Restate

We love it when users share how they think about Restate and the problems it solves, and how that thinking evolves. It's refreshing to see the project through different lenses, and it helps everyone in the community to understand it more deeply. Thank you Sebastian Hindhede for sharing!

In case sou missed it: Restate has a dedicated section of the docs on how to build 𝗱𝘂𝗿𝗮𝗯𝗹𝗲 𝗮𝗴𝗲𝗻𝘁𝘀 and AI workflows. * Recipes for common patterns and tools * Integrations with many popular AI SDKs (make your existing apps durable) * End-to-end examples https://lnkd.in/eW379VXh

Customers often tell us they started implementing in-house solutions for resiliency and workflow orchestration, typically backed by a database, queues and cron jobs. But this is extremely hard to get right. Even for the best engineering teams. Unkey, an API management platform, recently shared their experience with us. They built their own system for multi-step deployment workflows. But they started hitting classical distributed systems challenges. Old jobs didn't get picked up after a node crash. Each worker would grab a lease (row in a database) to claim a task, but a timing bug made it think another worker still held the lease even when it didn't. Debugging these subtle bugs takes hours. This is exactly why we built Restate. They replaced their entire custom retry and state-tracking system with Restate in less than a day by swapping their internal run functions with restate.run and it just worked. Now they use Restate for: - Multi-step deployment workflows across regions - Concurrency control for critical operations (domain switches happen exactly once) - Durable promises for external events like S3 notifications - Time-delayed operations like scheduled deletions No more race conditions, nasty timing bugs, and other distributed systems complexity. Read the full story: https://lnkd.in/dp4SiXkE What was the toughest distributed systems bug you encountered?

🎉 𝗔𝗻𝗻𝗼𝘂𝗻𝗰𝗶𝗻𝗴 𝗿𝗲𝘀𝘁𝗮𝘁𝗲-𝗼𝗽𝗲𝗿𝗮𝘁𝗼𝗿 𝘃𝟮.𝟭.𝟬! We're excited to release a new version of the Restate Kubernetes operator with a headline feature that simplifies cluster deployment significantly. ✨ 𝗢𝗽𝗲𝗿𝗮𝘁𝗼𝗿-𝗱𝗿𝗶𝘃𝗲𝗻 𝗰𝗹𝘂𝘀𝘁𝗲𝗿 𝗽𝗿𝗼𝘃𝗶𝘀𝗶𝗼𝗻𝗶𝗻𝗴 Previously, deploying a multi-node Restate cluster required manual provisioning via `restatectl provision` after the pods were running. This added friction to the deployment process and made fully automated GitOps workflows more challenging. With v2.1.0, simply set 𝗰𝗹𝘂𝘀𝘁𝗲𝗿.𝗮𝘂𝘁𝗼𝗣𝗿𝗼𝘃𝗶𝘀𝗶𝗼𝗻: 𝘁𝗿𝘂𝗲 in your RestateCluster resource, and the operator handles everything automatically: 1. Waits for pods to be ready 2. Calls the Restate gRPC ProvisionCluster API 3. Tracks provisioning state to prevent duplicate attempts 🎯 This is now the recommended approach for provisioning Restate clusters on Kubernetes. Check out the full release notes and upgrade instructions: 👉 https://lnkd.in/diahdNij #Kubernetes #Restate #DevOps #OpenSource

We're excited to announce that Restate now integrates with Google's Agent Development Kit (ADK) 🎉 AI agents suffer from the same issues as any other backend: rate limits, network failures, and service crashes can kill your agents mid-execution. And maintaining conversation state across interactions usually means managing external databases and complex retry logic. With Restate + Google ADK, you get: ✅ Automatic recovery from failures: your agents pick up exactly where they left off ✅ Persistent conversation history without managing state stores ✅ Human-in-the-loop workflows that survive crashes ✅ Built-in observability across all agent operations and state ✅ Resilient building blocks to easily customize your agents Check out the announcement: https://lnkd.in/dbBG5A7Y Questions? Join us on Discord or Slack. We'd love to help you get started. Code example in the comments! #AI #AIAgents #GoogleADK #DurableExecution #Python #TypeScript

Many application patterns are so much harder to implement than they should be. Here's a perfect example of how Restate helps you out:

