跪拜 Guibai
← All articles
Database

Four Steps That Cut a Database Shortlist from 11 to 2

By 倔强的石头_ ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Any team facing a crowded database market can apply the same order: align constraints before comparing features. The compliance gate is specific to China's domestic IT push, but the principle of turning regulatory and operational reality into hard filters translates to any sector.

Summary

Database selection usually begins with a feature comparison table, but that approach led nowhere for a team picking a domestic replacement for its core system. After two weeks of benchmarking eleven products, the team switched to a four-step method that evaluates internal constraints first. Step one audits the source systems—Oracle, MySQL, and SQL Server—to define which compatibility details matter, from PL/SQL packages to system packages. Step two sizes the data: 3.2TB and under 4,000 concurrent connections place it squarely in centralized territory, despite industry numbers that push distributed adoption to 71% by 2026.

Step three applies a hard compliance gate for government and state-owned enterprise clients: Xinchuang certification, national cryptography algorithms, domestic CPU and OS support, cutting the list to five. Step four assesses the operational team—two DBAs with no distributed experience—making toolchain completeness and failover automation more valuable than raw performance. The surviving two candidates then go through a rigorous POC: top production SQL, migration assessment tools, mirrored traffic stress tests, failover drills, and hands-on DBA exercises.

The final pick, KingbaseES, passes all four steps. The method's core lesson is order: constraints decide what you can even consider, and the final comparison only ever involves two or three viable products.

Takeaways
— The initial candidate list of eleven products was cut to two after applying four steps: source compatibility, data scale, compliance, and team capability.
— Oracle source migration requires deep checks on PL/SQL packages, system packages (DBMS_JOB, DBMS_OUTPUT, DBMS_LOB), and specialty syntax like ROWNUM and CONNECT BY.
— A centralized architecture is sufficient when single-table rows stay under 10 million, total storage under 10TB, and daily queries below 10,000; the team's 3.2TB and under-4,000 concurrency fit.
— Signals for going distributed include real write bottlenecks, single tables over 100 million rows, online expansion needs, and RPO=0 with second-level failover requirements.
— Compliance screening in China includes security and reliability evaluation (23 products in the fourth list, II-level certifications rising from 1 to 6), SM2/SM3/SM4 cryptography, and domestic CPU/OS compatibility.
— A two-DBA team with no distributed experience prioritizes operational tooling–one-click monitoring, simple PITR, automated failover–over feature complexity.
— POC must use production top-50 SQL, migration assessment reports, mirrored production traffic, failover drills, and hands-on DBA tasks, not sysbench or idle benchmarks.
— A provincial grid ran 30TB of data on a centralized database for 14 years with 99.999% availability, showing centralized headroom well beyond 10TB.
— A ship-industry platform with 30 billion rows and 30 million daily peak writes uses a 5-node distributed cluster, achieving 150 million new writes per day and millisecond queries.
— KingbaseES won after passing all steps, with over 95% of stored procedures requiring minimal adjustment and full Xinchuang compliance.
Conclusions

Comparing features before aligning constraints wastes months; the team's initial half-month effort produced no decision, while the four-step method took three months to signing.

Compliance can eliminate more products than technology comparison; the technical merits of non-compliant products are irrelevant in government-facing projects.

The industry's 71% distributed adoption figure is skewed by new write-heavy IoT and edge workloads, which obscures the large centralized base in core transaction systems.

POC design changes the outcome: using production traffic and real DBA operations reveals operational fit that benchmarks and sales slides cannot.

Concepts & terms
Xinchuang (信创)
A Chinese government initiative to replace foreign IT products with domestic ones, making compliance a hard requirement for public sector sales.
Multi-model database
A database that supports multiple data models (relational, document, time-series, vector) in a single engine, reducing component count.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