7.3 Bigtable Row Keys and Replication

Cloud Bigtable: Petabyte-Scale Wide-Column Storage

Cloud Bigtable is Google's massively scalable, low-latency NoSQL wide-column database. It powers Google Search, Google Maps, and YouTube, supporting low-latency reads and writes; latency depends on data model, load, storage type, and capacity for petabyte-scale workloads.

Architectural Fundamentals

• Sparse, Distributed Multi-Dimensional Map: Data is indexed by Row Key, Column Family, Column Qualifier, and Timestamp.
• Decoupled Compute and Storage: Bigtable nodes handle incoming client requests and indexing. Actual data files (SSTables) are persisted on Google's Colossus distributed filesystem. Because compute is decoupled from storage, adding or removing cluster nodes rebalances pointers without copying data, with rebalancing and performance stabilization requiring time.
• Ideal Workloads: Time-series telemetry, IoT sensor streams, financial order books, ad-tech clickstream analytics, and marketing event logs exceeding 1 TB in size and requiring thousands to millions of operations per second.

---

Bigtable Row Key Engineering

Bigtable organizes rows in lexicographical (alphabetical) order by row key. Because queries scan contiguous row key ranges, row key design is the single most critical architectural decision in Bigtable.

+--------------------------------------------------------------------------+
| Bigtable Row Key Design Anti-Patterns vs. Best Practices                 |
+--------------------------------------------------------------------------+
|                                                                          |
| ANTI-PATTERN: Timestamp Prefix                                           |
| Row Key: [2026-10-18T10:00:00]#[sensor_001]                              |
|   - All incoming writes share the current timestamp prefix!              |
|   - 100% of writes hit Tablet 1 -> Single Node Overload (Hotspotting)    |
|                                                                          |
| BEST PRACTICE: Field Promotion                                           |
| Row Key: [sensor_001]#[temperature]#[2026-10-18T10:00:00]                |
|   - High-cardinality sensor_id evenly distributes writes across tablets  |
|   - Scans can target a specific sensor's contiguous time range           |
|                                                                          |
| BEST PRACTICE: Reverse Timestamps for Latest-First Retrieval             |
| Row Key: [sensor_001]#[Long.MAX_VALUE - timestamp]                       |
|   - Most recent readings appear at the top of the sensor's key range     |
|   - Eliminates need to scan millions of historical rows to find latest   |
+--------------------------------------------------------------------------+

Core Row Key Design Principles

1. Field Promotion (High-Cardinality First): Place high-cardinality attributes (e.g., device_id, customer_id, or another sufficiently distributed entity identifier) at the beginning of the row key to disperse writes evenly across multiple tablet servers.
2. Avoid Sequential Timestamps as Prefixes: Prefixing row keys with monotonically increasing timestamps causes every new write to land on the same tablet server at the end of the keyspace, creating severe write hotspots.
3. Reverse Timestamps for Most-Recent Scans: When an application needs to read the most recent sensor reading or stock price, subtract the event timestamp from a maximum integer: Long.MAX_VALUE - timestamp. This causes new rows to sort lexicographically before older rows, allowing queries to scan the newest records instantly with LIMIT 1.
4. Salting / Hash Prefixing: When an entity has too much traffic for one key range, deliberately shard its events across several keys. Hashing one constant ID always produces the same bucket and does not fix a single hot row. Query all needed shards and merge results.

---

Bigtable Tablet Splitting, Garbage Collection & Replication

Tablet Splitting

Bigtable partitions tables into contiguous chunks called tablets (whose size and split timing are managed by the service). As tablets grow in size or experience sustained CPU pressure, Bigtable automatically splits the tablet into two smaller tablets and reallocates one to another worker node, balancing cluster load.

Garbage Collection Policies

Bigtable stores multiple timestamped versions of data within each cell. Garbage collection policies run during background compactions to purge unwanted versions, freeing storage:

• Age-Based Policy: Retains data written within a specific duration (e.g., expire data older than 30 days).
• Version-Based Policy: Retains a fixed number of recent versions per cell (e.g., keep the last 3 versions).
• Compound Policy: Combines rules using Boolean logic (AND or OR), such as keeping up to 5 versions, but only if they are younger than 14 days.

Single-Cluster vs. Multi-Cluster Replication

• Multi-Cluster Topology: Bigtable supports replicating data across multiple supported clusters in near-real-time (eventual consistency). Availability and SLA eligibility depend on the cluster/routing configuration; replication is eventually consistent.
• App Profiles & Routing:
  • Single-Cluster Routing: Directs client application traffic to a specific cluster. Provides Read-Your-Writes consistency for applications that require immediate read visibility after writing.
  • Multi-Cluster Routing: Automatically routes traffic to the nearest healthy cluster. If one cluster experiences an outage, client requests failover automatically to a surviving cluster without application intervention.

Row Atomicity and a Worked Access Pattern

Bigtable mutations are atomic within one row; the core row-oriented model does not provide a relational transaction across arbitrary rows. Put values that must change together in the same logical row when that matches the access pattern, and use conditional row mutations for an atomic check/update where appropriate. A broad relational join or a multi-entity balance transfer usually needs a different design or transactional datastore.

Consider a hypothetical fleet with many devices, each producing modest telemetry. A key beginning with the device ID followed by metric and a fixed-width sortable timestamp supports a contiguous time-range scan for one device and spreads different devices over the key space. If one device becomes exceptionally hot, repeating its hash prefix will not help; deliberate event sharding requires the reader to fan out and merge shards. Test both the normal range query and the hottest entity, because average fleet throughput can hide a local hotspot.

Garbage collection also has a visibility boundary. Age/version rules are applied during compaction, not as an immediate deletion deadline for every read. If a query must never return older versions, set an appropriate read filter as well as the storage cleanup policy. Keep this retention behavior separate from IAM authorization and backup/recovery requirements.