These past months, I’ve been working a lot on agents. At their core, agents are interactive backend systems. Agents interact with: - UI / User: sending messages back and forth - LLMs from several providers - Other agents via local handoffs or HTTP calls - MCP servers or remote tools - Humans for tool approvals and feedback - APIs, CRMs, databases: tools can interact with anything - Session stores to persist events - AI Observability tools to persist user feedback Underneath every Agent SDK is a big while-loop that first calls an LLM, that then orders to call one or multiple items on this list, and then calls the LLM again to interpret the results. The ability of agents to interact and act is what makes them powerful but also fragile. Each of these interactions can fail, experience network outages, hit timeouts and rate limits, or take more time to respond. The same issues as with any other backend. Before we can deploy agents to production we need: - Retry logic for each interaction - Persistence and recovery of progress so users don't need to start over - Message queues: for user messages, human-in-the-loop callbacks, agent responses, ... - Storage for events and context like user preferences - Idempotency with deduplication for critical actions - Safe versioning to upgrade agents without breaking ongoing work This is not only for agents which do high-profile stuff like payments, it’s the basics to keep them working in production. And then we are not even talking yet about regional failures, race conditions, zombie processes, network partitions, and all the other super hard stuff. An alternative to all of this, is just using Restate. It handles all of this for you. Each interaction becomes durable and recoverable. 🎁 Restate now has integrations for Vercel AI SDK, OpenAI Agents SDK, and Google ADK. Or build your own agent loops with Restate + LLM calls.

$27,000 in LLM costs. Three weeks of development. Just Gone...puff. Because a pod restarted during human review. This is the hidden tax of stateless agents. Here's what happened: Agent processes a complex claim. Needs human approval. Human is in meetings all day. Server does a routine restart overnight. Morning comes. Agent has amnesia. No memory of the claim. No context. No progress. Three weeks of orchestrated work - just vanished. So the team rebuilds. Adds database checkpoints. State serialization. Recovery logic. Retry handlers. Suddenly your "AI agent" is 80% infrastructure code and 20% actual intelligence. There's a better way. Introducing Durable Execution with Google ADK + Restate 🔥 We've integrated Restate into the Agent Development Kit to give your agents a persistent brain. This isn't "saving to a database." This is code that is fundamentally un-killable. Here's how it works: Restate journals every step of your agent's execution. ⏸️ Pause for days Agent hits request_confirmation() and goes to sleep. Actually sleeps. Zero resources consumed. 💥 Survive anything Server restarts? Pod dies? Network blip? Agent wakes up on the exact line of code where it stopped. 🧠 Context intact No re-prompting the LLM to "remember." State is baked into the execution itself. We built a Claims Processing Agent to prove it. With the Restate Plugin, your agents: ✅ Never lose progress ✅ Sleep for days without burning costs ✅ Recover gracefully from any failure ✅ Turn complex state logic into simple procedural code Deploy it in 5 minutes. Stop paying the hidden tax of fragile agents. Start building agents that survive the real world. This is Day 23 of 25 in Google Cloud's Advent of Agents. Missed previous days? The archive is live. Catch up anytime. ♻️ Repost to share this free and interactive course with your network. And follow this space to stay updated for what's to come.

$27,000 in LLM costs. Three weeks of development. Just Gone...puff. Because a pod restarted during human review. This is the hidden tax of stateless agents. Here's what happened: Agent processes a complex claim. Needs human approval. Human is in meetings all day. Server does a routine restart overnight. Morning comes. Agent has amnesia. No memory of the claim. No context. No progress. Three weeks of orchestrated work - just vanished. So the team rebuilds. Adds database checkpoints. State serialization. Recovery logic. Retry handlers. Suddenly your "AI agent" is 80% infrastructure code and 20% actual intelligence. There's a better way. Introducing Durable Execution with Google ADK + Restate 🔥 We've integrated Restate into the Agent Development Kit to give your agents a persistent brain. This isn't "saving to a database." This is code that is fundamentally un-killable. Here's how it works: Restate journals every step of your agent's execution. ⏸️ Pause for days Agent hits request_confirmation() and goes to sleep. Actually sleeps. Zero resources consumed. 💥 Survive anything Server restarts? Pod dies? Network blip? Agent wakes up on the exact line of code where it stopped. 🧠 Context intact No re-prompting the LLM to "remember." State is baked into the execution itself. We built a Claims Processing Agent to prove it. With the Restate Plugin, your agents: ✅ Never lose progress ✅ Sleep for days without burning costs ✅ Recover gracefully from any failure ✅ Turn complex state logic into simple procedural code Deploy it in 5 minutes. Stop paying the hidden tax of fragile agents. Start building agents that survive the real world. This is Day 23 of 25 in Google Cloud's Advent of Agents. Missed previous days? The archive is live. Catch up anytime. ♻️ Repost to share this free and interactive course with your network. And follow this space to stay updated for what's to come.