Datastores

In order to reduce operational complexity, SpiceDB leverages existing, popular systems for persisting data.

AuthZed has standardized our managed services on CockroachDB, but we give self-hosted customers the option to pick the datastore that best suits their operational requirements.

• CockroachDB - Recomended for self hosted deployments with high throughput and/or multi-region requirements
• Cloud Spanner - Recommended for self-hosted Google Cloud deployments
• PostgreSQL - Recommended for self-hosted single-region deployments
• MySQL - Not recommended; only use if you cannot use PostgreSQL
• memdb - Recommended for local development and integration testing against applications

Migrations

Before a datastore can be used by SpiceDB or before running a new version of SpiceDB, you must execute all available migrations.

In order to migrate a datastore, run the following command with your desired values:

spicedb migrate head \
    --datastore-engine $DESIRED_ENGINE \
    --datastore-conn-uri $CONNECTION_STRING

For more information, refer to the Datastore Migration documentation.

CockroachDB

Usage Notes

• Recommended for multi-region deployments, with configurable region awareness
• Enables horizontal scalability by adding more SpiceDB and CockroachDB instances
• Resiliency to individual CockroachDB instance failures
• Query and data balanced across the CockroachDB
• Setup and operational complexity of running CockroachDB

Developer Notes

• Code can be found here (opens in a new tab)
• Documentation can be found here (opens in a new tab)
• Implemented using pgx (opens in a new tab) for a SQL driver and connection pooling
• Has a native changefeed
• Stores migration revisions using the same strategy as Alembic (opens in a new tab)

Configuration

Required Parameters

Optional Parameters

Overlap Strategy

In distributed systems, you can trade-off consistency for performance.

CockroachDB datastore users that are willing to rely on more subtle guarantees to mitigate the New Enemy Problem can configure --datastore-tx-overlap-strategy.

The available strategies are:

For more information, refer to the CockroachDB datastore README (opens in a new tab) or our blog post "The One Crucial Difference Between Spanner and CockroachDB (opens in a new tab)".

Garbage Collection Window

SpiceDB warns if the garbage collection window as configured in CockroachDB is smaller than the SpiceDB configuration.

If you need a longer time window for the Watch API or querying at exact snapshots, you can adjust the value in CockroachDB (opens in a new tab):

ALTER ZONE default CONFIGURE ZONE USING gc.ttlseconds = 90000;

Relationship Integrity

Relationship Integrity is a new experimental feature in SpiceDB that ensures that data written into the supported backing datastores (currently: only CockroachDB) is validated as having been written by SpiceDB itself.

• What does relationship integrity ensure? Relationship integrity primarily ensures that all relationships written into the backing datastore were written via a trusted instance of SpiceDB or that the caller has access to the key(s) necessary to write those relationships. It ensures that if someone gains access to the underlying datastore, they cannot simply write new relationships of their own invention.

• What does relationship integrity not ensure? Since the relationship integrity feature signs each individual relationship, it does not ensure that removal of relationships is by a trusted party. Schema is also currently unverified, so an untrusted party could change it as well. Support for schema changes will likely come in a future version.

Setting up relationship integrity

To run with relationship integrity, new flags must be given to SpiceDB:

spicedb serve ...existing flags...
--datastore-relationship-integrity-enabled
--datastore-relationship-integrity-current-key-id="somekeyid"
--datastore-relationship-integrity-current-key-filename="some.key"

Place the generated key contents (which must support an HMAC key) in some.key

Deployment Process

1. Start with a clean datastore for SpiceDB. At this time, migrating an existing SpiceDB installation is not supported.
2. Run the standard migrate command but with relationship integrity flags included.
3. Run SpiceDB with the relationship integrity flags included.

Cloud Spanner

Usage Notes

• Requires a Google Cloud Account with an active Cloud Spanner instance
• Take advantage of Google's TrueTime. The Spanner driver assumes the database is linearizable and skips the transaction overlap strategy required by CockroachDB.

Developer Notes

• Code can be found here (opens in a new tab)
• Documentation can be found here (opens in a new tab)
• Starts a background GC worker (opens in a new tab) to clean up old entries from the manually-generated changelog table

Configuration

• The Cloud Spanner docs (opens in a new tab) outline how to set up an instance
• Authentication via service accounts: The service account that runs migrations must have Cloud Spanner Database Admin; SpiceDB (non-migrations) must have Cloud Spanner Database User.

