Four Steps That Cut a Database Shortlist from 11 to 2
Four-Step Database Selection: From a Dozen Candidates to Two
At the end of last year, the company asked me to lead the database selection for the core system, with the direction of switching to a domestic database. When I took over, the candidate list had eleven products, and the documentation piled half a person high. Someone in the team suggested building a comparison table, comparing features, benchmarks, and compatibility rates one by one. We followed that approach for half a month, and the more we compared, the messier the conclusion became. Every product had a pile of advantages, and no one could convince anyone else.
Later, I changed the method. Look at yourself first, then at the product. Before aligning the constraints, comparing products is a waste of time. After four steps, the eleven candidates were narrowed down to two. This article writes down the complete method.
@[toc]
Step 1: Look at the source system, determine the compatibility direction
The first thing in selection is to thoroughly understand the current system's baseline. What the source database is determines which direction you should verify compatibility.
Our source systems are somewhat mixed. The core runs on Oracle, with several MySQL databases on the periphery, and an old system on SQL Server. Three source types, three ways to verify. This step took a week to thoroughly inspect the old system. During the inspection, there was a finding worth mentioning. Many compatibility issues are hidden in the corner code that "nobody uses often." The longer a module hasn't been touched, the more it needs to be reviewed.
Oracle source: The key verification goes beyond just SQL syntax. The real deep water is here:
- PL/SQL objects: stored procedures, functions, packages, triggers, custom types. Among them, "packages" are the easiest place to trip up. Many domestic databases support stored procedures, but package support is incomplete. We have over forty business packages, and each one has to be checked.
- Oracle-specific syntax: ROWNUM, CONNECT BY, MERGE INTO, DECODE, NVL — these are everywhere in the business code.
- System packages: DBMS_JOB, DBMS_OUTPUT, DBMS_LOB, etc. The business code uses them heavily. Missing any one is a problem.
- Data type details: NUMBER precision mapping, whether VARCHAR2 length is calculated by bytes or characters, behavioral differences between DATE and TIMESTAMP.
After this round of inspection, our conclusion is that Oracle compatibility capability is the first screening criterion. In public material, KingbaseES (hereinafter referred to as KES) supports PL/SQL packages, autonomous transactions, cursor loops, and other features relatively completely. In actual migration projects, over 95% of stored procedures run directly with minor adjustments. The direction is right, but the specific numbers still need to be verified by yourself in POC.
MySQL source: Focus on protocol compatibility. Can the migration be completed only by replacing the JDBC driver? Does the CDC tool recognize the Binlog format? Are there dialect issues in SQL generated by MyBatis?
SQL Server source: Check T-SQL syntax and system function compatibility, and alternatives for surrounding tools like SSIS and SSRS.
The key points from this step were organized into a diagram:
Step 2: Look at scale, determine the architecture path
Once the compatibility direction is set, the second step is to "verify the true identity" of your data volume and concurrency. Don't make a mountain out of a molehill, and don't overextend.
The industry's empirical guideline is: when a single table has fewer than 10 million rows, total volume is under 10TB, and daily queries are under 10,000, a centralized architecture is usually sufficient and has the lowest operational cost. Our core database actually has 3.2TB of data and a peak concurrency of less than 4,000. No matter how you calculate it, we don't reach the threshold for distributed.
I also note the signals for "when to consider distributed" so I don't forget to check in two years. Real write bottlenecks, a single table exceeding 100 million rows causing DDL and backup windows to fail, clear online expansion requirements, and a high-availability hard requirement of RPO=0 plus second-level failover. If none of the four signals are present, feel comfortable using centralized.
Let me add an industry figure to prevent someone from persuading you with "the trend." One industry analysis puts the adoption rate of distributed systems among new systems in 2026 at about 71%. But this number needs to be broken down. Among new scenarios, a large proportion are write-heavy businesses naturally suited to distributed, such as IoT collection, logging, and edge nodes, which inflate the ratio. The existing centralized base of core transaction databases is much larger than this number suggests. Others have their reasons for going distributed. Your reason has to be found in your own four signals.
There is another new variable to mention: multi-model. Our company, besides relational data, also has document-type customer service records and vector retrieval demanded by the AI team. Should we configure a dedicated database for each of these needs, or find a product that can house relational, document, time-series, and vector data? Converged multi-model architectures like KingbaseES V9 offer the latter option. Five data models in one kernel, and the product line covers centralized, distributed, and shared-storage modes. For a small team like ours with tight operational staffing, one less component means one less scheduling burden.
Beyond centralized and distributed, cloud-native and converged multi-model routes are also often brought to the table. A comparison table makes the boundaries of the four routes clear at a glance:
I also turned the judgment basis of step two into a self-check list. On one side are our own numbers, on the other are the four signals:
Step 3: Look at compliance, remove non-qualified products
This step has the highest filtering efficiency and is also the easiest for technical people to overlook.
