Architecture / Ideal Workloads
Distributed SQL, shared-nothing, peer-to-peer architecture. All nodes symmetrical; any node can handle reads/writes. Cluster uses distributed consensus: No matter where data lives, every node can access data anywhere in cluster
SYSTEM OF RECORD. Optimized for transactional workloads that require strong consistency and global distribution, such as AI innovators, cybersecurity, eCommerce & retail, financial services, fintech/payments, gaming, quant/trading & research, and online travel
Globally distributed relational system; absolute consistency via TrueTime (atomic clocks + GPS); proprietary GCP infrastructure dependency
GCP USERS RUNNING CLOUD-NATIVE APPS. GCP-commited enterprises
Auto-Sharding (Dynamic Re-Sharding Online)
NATIVE & AUTOMATIC. Automatically shards data into ranges and dynamically splits, merges, and rebalances online across nodes based on load and size. Zero downtime, fully transparent
Splits rebalanced automatically; less operator-visible; online but less configurable
Automatic Geo-Partitioning (Multi-Region Data Affinity)
NATIVE AND AUTOMATIC. Declarative SQL schema adjustments automatically route, partition, and anchor data close to the user's location, automatically moving data to the region where it is most frequently accessed. Supports geo-partitioning with zone configurations for data locality, compliance, and low latency
Geo-partitioning available; requires careful interleaved schema design; less intuitive DDL
Availability including Multi-Cloud and Hybrid
Available on all public clouds, e.g., AWS-Google Cloud-Azure; can run a single logical cluster spanning multiple clouds. Can run on prem/local, and cloud + prem hybrid deployments
Historically locked to Google Cloud. Now available (Preview version, April 2026) as Spanner Omni, a managed GCP service, or self-managed for on-prem or other clouds; less mature and operationally flexible
Change Data Capture (CDC)
NATIVE. CHANGEFEED command enables scalable, resilient streaming of data changes to Kafka, cloud storage, and webhooks; no third-party CDC tool needed. CDC Queries enable SQL-based filtering and transformation of streams
Spanner Change Streams available; no native SQL-based stream transformation; typically requires Dataflow or Pub/Sub for downstream delivery
Data Anomalies
ZERO under Serializable isolation: all standard SQL anomalies (dirty reads, non-repeatable reads, phantom reads, lost updates, and write skew) are prevented by default with no additional developer configuration. Read Committed is also available for workloads where some consistency relaxation is an acceptable tradeoff for reduced latency
ZERO under Serializable isolation. all standard SQL anomalies prevented by default. Repeatable Read (snapshot) also available
Data Integrity & Foreign Keys Support
Provides strict ACID enforcement at the storage layer and full referential integrity. NATIVELY VALIDATES FOREIGN KEYS, explicit CHECKs, and transactional constraints to ensure absolute global correctness across global nodes
Hardened data integrity via Paxos replication and two-phase commits. Guarantees absolute correctness across global nodes. Enforced and informational foreign keys
Data Model Complexity
LOW. Relational model with strict schemas, normalized tables, joins, and referential integrity. Ideal for managing complex relationships and transactional systems of record; adapts easily to microservices and enterprise legacy systems
MODERATE TO HIGH. Requires intentional schema hierarchical design (interleaving tables) to ensure localized performance and avoid hotspotting
Data Residency
STRONG, INTUITIVE, AND LOCALITY-AWARE. Helps fulfill compliance (e.g., GDPR, CCPA) with Row-Level Control: can pin specific rows to specific geographic regions using REGIONAL BY ROW command, while preserving single logical data platform. Business and compliance teams can use simple SQL commands to ensure customer data never leaves specific geographic borders
Complex & Rigid. Lacks native, row-level data pinning out-of-the-box. To satisfy strict regional data-sovereignty laws, organizations must split data into separate, complex database instances or manually manage it. Geo-partitioning for data residency; supported but at less granular levels
Developer Tools / Experience / Ease of Use
Rich ecosystem: Local CLI, web console UI, ORMS, BI tools, SQL clients, native DB migration toolkits, language‑specific drivers, and compatibility with standard PostgreSQL developer tools like psql
PostgreSQL wire protocol-compatible; feels exactly like developing on standard PostgreSQL. Fits effortlessly into existing ORMs, drivers, and frameworks
Can be spun up instantly in any environment (AWS, GCP, on-prem) with the exact same management interface. The cluster manages its own data balancing, scaling, and hardware survival automatically; DBAs do not need to be distributed systems experts to keep it running smoothly.
