Dapr offers key/value storage APIs for state management. If a microservice uses state management, it can use these APIs to leverage any of the supported state stores, without adding or learning a third party SDK.
When using state management your application can leverage several features that would otherwise be complicated and error-prone to build yourself such as:
- Distributed concurrency and data consistency
- Retry policies
- Bulk CRUD operations
See below for a diagram of state management's high level architecture.
- State Management API
- State Store Behaviors
- Concurrency
- Consistency
- Retry Policies
- Bulk Operations
- Querying State Store Directly
Developers can use the state management API to retrieve, save and delete state values by providing keys.
Dapr data stores are components. Dapr ships with Redis out-of-box for local development in self hosted mode. Dapr allows you to plug in other data stores as components such as Azure CosmosDB, SQL Server, AWS DynamoDB, GCP Cloud Spanner and Cassandra.
Visit State API for more information.
NOTE: Dapr prefixes state keys with the ID of the current Dapr instance. This allows multiple Dapr instances to share the same state store.
Dapr allows developers to attach to a state operation request additional metadata that describes how the request is expected to be handled. For example, you can attach concurrency requirements, consistency requirements, and retry policy to any state operation requests.
By default, your application should assume a data store is eventually consistent and uses a last-write-wins concurrency pattern. On the other hand, if you do attach metadata to your requests, Dapr passes the metadata along with the requests to the state store and expects the data store to fulfill the requests.
Not all stores are created equal. To ensure portability of your application, you can query the capabilities of the store and make your code adaptive to different store capabilities.
The following table summarizes the capabilities of existing data store implementations.
| Store | Strong consistent write | Strong consistent read | ETag |
|---|---|---|---|
| Cosmos DB | Yes | Yes | Yes |
| PostgreSQL | Yes | Yes | Yes |
| Redis | Yes | Yes | Yes |
| Redis (clustered) | Yes | No | Yes |
| SQL Server | Yes | Yes | Yes |
Dapr supports optimistic concurrency control (OCC) using ETags. When a state is requested, Dapr always attaches an ETag property to the returned state. And when the user code tries to update or delete a state, it's expected to attach the ETag through the If-Match header. The write operation can succeed only when the provided ETag matches with the ETag in the state store.
Dapr chooses OCC because in many applications, data update conflicts are rare because clients are naturally partitioned by business contexts to operate on different data. However, if your application chooses to use ETags, a request may get rejected because of mismatched ETags. It's recommended that you use a Retry Policy to compensate for such conflicts when using ETags.
If your application omits ETags in writing requests, Dapr skips ETag checks while handling the requests. This essentially enables the last-write-wins pattern, compared to the first-write-wins pattern with ETags.
NOTE: For stores that don't natively support ETags, it's expected that the corresponding Dapr state store implementation simulates ETags and follows the Dapr state management API specification when handling states. Because Dapr state store implementations are technically clients to the underlying data store, such simulation should be straightforward using the concurrency control mechanisms provided by the store.
Dapr supports both strong consistency and eventual consistency, with eventual consistency as the default behavior.
When strong consistency is used, Dapr waits for all replicas (or designated quorums) to acknowledge before it acknowledges a write request. When eventual consistency is used, Dapr returns as soon as the write request is accepted by the underlying data store, even if this is a single replica.
Dapr allows you to attach a retry policy to any write request. A policy is described by an retryInterval, a retryPattern and a retryThreshold. Dapr keeps retrying the request at the given interval up to the specified threshold. You can choose between a linear retry pattern or an exponential (backoff) pattern. When the exponential pattern is used, the retry interval is doubled after each attempt.
Dapr supports two types of bulk operations - bulk or multi. You can group several requests of the same type into a bulk (or a batch). Dapr submits requests in the bulk as individual requests to the underlying data store. In other words, bulk operations are not transactional. On the other hand, you can group requests of different types into a multi-operation, which is handled as an atomic transaction.
Dapr saves and retrieves state values without any transformation. You can query and aggregate state directly from the underlying state store. For example, to get all state keys associated with an application ID "myApp" in Redis, use:
KEYS "myApp*"NOTE: See How to query Redis store for details on how to query a Redis store.
If the data store supports SQL queries, you can query an actor's state using SQL queries. For example use:
SELECT * FROM StateTable WHERE Id='<app-id>||<actor-type>||<actor-id>||<key>'You can also perform aggregate queries across actor instances, avoiding the common turn-based concurrency limitations of actor frameworks. For example, to calculate the average temperature of all thermometer actors, use:
SELECT AVG(value) FROM StateTable WHERE Id LIKE '<app-id>||<thermometer>||*||temperature'NOTE: Direct queries of the state store are not governed by Dapr concurrency control, since you are not calling through the Dapr runtime. What you see are snapshots of committed data which are acceptable for read-only queries across multiple actors, however writes should be done via the actor instances.
- Follow these guides on
- Read the state management API specification
- Read the actors API specification