Understanding serverless relational data at scale

The presenters consider Cloud Spanner for a hypothetical NoSQL-backed application that now needs ACID transactions, cross-record aggregation, stronger consistency, and managed scaling. Product capability, scale, availability, pricing, and tuning statements are April 2022 claims rather than demonstrated guarantees.

Key points

• Managed does not mean responsibility-free: The presenter says Spanner handles patching, server maintenance, load-balancing and failover architecture, and default encryption, while customers retain access control, backup scheduling, schema migrations, and optional key management [00:01:03]-[00:03:07].
• The scale figure is product-wide and unqualified: The presenter says Spanner processes more than one billion requests per second at peak [00:02:05]-[00:03:07]. The episode does not frame that as Google's aggregate traffic, a per-instance capacity, or a result available to the sample application.
• Consistency and availability need their original qualifications: The presenter attributes global consistency to TrueTime and states an SLA of up to 99.999% [00:03:07]-[00:04:12]. The console walkthrough separately says a multi-region configuration is appropriate when five-nines availability is required [00:04:12]-[00:05:16].
• The price and scaling demo are point-in-time: A regional development instance with 100 processing units is quoted at $0.09 per hour, about $65 per month. The load test adds processing units after CPU crosses an unnamed recommended threshold, then observes more requests without reporting reproducible throughput or latency [00:04:12]-[00:07:21].

Validate current Spanner editions, pricing, processing-unit guidance, CPU thresholds, scaling behavior, SLA terms, backup policy, IAM, and migration requirements. The demo does not establish instance capacity or an application's achievable performance.

Full video: https://www.youtube.com/watch?v=IqAe1WRPjfU(opens in a new tab)