政策 Serverless 双轨计费
- 一句话
- On-demand 按扫描字节计费(闲置零成本)+ Enterprise flat-rate slots 封顶可预测负载——同 workload 实测比 Snowflake 便宜 71%。On-demand billing per byte scanned (zero idle cost) + Enterprise flat-rate slots that cap predictable loads — measured 71% cheaper than Snowflake on the same workload.
- 窄场景
- 分析师驱动的突发、不可预测查询(白天 9-5 有人、夜间周末接近零);可预测的持续负载(定时 ETL + 稳定 BI)切 flat-rate。GCP 原生、无专职 FinOps 的中小数据团队——"不用调仓库大小、不用管自动 suspend"本身就是生产力。Analyst-driven, bursty, unpredictable queries (people 9-to-5, near zero nights/weekends); predictable continuous loads (scheduled ETL + steady BI) moved to flat-rate. GCP-native small-to-mid data teams with no dedicated FinOps — "no warehouse sizing, no auto-suspend babysitting" is itself productivity.
- 机制
- on-demand 按扫描字节($6.25/TB)而非计算时长计费,列存只收实际碰的列;slot 模式是 Borg 集群上的可抢占容量池。关键在闲置成本为零:Snowflake 按 warehouse 按秒计费(auto-suspend 只能缓解不能消除),Databricks 按 DBU+EC2 计费。对"查询量波动大、夜间无人"的 workload,这是结构性成本优势,不是调优能追平的。On-demand bills per byte scanned ($6.25/TB), not compute time — columnar storage means you only pay for the columns you touch; the slot mode is a preemptible capacity pool on Borg. The structural edge is zero idle cost: Snowflake bills warehouses per second (auto-suspend only softens it), Databricks bills DBU+EC2. For volatile workloads this is structural, not something tuning can close.
- 生产验证
- LeanOps 独立第三方实测(2026):某 SaaS Snowflake 账单 $108K/月,同 workload BigQuery 混合计费 $31K/月,省 71%、年省 $92.4 万,查询延迟因重做分区反而改善(https://leanopstech.com/blog/snowflake-vs-bigquery-vs-databricks-vs-redshift-cost-2026/);SmarterX 从 Snowflake 迁 BQ 成本减半(Google 官方博客,厂商口径)(https://cloud.google.com/blog/products/data-analytics/smarterx-migrating-to-bigquery-from-snowflake-cut-costs-in-half);Bonnier News 选 BQ over Snowflake,存储 $15/TB/月 vs $23/月(https://medium.com/bonniernewstech/why-we-picked-google-bigquery-over-snowflake-as-our-new-data-warehouse-solution-dddf7a441d5f)。LeanOps independent third-party benchmark (2026): one SaaS company's $108K/month Snowflake bill vs $31K/month on BigQuery's mixed model for the same workload — 71% saved, $924K/year, with query latency actually improving after re-partitioning (https://leanopstech.com/blog/snowflake-vs-bigquery-vs-databricks-vs-redshift-cost-2026/); SmarterX halved costs migrating from Snowflake to BigQuery (Google Cloud official blog — vendor-sourced) (https://cloud.google.com/blog/products/data-analytics/smarterx-migrating-to-bigquery-from-snowflake-cut-costs-in-half); Bonnier News chose BigQuery over Snowflake — storage $15/TB/month vs $23 (https://medium.com/bonniernewstech/why-we-picked-google-bigquery-over-snowflake-as-our-new-data-warehouse-solution-dddf7a441d5f).
- 竞品差距
- Snowflake warehouse 空转成本,24/7 连续 ETL 贵 3 倍;Databricks DBU+EC2 双重计费,同 workload 贵约 18-27%;Redshift provisioned 24/7 计费,只有 RA3+RI 在"连续可预测 ETL"这个窄场景能反超。"闲置零成本 + 按扫描字节计费"的组合 31 款中没有第二家。Snowflake's warehouse idle cost makes 24/7 continuous ETL 3x more expensive; Databricks' DBU+EC2 double billing is ~18-27% more on the same workload; Redshift provisioned bills 24/7 — only RA3+reserved pricing beats BigQuery in the narrow "continuous predictable ETL" case. The "zero idle cost + pay-per-byte-scanned" combo has no second product among the 31.
- 证据等级
- 独立实测(LeanOps 2026,同 workload 四平台重跑)+ 社区共识;SmarterX 为厂商口径Independent benchmarks (LeanOps 2026, same workload replayed on four platforms) + community consensus; SmarterX is vendor-sourced
- 最后核验
- 2026-10-012026-10-01