Store per-second CPU and memory time series for millions of machines with real-time ad-hoc analytics, low per-query cost, and scalable schema. Which database and data model do you choose?

The correct choice is a narrow Bigtable table with row keys combining VM identifier and per-second timestamp. Bigtable is ideal for high-throughput time-series data due to its excellent performance for reads and writes at scale, low latency, and cost-effectiveness for large datasets. A narrow table design, where each row represents a single time-series sample (VM + timestamp), is efficient for Bigtable, allowing for fast lookups and range scans for ad-hoc analytics. Appending to a BigQuery table would work for storage but BigQuery is optimized for analytical queries over large datasets, not high-volume, low-latency point writes and real-time ad-hoc queries on individual data points. Creating a wide BigQuery table and updating rows per interval is inefficient and costly due to BigQuery's update mechanisms. A wide Bigtable table keyed by VM+per-minute timestamp, storing per-second samples as columns, would lead to very wide rows, which are less efficient for Bigtable and make ad-hoc queries across specific seconds more complex.