Cosmos: flow transactional batch failures through the execution strategy so they can be retried

[Copilot]

Cosmos transactional batch failures returned a result object from ProcessBatchResponse rather than throwing, so the failure surfaced as a DbUpdateException outside any execution-strategy scope and could never be retried by an ExecutionStrategy (ShouldRetryOn was never called).

Changes

• Throw wrapped exception on batch failure — CosmosClientWrapper.ProcessBatchResponse now computes the errored entries (those whose per-item status code matches the batch-level error code), wraps the CosmosException into a DbUpdateConcurrencyException (PreconditionFailed) or DbUpdateException (Conflict / other), and throws it directly. CosmosTransactionalBatchResult and CosmosTransactionalBatchException are removed.
• Execute batches once in the client — ExecuteTransactionalBatchAsync no longer wraps itself in the execution strategy; it runs once and returns Task (interface updated).
• Drive retries from SaveChangesAsync — CosmosDatabaseWrapper takes an injected IExecutionStrategy and wraps the whole set of batches in a single ExecuteAsync. A BatchExecutionState.CommittedOperations counter lets a retry skip already-committed operations while preserving the original ordering of single-entry saves interleaved with transactional batches. The execution strategy unwraps DbUpdateException via ExecutionStrategy.CallOnWrappedException to reach the inner CosmosException before calling ShouldRetryOn, so transient failures are correctly identified and retried.
• Suppressed concurrency consistency — when a DbUpdateConcurrencyException from a batch is suppressed via OptimisticConcurrencyExceptionAsync, RowsAffected is not incremented (matching SaveAsync single-entry behavior), while CommittedOperations still advances so a subsequent retry skips the already-processed operation.

Because a transactional batch is atomic, a failed batch commits nothing and is retried in place; deterministic transaction serialization across attempts makes index-based skipping of committed operations valid.

Result

The repro from the issue now behaves as expected — a transient batch failure reaches the strategy and is retried:

protected override bool ShouldRetryOn(Exception exception)
{
    Console.WriteLine($"Exception: {exception.Message}"); // now written
    return true;
}

Two functional tests verify the behavior:

• SaveChanges_transactional_batch_failure_flows_through_execution_strategy — a conflicting batch reaches ShouldRetryOn and yields RetryLimitExceededException, with nothing committed.
• SaveChanges_suppressed_concurrency_exception_in_transactional_batch_reports_zero_rows_affected — a stale-ETag batch update suppressed by an interceptor returns 0 rows affected and leaves the store unchanged.

[Copilot]

Pull request overview

This PR fixes Cosmos transactional batch failures bypassing the execution strategy retry pipeline by throwing on batch failure (instead of returning a “failed result”) and moving the execution-strategy scope to wrap the full SaveChanges batch sequence, allowing transient failures to reach ShouldRetryOn and be retried.

Changes:

• Replace the transactional batch “result” return pattern with a CosmosTransactionalBatchException thrown from ProcessBatchResponse.
• Change ICosmosClientWrapper.ExecuteTransactionalBatchAsync to return Task and execute once (no per-batch execution strategy wrapping in the client).
• Wrap the full set of transactional batches in a single execution strategy invocation inside CosmosDatabaseWrapper, with retry skipping of already-committed operations, and add a functional test verifying the exception reaches the execution strategy.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
test/EFCore.Cosmos.FunctionalTests/CosmosTransactionalBatchTest.cs	Adds a functional test asserting transactional batch failures flow through the execution strategy and nothing is committed.
src/EFCore.Cosmos/Storage/Internal/ICosmosClientWrapper.cs	Updates the transactional batch API to return Task rather than a result object.
src/EFCore.Cosmos/Storage/Internal/CosmosTransactionalBatchException.cs	Introduces an internal exception type carrying errored entries and the underlying CosmosException.
src/EFCore.Cosmos/Storage/Internal/CosmosDatabaseWrapper.cs	Moves execution strategy retry scope to encompass the entire batch set and adds committed-operation skipping across retries.
src/EFCore.Cosmos/Storage/Internal/CosmosClientWrapper.cs	Executes transactional batches once and throws CosmosTransactionalBatchException on failure (instead of returning a failure result).

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.

[AndriySvyryd]

@copilot Remove the last comment line

[Copilot]

Done — removed the // Signature (SDK 3.x): ... comment line.

[AndriySvyryd]

@copilot Make sure this pattern works in NativeAOT mode

[Copilot]

Fixed. The field initializer now checks RuntimeFeature.IsDynamicCodeSupported before running reflection, so the lookup is skipped entirely in NativeAOT builds (where the private constructor would be trimmed anyway). The reflection itself is extracted into a private helper method FindCosmosExceptionInternalCtor annotated with [UnconditionalSuppressMessage("ReflectionAnalysis", "IL2075", ...)] to suppress any trimmer warning. In both NativeAOT and trimmed builds, CosmosExceptionInternalCtor will be null and CreateCosmosException will use the public constructor as the fallback.