GCP Console UI, gcloud CLI tools, emulator for local testing, and tight integration with Google Cloud Code and Vertex
Requires learning custom schema concepts (e.g., table interleaving) and specific Google SQL dialects or custom APIs
FAST BUT RESTRICTIVE. If already fully committed to GCP, it can be a single-click deployment; if operating across multiple clouds or on-premise, introducing Spanner adds heavy infrastructure, architecture, and procurement friction. Once running in the cloud; administrative maintenance is low with Google handling the underlying hardware patching, backups, and sharding; uncertain with Spanner Omni
Distributed ACID Transactions
Fully distributed, multi-row, multi-table ACID transactions out-of-the-box. Fully supported with serializable isolation using distributed consensus (Raft Protocol) across tables, ranges, and regions; strong ACID guarantees
Fully supported; requires two-phase commit with TrueTime timestamps
Enterprise Support
Dedicated 24/7/365 enterprise support directly from Cockroach Labs with strict SLAs and custom engineering channels. Offers global follow-the-sun support (TSE+SRE) with proven reliability and global partnerships with industry leaders. Single Global Incident Management integrates Engineering + Support + Customer Success in one channel for consistency/immediacy
Tiered Google Cloud premium support; specialized Spanner support requires a GCP enterprise-level organization agreement
FinOps Support
HIGH. Straightforward pricing based on predictable node usage or consumption metrics. Avoids hidden, fluctuating network traps when moving data across different infrastructure regions. Supports financial governance/FinOps
LOW. Highly-complext, multi-layered a la carte billing combines processing nodes, storage tiers (SSD vs HDD), and highly variable cross-region network egress fees. Harder for finance teams to forecast
Follower Reads
SUPPORTED. Supports follower/replica reads with Bounded (controlled) Staleness, allowing low‑latency local reads from nearby replicas while keeping strong global ordering
Stale/snapshot reads available; less granular control
FREEDOM
ZERO VENDOR LOCK-IN. Runs on any public or private cloud, across multiple clouds, via CockroachDB's Bring Your Own Cloud (BYOC) offering, on-premises, bare metal, Kubernetes, self-hosted, or in a hybrid deployment encompassing some or all of these. Business Source License (BSL) but Source Available. Full commercial-grade support directly from CockroachDB
HIGH VENDOR LOCK-IN. Architecture tied to GCP's TrueTime hardware by design; Spanner Omni offers limited off-GCP deployment but is a newer, more constrained offering
Joins
Executes fully distributed hash, merge, and lookup joins across arbitrary nodes with CockroachDB's advanced Cost-based Optimizer. Full standard SQL support for complex INNER, OUTER, LEFT, RIGHT joins across distributed tables
Efficiently executes distributed joins, which are optimized when using parental table interleaving
LDAP Support
NATIVE. Direct native support for external authentication systems like LDAP, Active Directory, GSSAPI, and OIDC
Indirect. Does not connect to LDAP natively; authentication must be mapped through Google Cloud IAM, single sign-on (SSO), or Workforce Identity
Migrations
Uses MOLT (Migration Off Legacy Technology) Toolkit & change data capture (CDC): MOLT handles schema conversion/verification and CDC moves data out. PostgreSQL wire protocol compatibility enables lift-and-shift; shadow mode testing
Spanner Migration Tool; Cassandra interface; library of MySQL functions for compatibility; more re-engineering typically required
Multi-Active
YES: FULLY MULTI-ACTIVE/MULTI-REGION; read/write and handle connection requests from any node in the cluster. All nodes are equal and active; any node can accept read and write traffic simultaneously.
