of MySQL ecology, with some cloudy teams using PostgreSQL, and some places with SQL Server. You want to go in there, there's only one way: get the door in first.
The KINGbaseES V9 approach allows for the initialization of compatible models with original compatible SQL standards, Oracle, MySQL, SQL Server and PG ecology. The data given in the Kimakura technology blog are that the Oracle object compatibility rate is more than 95% and PL/SQL compatibility is over 90% in typical migration settings. This figure means that a large number of storage processes can be reduced or not.
MySQL compatible with this line is much more important in overseas than in the interior. MySQL is the largest open source database of global deployment, with driving, ORM, surveillance tools, cloud-manufacturing hosting all around it. You're compatible with it, which means you don't have to allow tens of thousands of overseas developers to redo a database dialect — the most expensive migration cost and not on the quotation.
The accompanying tools will have to be kept up, otherwise the compatibility rate will not fall. The gold silos are DTS, DMS for evaluation and object conversion, KFS for full relocation and incremental homing, and follow the standard path of "Assessing full retrofitment and retromovation for flat ash cutting". The domestic project has run many times, leaving the sea is a different environment.
But compatible just enters the door. The gold-kilo solutions experts were clear on Sharon: compatibility was only the basis threshold and true competitiveness was elsewhere. I agree, because compatibility only makes you "not lose" and you can't win the list.
Second: compliance and supply chain credibility is the hard threshold
Many skilled people in the country are easy to ignore, but overseas clients are particularly interested.
They'll ask questions that are rarely asked at home: who wrote your kernel code? What open source components are used? Is the construction process traceable? How long can you give me a list of patches and effects?
It's not hard. Global legislation on software supply chain security has been tightened, and in March 2026, the State Council Regulation on industrial chain supply chain security was introduced. In selecting models, key overseas industry clients have gone beyond reference test scores to produce full life-cycle safety certificates.
The base of the gold warehouse in this area is: kernel 100% self-research, core code self-control; KingbaseES V9 passed a Level 3 certification of software supply chain security capabilities at the China Software Assessment Centre (which the manufacturer claims was the first of its own database manufacturers). The two items are not always visible, but the difference between whether or not they are on the shortlist is the point at which international bids are made and compliance audits are made.
Data sovereignty is the other half. Requirements for data exits vary from country to country, with fine permission, transparent encryption, full-chain auditing capabilities, which are internal and overseas entry conditions.
The supply chain trust this thing cannot be achieved quickly, but only by time and delivery once. And that's the slowest part of what I think is going to happen to a database.
Third course: Localized services determine whether you can survive.
The product was sold out just to start. At the Hong Kong salon, the round table was the one where the trust of local clients in external suppliers was built over a long period of time and needed to be sustained by a real case; and where Hong Kong ' s requirements for data security and compliance were different from those of the Mainland, it required a more detailed permission management and auditing capacity.
The local adaptation test methodology, described by Sun Lixia, is essentially a solution to the problem of "one kilometre prior to delivery" — cross-border deployment, high-frequency adaptation failures in the migration of former operations, and how to detect them in advance.
The service network is also hard work. 7x24 responses across time zones, across geographical areas, centralized unified surveillance and transport (KEMCC on the side of the vault), which don't sound sexy, but this is the only thing that clients recognize when things happen in the middle of the night.
What's the real value card in the vault?
If you only play "Oracle + cheap", there's really little chance that overseas markets will be able to do. Oracle's alternatives, the PG's distribution, there's a big one over there, and it's more local than you.
The real thing that can be played out is this publish meeting -- AI multi-modular integration. These two things are worth more in overseas than in the country.
The three capabilities described at the publish meeting were all described in Couné: the Cuen semantic construction, the Coone reasoning for the calculation of the Cooney, and the control of the Cooney native security. The core sentence is "Data is not available, AI works." The measured figures are 100 million 768 dimensions, 98 per cent recall rates, 95.4 per cent of peak memory costs and 80 per cent of queries less than industry programmes.
The actual meaning of the 95% reduction in memory is to change from "to pile TB-level memory to "to run by a regular server." For overseas clients, this is not just a cost saving, but a difference as to whether they can land in their own rooms.
Domu, a live demonstration of a SQL operating relationship data, vector data, JSON semi-structured data and graphic data in Hong Kong's Sharon. The traditional practice is four repositories, four interfaces and four data moves, which also address consistency issues.
Why does this thing sell better overseas? Two reasons.
The first is the long history of computerization of overseas enterprises, the multiplicity of systems and the demand for technology storage. They do not lack single-point products, and they lack a base that captures relationships, documents, time series, vectors into the same core.
Second, GDP R, data sovereignty, etc., are becoming more and more demanding, and the "data non-bank" has changed from a technology selling point to a compliance requirement. You hang the vector bank outside, and the data has already crossed the border once.
V. Calm down: It's a long road.
I don't want to be too full.
This track is already running and not slow. In October 2025, OceanBase announced full-scale globalization, with Japanese markets outside Hong Kong, Australia and South-East Asia, with open numbers reaching more than 70 overseas clients and 40 contracted paid. The national production database is not a story of the Kinkura family, but a group session.
The gap is also clear: ecology. Drivers, communities, cloud markets, ISV appliance, university courses have been in place for 20 to 30 years. Kimura's own technical blog also writes about the judgment that, for the next 1 to 3 years, the national production database is going to sail from the "usable" validation period to the "good" and "credible" depths. I think it's honest.
There is also a more fundamental question: the hardest part of globalization is not to move its products out, but to let others give you the core system. It's not about running points. It's about 10 or 20 years of rhetoric.
Six: If you're making choices, you can ask three questions.
Whether you're the technical head of the offshore company or the team of foreign companies that evaluate the national production database, I think these three questions are more useful than looking at the parameter table:
First, is it kernel integration or assembly? Allows it to demonstrate a relationship filter and vector calculation in a SQL. If the external interface is spelled out, it is not materially different from the outer one, but is better packaged.
Second, can the chain of evidence in the supply chain come out? List of components, construction traceability, security authentication, leak response SLA. The international project will hardly be able to take the technical evaluation step.
Third, are there local deliveries and real cases in the target market? Instead of translating Headquarters cases into English, there are local teams, local clients and local services. This article is the most easily bypassed and the easier to cross the project.
Summary
It is clear by putting two things together: the shift of slogan from a "database parlour" to an "AI database warehouse" is an upgrade in product positioning; the globalization strategy publish is an upgrade in market positioning.
Essentially, it is one thing — to stop chasing standards.
Substitute is "you have me and you have me." Domestic warfare is a substitute and overseas is a definitional war. The latter is much more difficult, but only the latter is worthy of the expression “world excellence”.
I don't think it's possible. It is the habit of technicians to look at the evidence: look how many bills they sign overseas next year, see how many local teams have built, and see how many customers dare to put the core system on.
These are more reliable than what publish said.
Je's got it, SQL.
Anything else you want to know, please leave a message! I'm sure Yao knows everything... I'll see you next time.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.