政策 RA3 存算分离 + 预留价:AWS 锁定团队的成本最优解
- 一句话
- RA3 节点(本地 SSD 缓存 + S3 managed storage)配合 1/3 年预留(最高 56% off)——在"7×24 连续 ETL + 稳定 BI" workload 上实测比 Snowflake 便宜 5 倍。RA3 nodes (local-SSD cache + S3 managed storage) with 1/3-year reservations (up to 56% off) — measured 5x cheaper than Snowflake on "7x24 continuous ETL + steady BI" workloads.
- 窄场景
- 连续 ETL 管道(每 15 分钟跑一批、24/7 不停)+ 每小时刷新的稳定 BI;深度 AWS 锁定(已有 EDP 折扣、Lake Formation/Glue 资产)、数据量大且持续增长、不想为"存储"多买"计算"的团队。Continuous ETL pipelines (a batch every 15 minutes, 24/7) plus hourly-refreshed steady BI; deeply AWS-locked teams (existing EDP discounts, Lake Formation/Glue assets) with large, growing data that don't want to buy "compute" just to get "storage".
- 机制
- RA3 把 ParAccel 血统的 MPP 计算层和 S3-backed managed storage 解耦:热数据在节点本地 SSD 缓存,冷数据走 S3;扩存储不再需要加节点(DC2 时代"为存数据被迫买计算"的痛点消失)。provisioned 节点 24/7 计费看似笨拙,但对"本来就要 24/7 跑"的连续 ETL,1 年 RI(33% off)/3 年 RI(56% off)后的单价是全场最低。RA3 decouples the ParAccel-lineage MPP compute layer from S3-backed managed storage: hot data sits in node-local SSD cache, cold data goes to S3; adding storage no longer requires adding nodes (the DC2-era pain of "buying compute to store data" is gone). Provisioned 24/7 billing looks clumsy, but for continuous ETL that was going to run 24/7 anyway, 1-year RI (33% off) / 3-year RI (56% off) yields the lowest unit price in the field.
- 生产验证
- LeanOps 独立第三方建模(2026,5 月定价):Workload B(200TB 存、100TB 查/月、15 分钟一次 ETL、100 用户 BI)四平台重跑,4×ra3.4xlarge(1 年 RI)$6,687/月 vs Snowflake $34,360/月、BigQuery flat-rate $12,828/月——"For pure continuous ETL on AWS, Redshift is genuinely the right answer."(https://leanopstech.com/blog/snowflake-vs-bigquery-vs-databricks-vs-redshift-cost-2026/);社区 FinOps 知识库(costly-oss):ra3.4xlarge $3.26/hr 按需 / $1.424/hr 3 年 RI,"Use RA3 for steady workloads with reserved pricing"(https://github.com/njain006/costly-oss/blob/HEAD/backend/app/knowledge/aws.md)。诚实备注:AWS 官网营销 RA3 的措辞是"性能 2 倍、存储 8 倍"(2019 年发布会口径);用户社区真正留住 Redshift 的理由只有一个字:便宜(在那个窄场景里)。性能叙事是厂商的,成本叙事是用户的。LeanOps independent third-party modeling (2026, May pricing): Workload B (200TB stored, 100TB scanned/month, ETL every 15 min, 100 BI users) replayed on four platforms — 4x ra3.4xlarge (1-year RI) at $6,687/month vs Snowflake $34,360/month and BigQuery flat-rate $12,828/month: "For pure continuous ETL on AWS, Redshift is genuinely the right answer." (https://leanopstech.com/blog/snowflake-vs-bigquery-vs-databricks-vs-redshift-cost-2026/); the community FinOps knowledge base (costly-oss): ra3.4xlarge $3.26/hr on-demand / $1.424/hr 3-year RI, "Use RA3 for steady workloads with reserved pricing" (https://github.com/njain006/costly-oss/blob/HEAD/backend/app/knowledge/aws.md). Honest note: AWS markets RA3 as "2x performance, 8x storage" (2019 launch messaging); the reason the community keeps Redshift is one word: cheap (in that narrow scenario). Performance is the vendor's narrative; cost is the users'.
- 竞品差距
- Snowflake 按秒计费的 warehouse 对 24/7 连续负载是结构性劣势(同 workload 贵 5 倍);BigQuery flat-rate 同场景 $12,828/月,接近但仍贵一倍(且 BQ 要求 GCP 在场);Databricks DBU+EC2 双重计费 $20,900/月。结论:"AWS 锁定 + 连续可预测 ETL + 愿意买 RI"是 Redshift 唯一还能赢的窄场景,离开这个场景即输。Snowflake's per-second warehouse billing is a structural disadvantage for 24/7 continuous loads (5x on the same workload); BigQuery flat-rate is close at $12,828/month but still 2x (and requires GCP presence); Databricks' DBU+EC2 double billing lands at $20,900/month. Conclusion: "AWS-locked + continuous predictable ETL + willing to buy RIs" is the only narrow scenario where Redshift still wins — outside it, it loses.
- 证据等级
- 独立实测(LeanOps 建模)+ 社区共识Independent modeling (LeanOps) + community consensus
- 最后核验
- 2026-10-012026-10-01