Diagrid’s Dapr 1.19 lightens AI agent workflows and drops a control-plane service

Software development firm Diagrid unveiled Dapr 1.19, an update aimed at streamlining AI agent workflows and eliminating a control-plane service. The latest version curtails the overhead of durable workflows in AI agents, reducing memory usage and latency. Instead of transmitting the complete execution history on each step, workers now retain their existing data and receive new events.

This adjustment benefits workers running for extended periods or executing numerous parallel tasks, as it prevents the payload from escalating. The reduction in transferred data applies solely to workers that support it, which are primarily those using Dapr's Java and Python software development kits.

Another workflow enhancement, WorkflowsFastPath, is an opt-in preview feature that discontinues the transmission of every workflow event through a durable scheduler reminder. Instead, wake-ups and local activities are managed on-site. The reminder pathway remains as a fallback for recovery. Benchmarks from Diagrid's engine demonstrate that enabling WorkflowsFastPath can reduce median step latency by 400 milliseconds.

Two services that constituted Dapr's control plane until now are being consolidated. Placement, which distributes the actor placement table, has been discontinued. The Scheduler service, which previously managed jobs, reminders, and workflow scheduling, now oversees actor placement as well. This change reduces the need for operators to deploy and maintain a stateful service and its Raft consensus cluster.

Consequently, reminders become faster as the Scheduler identifies which sidecar is responsible for each actor. Scaling events are also less disruptive since a new host joining or leaving only temporarily halts placement for the affected actor types, while the rest continue to serve traffic.

In addition to these workflow and control-plane improvements, Dapr 1.19 introduces new alpha building blocks designed for AI application developers. A Vector building block offers a portable application programming interface for storing and querying embeddings using similarity search, with Meilisearch as its initial supported component.

Binary Store facilitates the streaming of large files, such as model outputs, in chunks to prevent them from occupying the entire memory. Currently, its supported components include Amazon S3 and object stores from Microsoft Azure, Google Cloud, and Oracle. For developers working with .NET, Actors.Next, a new actor SDK, is included.

Actors.Next generates proxies and registration code at compile time, decreasing runtime overhead and enabling actor applications to compile with Native AOT. Dapr 1.19 is now available. WorkflowsFastPath is initially disabled, and the Scheduler placement mode can be enabled through a Helm setting on every sidecar running 1.19, without the need for sidecar restarts.

Diagrid's Dapr Ops Dashboard allows Kubernetes users to manage the upgrade, typically within 24 hours. The company also noted that the Dapr Ops Dashboard will support new releases rapidly.