Many of our downstream customers are in government and state-owned enterprises. Xinchuang (domestic IT innovation) is a hard gate, not a bonus point. The checklist to verify: security and reliability evaluation. After the fourth list was released in May 2026, 23 products were selected, and the number of II-level certifications rose from 1 to 6. National cryptography algorithm certification: SM2/SM3/SM4. Security protection level. Compatibility with domestic CPUs: Kunpeng, Feiteng, Hygon, Loongson. Compatibility with domestic operating systems: Kylin, UnionTech UOS.
After this round, the list was cut from eleven to five. To be honest, the previous five steps are more effective than any feature comparison table. A non-compliant product, no matter how good its features, cannot enter the bid.
Step 4: Look at the team, don't pick something you can't handle
The last step is to honestly evaluate your own people.
Our reality is two DBAs, no distributed operation experience, and only these two people are on night duty. What does that mean? It means the product with the most features and the most complex architecture is actually a negative asset for us. The completeness of the operation toolchain, the automation level of failover, and the ease of backup and restore operations carry more weight here than performance benchmarks.
What specific items does this step examine? I went through them one by one later, and I've copied the list here:
For a team of two DBAs, look at three things when choosing a product. Can the monitoring system show slow queries and replication lag in one click? Is PITR doable with a few button clicks? Does failover require manually running scripts? Kingbase's toolchain covers these capabilities relatively well. That's one reason it stayed in the candidate list.
This step also has a hidden benefit. After putting team capabilities on the table, the negotiation posture with vendors changes. When sales talk about advanced distributed, multi-model, multi-active capabilities, we just say, "First demonstrate your monitoring and one-click failover." The conversation immediately returns from PPT to the features we actually use. Knowing your boundaries keeps you from being led astray in negotiations.
POC: The Final Arbiter
After the four steps, two candidates remain. Don't make the final decision yet. POC must be done, and it must be done with the right method. We'll list the pitfalls we fell into in advance:
- Don't replace real business SQL with sysbench. Run the top 50 production SQL statements one by one, comparing execution plans and actual latency. That's the real test.
- Run a migration assessment tool on the source database first, generating a compatibility report. Focus on the automatic conversion rate of stored procedures, packages, and triggers. Every item marked red in the report is a debt you'll have to pay later.
- Use production traffic mirroring for stress testing, at least at the level of real business peak volume. Idle benchmarking is meaningless; you need to see performance under load.
- Failover drill: Simulate the primary node going down, time the failover, and check data consistency. Don't trust the RTO numbers on the PPT.
- Let the DBAs perform a round of operations themselves: backup, restore, expansion, slow query troubleshooting. Keep notes on where things get stuck. POC is the only chance to have the vendor accompany you through trial and error.
Two Comparable Reference Cases
The four-step method sounds abstract. Fortunately, there are numbers from two public projects that serve perfectly as references. All figures are from public sources, reproduced below.
One is a centralized example. A provincial power grid's measurement and control business covers 26 provinces and more than 700,000 sensors, with 30TB of legacy data. The entire system has run continuously on a domestic database for 14 years, achieving five nines of availability. As you can see, with 30TB of data, they didn't go distributed; they ran on centralized for 14 years. That's the other side of the "centralized is enough below 10TB" guideline. Even exceeding it a bit, centralized is far from the ceiling.
One is a distributed example. A ship industry regulatory platform has hundreds of thousands of ships and 200,000 terminals reporting data. Daily peak writes are 30 million records, with a total of 30 billion records stored. A 5-node distributed cluster writes 150 million new records per day and answers queries in milliseconds. This kind of write throughput and volume is the true home of distributed.
Text alone is not intuitive. I've transcribed the numbers from both cases into a comparison table:
| Comparison item | Provincial power grid · measurement and control | Ship industry · regulatory platform |
|---|---|---|
| Coverage scale | 26 provinces, 700,000+ sensors | Hundreds of thousands of ships, 200,000 terminals |
| Data volume | 30TB stored | 30 billion stored, 30 million daily peak writes |
| Architecture | Centralized, running for 14 years | 5-node distributed cluster, 150 million new writes per day |
| Performance result | 99.999% availability | Millisecond query response |
| What it shows | 30TB and beyond, centralized still far from ceiling | This throughput is where distributed truly belongs |
In other words, the criterion isn't "is distributed new or not," it's which side your scale falls on.
Putting the two cases together makes the point of step two intuitive. Scale determines architecture, not trends. Our 3.2TB sits between these two cases, and it's obvious which side we're on.
Result and Review
The eventual winner was KingbaseES, passing all four steps. The depth of Oracle compatibility meant most of our forty-plus packages could be automatically converted. The centralized architecture is manageable for two DBAs. The Xinchuang checklist was all green. In POC, the compatibility of the top 50 SQL was also the most stable. On the signing day, exactly three months had passed since the project was initiated.
Looking back at the half-month detour of "comparing features in a table," the problem wasn't diligence, it was order. The first step of selection isn't to look at products, it's to look at yourself. The source type determines the compatibility direction. Data scale determines the architecture route. Compliance gates remove non-qualified products. Team capability determines which one can finally land. After four steps, you'll find that there are never more than two or three products truly worth serious comparison.