YES-PARTIALLY. All nodes can serve traffic; Paxos leader still coordinates writes
Multi-Data-Center Support
FULL. Connects geographically isolated, heterogeneous data centers (AWS, GCP, Azure, on-prem) into a single logical cluster, supported by features such as Physical Cluster Replication (PCR) and Logical Data Replication (LDR)
Historically GCP regions only. Now supports multi-environment topologies via Spanner Omni, which allows self-managed local instances
Multi-region Functionality / Multi-region Writes
ACTIVE-ACTIVE: Read/Write from any node in any region; built-in low-latency local access patterns and Survival Goals (e.g., ALTER DATABASE...SURVIVE REGION FAILURE) commands configure fault tolerance intent
True multi‑region, multi‑active writes: any node in any region can serve reads and writes while preserving serializable consistency guarantees
Single-region, dual-region, or multi-region configurations available; less granular, more complex to configure
Leader-dependent; every write must pass through the Paxos group's designated primary read-write leader zone, introducing network latency
Replication
Built-in, automatic consensus replication using the Raft protocol; data is divided into ranges and replicated across nodes
Built-in, managed consensus replication driven by the Paxos protocol across zone/regional boundaries
Required Downtime
ZERO. Online schema changes, rolling upgrades, and cluster expansion occur without taking the data platform offline
Near zero: Managed by Google with automated maintenance
Resilience
Five 9s availability: Survives node/disk/rack/region failures automatically via Raft consensus, with zero data loss (RPO=0). Naturally resilient to outages with granular row-level control
Five 9s availability; massive scale (17+ exabytes, 6B+ queries/sec at peak); Google-managed resilience
Scale
Virtually unlimited horizontal scale-out. Automatic, seamless handling of growing datasets; increase storage and throughput capacity linearly simply by adding more nodes
Virtually unlimited horizontal scale; seamless handling of growing datasets; tied to GCP compute capacity
Schema Changes
FULLY ONLINE & NON-BLOCKING. Online transactional schema changes (add/alter columns, indexes, constraints) run in the background without locking tables with zero downtime. Designed for always‑on services
Online & Managed. Changes run asynchronously in the background via schema update DDL operations without application downtime
Security-Privacy-Compliance
RBAC, Encryption at Rest with Customer Managed Encryption Keys (CMEK), TLS encryption in transit, IAM integrations, column-level encryption, and robust data-masking natively. Fine-grained encryption at cluster, database, table, or partition levels. Certified SOC 2 Type II and SOC 3, PCI-DSS, HIPAA, and ISO 27001-27017-27018 compliant, with ISO 42001 (Responsible, Ethical, and Safe AI Governance) pending. CockroachDB CIS Benchmarks to deploy hardened CockroachDB configurations. Comprehensive support for GDPR and DORA compliance
GCP-native security: IAM integration, CMEK (Customer Managed Encryption Keys), VPC Service Controls, and Access Approval.
SQL Compatibility
HIGH. PostgreSQL Wire Compatible: Uses PG wire protocol; strong ANSI SQL with complex queries, joins, window functions, triggers, stored procedures, and UDFs. Supports spatial data, extensions, syntax; most apps connect with minimal or no changes
GoogleSQL and PostgreSQL dialects; not full PostgreSQL compatibility; gaps in spatial data, extensions, triggers, and stored procedures
Storage Engine
Built on Pebble, a highly tuned, GO-based, LSM-tree key-value storage engine developed by Cockroach Labs and inspired by RocksDB specifically for distributed SQL
Built on an internal, proprietary storage subsystem optimized by Google to interact directly with Colossus (Google's global file system)
Stored Procedures
SUPPORTED AND MATURE. PL/pgSQL and other languages such as Python and Perl support deep procedural logic, autonomous transactions, and complex business rule enforcement. Supports user-defined stored procedures
Only built-in system stored procedures; users cannot create custom stored procedures
Transaction Performance / Isolation Levels
Optimized for OLTP with strong consistency; cross‑region transactions maintain data correctness. Optimizations like Parallel Commits drop distributed execution overhead to a single network round-trip for most transactions
Enforces strict Serializable isolation exclusively, the strongest isolation level, to ensure zero data anomalies under heavy parallel traffic, and Read Committed
Enforces Serializable isolation by default, offering linearizable consistency guarantees
Distributed transactions globally optimized by Google's dedicated internal networking and TrueTime clock synchronization
Triggers & Deferrable Constraints
FULLY SUPPORTED. Supports triggers and deferrable constraints across all deployment models
No trigger support; workloads that rely on triggers require rework to run on Spanner
Vector Search
BUILT-IN NATIVE SUPPORT via VECTOR data type, scalable distributed HNSW/IVF indexing, and pgvector (the industry standard for vector similarity search). CockroachDB's C-SPANN provides distributed vector indexing (ANN) at scale; available across all tiers
KNN and ANN (ScaNN-based); Vertex AI integration; restricted to Enterprise and Enterprise Plus editions
Writes and Query Routing
Every node is a gateway to the entirety of the database for unlimited reads and writes in any region. Any node can accept SQL queries; a Distributed Optimizer routes work to the right ranges/replicas based on locality and cost
Handled at the network layer by Google's internal Cloud Load Balancing, routing requests to the nearest healthy replica zone
PRICING
Commercial Enterprise: Simple, straightforward pricing, plus the ability to tie data to a location to avoid egress costs. Free for single-node/dev. Free Community Tier
Node/processing-unit pricing from; premium for multi-region; standardized GCP pricing model with less negotiation flexibility
