/en/blog · writing
Orchestration Without Pain: The History, Architecture, and Triumph of Temporal.io
Deep dive into Temporal.io, a powerful platform for durable execution in distributed systems. Explore its origins at Uber as Cadence, its core architecture based on Event Sourcing, and its triumph as an industry standard, especially in the era of AI agents and LLM pipelines.
- Temporal.io
- Distributed Systems
- Orchestration
- Cadence
- Event Sourcing
- Durable Execution
- Microservices
- AI Agents
- LLM
- Workflow-as-Code
In the world of distributed systems, there are "hot path" tasks-for example, ingesting 100,000 GPS pings per second or quickly reading a line from a cache. Here, Kafka, Redis, and Cassandra reign supreme. But as soon as it comes to business logic that lasts longer than a few milliseconds and consists of a chain of steps, engineers find themselves in the hell of distributed transactions.
A taxi ride, booking a flight, issuing a loan, or launching a multi-stage AI agent-all these processes have one thing in common: they are discrete, can last for hours (or weeks), and must be fault-tolerant. If the network goes down or a microservice crashes in the middle of a transaction, the system must not forget at what stage it was.
For a long time, developers solved this with "reinventing the wheel": tons of queues (SQS/RabbitMQ), databases for storing intermediate states, and complex distributed sagas. Until Temporal.io emerged-a platform that changed the very paradigm of writing distributed code.
🚀 Part 1. How Uber Gave Birth to Cadence (Temporal's Prehistory)
In the mid-2010s, Uber was rapidly transitioning to a microservice architecture. Instead of a single monolith, hundreds of smaller services appeared. And then engineers hit a wall called "the state of distributed workflow".
It turned out that a regular ride request is an extremely complex distributed process:
- 📱 User tapped "Order".
- 💳 Server blocked funds on the card.
- 🔎 Dispatch engine started searching for a driver (this process can take several minutes).
- 🚗 Driver found, car is en route (process lasts 10-15 minutes).
- ✅ Ride completed, need to finalize the charge, send a receipt, and update statistics.
What if Uber's data center goes down at step 4? How to resume the ride exactly from the second where everything stopped?
Uber engineers Maxim Fateev (previously worked on Amazon SWF) and Samar Abbas decided that forcing developers to write retry logic, compensations, and state tracking in every microservice was a path to disaster. They created an internal project called Cadence.
The concept was revolutionary and called Durable Execution. The idea is simple: a developer writes ordinary, sequential code in Go or Java, and the Cadence platform guarantees that this code will execute to completion, even if the server it's running on physically explodes in the middle of execution.
💥 Part 2. The Great Split and the Birth of Temporal.io
Cadence quickly became the heart of Uber. Hundreds of critical processes began to run on it. However, it had two limitations: it was developed strictly for Uber's needs and was tightly coupled to the company's infrastructure.
In 2019, Maxim Fateev and Samar Abbas decided to leave Uber to make this technology available to the whole world. They founded a startup and forked Cadence, naming it Temporal.
What's the difference between Cadence and Temporal? Temporal is a deep evolution of Cadence. The developers completely rewrote the internal API, replaced the custom protocol with gRPC, significantly simplified the SDKs for programming languages, and removed the tight coupling to Cassandra, adding official support for PostgreSQL, MySQL, and SQLite.
The startup immediately attracted the attention of Silicon Valley. Top-tier funds (including Sequoia Capital) believed in Temporal, and investment rounds amounted to tens of millions of dollars. The industry urgently needed a standard for orchestration.
✨ Part 3. The Magic Under the Hood: How It Works?
Most orchestration systems (e.g., Airflow or Camunda) force you to describe processes as JSON/YAML files or to draw "boxes" in visual editors (BPMN).
Temporal took a different approach: Workflow-as-Code. You write ordinary code in your favorite language (Go, Java, TypeScript, Python, .NET).
The platform consists of two main parts:
- Temporal Server: The orchestrator itself (written in Go), which stores Event History and coordinates tasks.
- Workers: Your application, where the workflow code is running.
+-----------------------------------+ | Temporal Server | | (State, Event History, Queues) | +-----------------+-----------------+ | (gRPC) v +-----------------------------------+ | Your App Workers | | (Workflow Code & Activities) | +-----------------------------------+
🔑 The Main Secret: Event Sourcing and Determinism
How does Temporal manage to "resurrect" crashed code on another server without saving heavy memory dumps (snapshotting)?
The answer: Event Sourcing.
When your workflow executes a function (in Temporal terms, this is called an Activity-for example, "issue an invoice"), the Temporal Server records in its database: "Activity #1 started", and then "Activity #1 successfully completed with result X".
If the worker executing this code crashes, Temporal simply takes another worker and starts the same workflow from the beginning. But when the code reaches Activity #1, the Temporal SDK "intercepts" the call, looks into the server's history, sees that this action has already been performed, and instead of actually calling the function, simply returns the ready result from the history instantly!
This process is called Replay. For it to work, the workflow code must be strictly deterministic (you cannot use random numbers or directly query the current time-special Temporal functions are available for this).
📈 Part 4. The Current State of Temporal (2026)
Today, Temporal.io is not just a successful startup, but an industry standard at the peak of its maturity.
🤖 1. Expansion into the AI and Agent System Ecosystem
In 2024-2026, Temporal is experiencing a second rebirth thanks to the boom of AI agents and LLM pipelines.
The operation of modern AI agents is a classic long-running, unpredictable workflow. An agent needs to query an LLM, wait for a response, execute code, check the result, if it failed-try again, ask a human through an interface, etc. A workflow can "hang" for days. Temporal has become an ideal backend for frameworks like LangGraph, and companies like OpenAI and Cursor use Temporal to orchestrate their internal processes.
📦 2. Product Lineup
💻 Temporal Open Source: A free core version that any company can deploy in their Kubernetes cluster.
☁️ Temporal Cloud: A SaaS (Software-as-a-Service) product. Companies no longer need to administer a complex Temporal Server and databases-they pay for the number of executed actions, delegating support to the Temporal team.
🌐 3. Who Uses Temporal Today?
- Uber (continues to use Cadence and is smoothly migrating to Temporal).
- Netflix: Uses it for orchestrating heavy video encoding and content delivery pipelines.
- Stripe & HashiCorp: Entrusted Temporal with their financial transactions and infrastructure deployment processes.
- Snapchat, Coinbase, Datadog and thousands of other tech leaders.
👋 Conclusion
Temporal has brought about a mental shift in distributed systems development. It allowed engineers to stop thinking about "the network might flicker" and databases "timing out".
If in the 2010s, building complex message queue chains and manually configuring distributed sagas was a sign of good practice, by 2026, delegating this routine to specialized Durable Execution engines has become the gold standard. And Temporal.io is the undisputed leader here.