How DynaSharp compares

DynaSharp is one of three ways to reach DynamoDB from .NET with mapped types. The other two are the object persistence model of the AWS SDK, DynamoDBContext, and EfficientDynamoDb. This page compares what each has, so you can choose. It is about features: Performance has the cost of a call, and Migrating the way from either of them.

The versions compared are AWSSDK.DynamoDBv2 4.0.109 with AWSSDK.Core 4.0.102.12, and EfficientDynamoDb 0.9.19. The claims come from their packages: the assemblies, the XML documentation of the AWS SDK and the package metadata. Where a claim comes from the documentation of EfficientDynamoDb alone, it says "from its docs".

Feature matrix

	DynaSharp	AWS SDK DynamoDBContext	EfficientDynamoDb
Target frameworks	.NET 11	.NET Framework 4.7.2, .NET Standard 2.0, .NET Core 3.1, .NET 8	.NET 8
Runtime package dependencies	None	AWSSDK.Core	Microsoft.IO.RecyclableMemoryStream
Mapping	Source generator, at build time	Reflection at run time; asks DescribeTable for the key schema unless that is turned off	Reflection, with accessors emitted as IL at run time; maps only properties marked [DynamoDbProperty]
Native AOT	Yes, published and run in CI	Partly: marked trimmable and annotated for the trimmer, not marked AOT compatible	No: it emits IL at run time
Keys built from several values	Key templates, such as CUSTOMER#{Id} with ORDER#{Number:d6}	No; a global index key can span several attributes	No, one property per key
Several types in one table	A discriminator per entity, and one query that returns several types	A class hierarchy, with [DynamoDBPolymorphicType]	No; read the items as documents and convert them yourself
Typed conditions and filters	Typed handles, Attrs, or text with handles in its holes	Lambda filters with ContextExpression, a condition on save; older APIs take names as strings	Lambda based, Condition<T>.On(x => ...)
Updates of named attributes	Update(key).Set(...), Add, Remove, also into nested documents	No: SaveAsync writes the whole object; a transaction takes an update expression as text	Yes, UpdateItem<T>().On(x => ...), also into nested paths
Optimistic locking	[Version] on put, update, delete and transaction steps	[DynamoDBVersion] on save and delete	[DynamoDbVersion], on its SaveAsync and DeleteAsync only
Transactions	Typed steps; a marker says which step failed and why	Yes; the reasons come as a list in request order	Yes; the reasons come as a list in request order
Batches	Any size, split and retried for you; what is left ends in an exception	Any size, split and retried for you	Unprocessed items retried; more than 25 writes or 100 reads throw
One query over several partitions	Fan-out, merged in the order of the sort key	No	No
Paging token for a later request	ContinuationToken	PaginationToken	Pagination token
Parallel scan	Scan().Parallel(segments)	Segments, run and merged by you	Yes
Projection	Project(...), Select(attribute)	Attribute names in the operation config of the document model	Typed, also into a type of its own
Values back from a write	ReturnNew(), ReturnOld(), ReturnOldOnFailure()	Not in the context; in the document model	Yes
Creating a table from the model	CreateTable(), with the indexes the entities declare	No, the low-level client only	No
DynamoDB Streams	DynamoDbStreamsClient, in the package, with the changes of a table as entities	A client in another package	No
Locks and leader election	DynaSharp.Coordination, a package of its own, with fencing tokens	No	No
Time to live in seconds	EpochSecondsConverter, and [TimeToLive], which CreateTable() turns on	StoreAsEpochLong	No built-in converter
Credentials	Environment, web identity (IRSA), profiles with keys, assume role, SSO and credential_process, ECS and EKS Pod Identity, EC2	The same; SSO and assume role need further AWS packages at run time	Keys you pass, or a provider of your own; the chain of the AWS SDK through the package EfficientDynamoDb.Credentials.AWSSDK, from its docs
Retries	Attempts and a timeout per attempt for the client, attempts per command	Standard and adaptive retry modes, number of retries	A strategy per kind of error
Telemetry	ActivitySource and Meter named DynaSharp, logs through ILoggerFactory	An SDK telemetry provider, no-op unless one is set	None
Compile-time checks	Analyzers for entities, keys, commands and expressions	Analyzers for values of the low-level request types	None
Tests without DynamoDB	DynaSharp.Testing fakes the wire; no interfaces	IDynamoDBContext and interfaces for its operations	IDynamoDbContext, but its builders expect the real context
Below the mapped API	Items without a model, a request type per operation, raw JSON for any operation	The full low-level client	A low-level context for item operations and PartiQL
PartiQL	No	The low-level client only	The low-level context only

None of the three has a unit of work: a call is sent when it is made, and a batch or a transaction when it is executed. None has a LINQ provider over IQueryable.

Not in DynaSharp

What the other two have, or what you may look for, and DynaSharp does not do:

• Older .NET. It targets .NET 11 only.
• A class hierarchy in one table. Several types are separate entities with a discriminator each, read together through Of<A, B>(), not derived classes of one base.
• Interfaces to mock. The client and the tables are sealed classes. A test fakes the wire with DynaSharp.Testing instead, see Testing.
• MFA in a profile. A profile with mfa_serial fails, because the code of the device would be needed at every refresh. See Assume a role.
• PartiQL, DAX and vector search. The AWS SDK has PartiQL on its low-level client and SearchVectorsAsync on the context.
• Every control-plane operation as a type. The request API covers 17 operations, the item operations and table administration such as CreateTable, UpdateTable and time to live. Backups, exports, tags and global tables go through SendRawAsync with the JSON of the operation, see API levels.
• A non-throwing mode. A failure is an exception, except a condition that did not hold for TrySendAsync, which is an answer, see Conditions; EfficientDynamoDb can return a result with any error instead.
• Change tracking. There is no SaveChanges: a command writes what it names, see Migrating.