Required Parameters

Optional Parameters

PostgreSQL

Usage Notes

• Recommended for single-region deployments
• Postgres 15 or newer is required for optimal performance
• Resiliency to failures only when PostgreSQL is operating with a follower and proper failover
• Carries setup and operational complexity of running PostgreSQL
• Does not rely on any non-standard PostgreSQL extensions
• Compatible with managed PostgreSQL services (e.g. AWS RDS)
• Can be scaled out on read workloads using read replicas

Developer Notes

• Code can be found here (opens in a new tab)
• Documentation can be found here (opens in a new tab)
• Implemented using pgx (opens in a new tab) for a SQL driver and connection pooling
• Stores migration revisions using the same strategy as Alembic (opens in a new tab)
• Implements its own MVCC (opens in a new tab) model by storing its data with transaction IDs

Read Replicas

SpiceDB supports Postgres read replicas and does it while retaining consistency guarantees. Typical use cases are:

• scale read workloads/offload reads from the primary
• deploy SpiceDB in other regions with primarily read workloads

Read replicas are typically configured with asynchronous replication, which involves replication lag. That would be problematic to SpiceDB's ability to solve the new enemy problem but it addresses the challenge by checking if a revision has been replicated into the target replica. If missing, it will fall back to the primary.

All API consistency options will leverage replicas, but the ones that benefit the most are those that involve some level of staleness as it increases the odds a revision has replicated. minimize_latency, at_least_as_fresh, and at_exact_snapshot consistency modes have the highest chance of being redirected to a replica.

SpiceDB supports Postgres replicas behind a load-balancer, and/or individually listing replica hosts. When multiple URIs are provided, they will be queried using a round-robin strategy.
Please note that the maximum number of replica URIs to list is 16.

Read replicas are configured with the --datastore-read-replica-* family of flags.

SpiceDB supports PgBouncer (opens in a new tab) connection pooler and is part of the test suite.

Transaction IDs and MVCC

The Postgres implementation of SpiceDB's internal MVCC mechanism involves storing the internal transaction ID count associated with a given transaction in the rows written in that transaction. Because this counter is instance-specific, there are ways in which the data in the datastore can become desynced with that internal counter. Two concrete examples are the use of pg_dump and pg_restore to transfer data between an old instance and a new instance and setting up logical replication between a previously-existing instance and a newly-created instance.

If you encounter this, SpiceDB can behave as though there is no schema written, because the data (including the schema) is associated with a future transaction ID and therefore isn't "visible" to Spicedb. If you run into this issue, the fix is documented here (opens in a new tab)

Configuration

Required Parameters

Optional Parameters

MySQL

Usage Notes

• Recommended for single-region deployments
• Setup and operational complexity of running MySQL
• Does not rely on any non-standard MySQL extensions
• Compatible with managed MySQL services
• Can be scaled out on read workloads using read replicas

Developer Notes

• Code can be found here (opens in a new tab)
• Documentation can be found here (opens in a new tab)
• Implemented using Go-MySQL-Driver (opens in a new tab) for a SQL driver
• Query optimizations are documented here (opens in a new tab)
• Implements its own MVCC (opens in a new tab) model by storing its data with transaction IDs

Read Replicas

SpiceDB supports MySQL read replicas and does it while retaining consistency guarantees. Typical use cases are:

• scale read workloads/offload reads from the primary
• deploy SpiceDB in other regions with primarily read workloads

Read replicas are typically configured with asynchronous replication, which involves replication lag. That would be problematic to SpiceDB's ability to solve the new enemy problem but it addresses the challenge by checking if the revision has been replicated into the target replica. If missing, it will fall back to the primary. Compared to the Postgres implementation, MySQL support requires two roundtrips instead of one, which means it adds some extra latency overhead.

All API consistency options will leverage replicas, but the ones that benefit the most are those that involve some level of staleness as it increases the odds a revision has replicated. minimize_latency, at_least_as_fresh, and at_exact_snapshot consistency modes have the highest chance of being redirected to a replica.

SpiceDB does not support MySQL replicas behind a load-balancer: you may only list replica hosts individually. Failing to adhere to this would compromise consistency guarantees. When multiple host URIs are provided, they will be queried using round-robin. Please note that the maximum number of MySQL replica host URIs to list is 16.

Read replicas are configured with the --datastore-read-replica-* family of flags.

Configuration

Required Parameters

Optional Parameters

memdb

Usage Notes

• Fully ephemeral; all data is lost when the process is terminated
• Intended for usage with SpiceDB itself and testing application integrations
• Cannot be ran highly-available as multiple instances will not share the same in-memory data

Developer Notes

• Code can be found here (opens in a new tab)
• Documentation can be found here (opens in a new tab)
• Implements its own MVCC (opens in a new tab) model by storing its data with transaction IDs

Configuration

Required Parameters

Optional Parameters