Why it matters

For time-series workloads with millions of writes per second and single-digit-ms read latency requirements, Bigtable is often the only viable managed choice on GCP.

Advertisement

The architecture

Bigtable data model is similar to HBase: rows keyed by lexicographic row keys, column families containing many columns, cells with timestamps and versions.

An instance has nodes (compute+storage units). Nodes scale up for more throughput; storage scales independently. Zonal or replicated clusters for HA and geo-distribution.

Bigtable data + physical modelRow keysorted, hot designColumn familiesphysical groupingNodesthroughput unitRow key design determines throughput and hot spots — the single most important schema choice
Bigtable structure.
Advertisement

How it works end to end

Row key design is critical. Monotonic keys hit one node; well-distributed keys spread across nodes. Salting or hashing helps for hot writes.

Column families group columns by access pattern and TTL. Tables have limited families (default 100).

Bigtable integrates with HBase API for compatibility, and with BigQuery via federated queries.