方法论
我们如何保证对比的中立与可用
覆盖粒度一致（中立性护栏）
每个档案都必须逐项回答固定的 47 个硬维度，给出"有 / 无 / 部分支持 / 未找到证据 / 查证为无"的结论并标注证据等级，不允许"因为不是特色就不写"。这条规则的存在理由：特性天然会被相对各自基线放大（"对社区 PG 而言 TDE 是加分项"），但同一能力在不同档案里覆盖粒度不同，就是偏颇。护栏保证的是覆盖一致，权重可以不同。
Every profile must answer the same fixed 47 hard dimensions, each with a verdict (supported / not supported / partially supported / no evidence found / verified absent) and an evidence level. Skipping a dimension "because it is not a highlight" is not allowed. The rationale: a feature always looks bigger against its own baseline, but inconsistent coverage of the same capability across profiles is bias. The guardrail guarantees consistent coverage; weights may differ.
选型权重自定：不做综合总分、不做排名
本站不做综合总分、不做排名。对比页只做维度级并排，不做跨库加权。原因是：权重取决于你的场景——金融核心系统里"复制与一致性"的权重可能是初创 MVP 的十倍；任何替你定权重的做法都是在替你做决定。正确用法：先列出你自己的必备维度与否决项，再到对比矩阵里按行（产品）逐项核对。
This site produces no aggregate scores and no rankings. The comparison page only places dimensions side by side; it never computes a cross-database weighted total. The reason: weights depend on your scenario — "replication & consistency" may weigh ten times more for a financial core system than for a startup MVP. Anyone who sets the weights for you is making your decision for you. The right way to use this site: list your own must-have dimensions and deal-breakers first, then check them row by row (per product) in the matrix.
证据分级
档案中每条关键结论都标注证据等级（彩色徽章）：
- 官方文档：来自厂商官方文档、论文或 release notes，可信度最高，但注意"官方口径"可能只讲上限；
- 厂商口径：厂商发布、新闻稿、定价页中的说法，未经独立验证，引用时请打折；
- 社区实测：来自生产事故报告、benchmark、故障复盘等一手实践，有具体版本与场景才可信；
- 社区共识：多份独立社区资料一致指向的结论，单篇博客不算共识；
- 待验证：信息不全或存在冲突，明确标出，不写进判决依据。
AI 顾问（待接入）在引用档案回答时，必须原样携带证据等级徽章。
Key claims in every profile carry an evidence level (colored badge):
- 官方文档: from vendor documentation, papers or release notes — the most trustworthy, but note that official statements may only describe the ceiling;
- 厂商口径: from vendor announcements, press releases or pricing pages — not independently verified, discount accordingly;
- 社区实测: from production incident reports, benchmarks or postmortems — only credible with a concrete version and scenario;
- 社区共识: pointed to by multiple independent community sources — a single blog post is not consensus;
- 待验证: incomplete or conflicting information, explicitly marked, never used as a basis for verdicts.
The AI advisor (pending integration) must carry these badges unchanged whenever it cites profile evidence.
“未找到证据”不等于“不支持”
档案中"未找到证据"只表示：在本次信息采集范围内没有找到可靠出处，不能反推出"该产品不支持"。同样，"查证为无"是经过核实确认缺失的结论，证据效力完全不同。做选型否决项判断时，请区分这两者；拿不准的，请以官方文档最新版本为准，或在 POC 中实测验证。
"No evidence found" in a profile only means no reliable source was found within this research pass — it must not be read as "not supported". "Verified absent", by contrast, is a confirmed gap. These two carry completely different evidential weight when you are judging deal-breakers. When in doubt, check the latest official documentation or verify with a POC.
信息时效
全部档案信息截至 2026-09-29，各档案头部标注了评审版本。数据库迭代快，结论会过期：看到"深水区"里的版本号与调优参数时，请先对一下你手上的版本再照抄。
All profiles are current as of 2026-09-29, and each profile header notes the reviewed version. Databases iterate fast and conclusions expire: before copying tuning parameters or workarounds from the "deep dives", check them against the version you actually run.
招牌能力 vs 客户经验：两种视角
每个档案有两个并列大区，回答的是两个不同的问题：
- 招牌能力：厂商挂在门口的招牌——架构级、官方主打的硬核能力；
- 客户经验：用户用生产实践投票选出来的、真正离不开的能力，含 [避坑] 标签的反向招牌（杀手级代价）。
研究方法：来源优先级为社区实践 > 迁移总结 > 故障复盘 > 独立基准 > 官方实践指南 > 官网营销。官网首页营销几乎不产生有效发现，但可作为"厂商吹 vs 用户用"的错位检测器。每个能力标注窄场景、机制解释、生产验证（含来源链接）、竞品差距与证据等级；证据不足就降级或不收录，绝不硬凑。两区重叠是最强共识，分歧处（只有招牌没有经验，或反之）往往是最有价值的信息。
Each profile carries two parallel sections answering two different questions:
- Signature strengths: what the vendor puts on the signboard — architecture-level, officially marketed capabilities;
- Customer experiences: what users actually cannot live without, validated by production practice — including reverse signatures tagged [Pitfall] (killer costs).
Research method: source priority is community practice > migration write-ups > postmortems > independent benchmarks > official guides > homepage marketing. Homepage marketing yields almost no valid findings, but works as a mismatch detector between vendor claims and user reality. Each capability notes its narrow scenario, mechanism, production evidence (with source links), gap vs rivals, and evidence grade; weak evidence is downgraded or dropped, never padded. Overlap between the two sections is the strongest consensus; divergence (signboard-only or experience-only) is often the most valuable signal.