Dapr 1.19 Adds Workflow History Deltas for Cached Instances

The Dapr project announced version 1.19 on Oct. 22, adding incremental workflow-history updates and optional Scheduler-based actor placement for deployments using the runtime.

A worker receives only newly added workflow events when it supports stateful history and already has the instance cached. It rebuilds the history locally instead of receiving the full record on each turn. The update supports the Python, Java, .NET and Go SDKs. If the worker reconnects, the sidecar sends the full history.

A separate preview feature, WorkflowsFastPath, changes how the runtime drives workflow wake-ups and activities. It is off by default. Unlike history caching, which changes what is sent to a worker, the fast path changes how the runtime handles those wake-ups and activities. A benchmark recorded median workflow-step latency of 153 milliseconds, down from 553 milliseconds.

Separate performance tests compared nightly v1.19 results with the published v1.18.0 run across eight runs from Sept. 23 to Oct. 2. Workflow throughput was 40% higher in TestWorkflowWithConstantVUs and 32% higher in TestSeriesWorkflowWithMaxVUs. Jobs-scheduling p99 latency fell from 39.5 milliseconds to 26.3 milliseconds. Actor-reminder registration p99 latency declined from 148 milliseconds to 43 milliseconds.

Operators can also configure Scheduler to handle actor placement, removing the standalone Placement service from that deployment. The option is off by default, and every sidecar must be upgraded to v1.19 or later before it is enabled. This setting is separate from the workflow-history update and fast path.

Version 1.19 also introduced alpha Binary Store, search and vector APIs.