The workflow is the implementation of the coordination logic. The Temporal programming framework (aka client library) allows you to write the workflow coordination logic as simple procedural code that uses standard Go data modeling. The client library takes care of the communication between the worker service and the Temporal service, and ensures state persistence between events even in case of worker failures. Furthermore, any particular execution is not tied to a particular worker machine. Different steps of the coordination logic can end up executing on different worker instances, with the framework ensuring that the necessary state is recreated on the worker executing the step.
However, in order to facilitate this operational model, both the Temporal programming framework and the managed service impose some requirements and restrictions on the implementation of the coordination logic. The details of these requirements and restrictions are described in the Implementation section below.
The sample code below shows a simple implementation of a workflow that executes one activity. The workflow also passes the sole parameter it receives as part of its initialization as a parameter to the activity.
In the Temporal programing model, a workflow is implemented with a function. The function declaration specifies the parameters the workflow accepts as well as any values it might return.
Let’s deconstruct the declaration above:
- The first parameter to the function is ctx workflow.Context. This is a required parameter for all workflow functions and is used by the Temporal client library to pass execution context. Virtually all the client library functions that are callable from the workflow functions require this ctx parameter. This context parameter is the same concept as the standard context.Context provided by Go. The only difference between workflow.Context and context.Context is that the Done() function in workflow.Context returns workflow.Channel instead the standard go chan.
- The second parameter, string, is a custom workflow parameter that can be used to pass data into the workflow on start. A workflow can have one or more such parameters. All parameters to a workflow function must be serializable, which essentially means that params can’t be channels, functions, variadic, or unsafe pointers.
- Since it only declares error as the return value, this means that the workflow does not return a value. The error return value is used to indicate an error was encountered during execution and the workflow should be terminated.
In order to support the synchronous and sequential programming model for the workflow implementation, there are certain restrictions and requirements on how the workflow implementation must behave in order to guarantee correctness. The requirements are that:
- Execution must be deterministic
- Execution must be idempotent
A straightforward way to think about these requirements is that the workflow code is as follows:
- Workflow code can only read and manipulate local state or state received as return values from Temporal client library functions.
- Workflow code should not affect changes in external systems other than through invocation of activities.
- Workflow code should interact with time only through the functions provided by the Temporal client library (i.e. workflow.Now(), workflow.Sleep()).
- Workflow code should not create and interact with goroutines directly, it should instead use the functions provided by the Temporal client library (i.e., workflow.Go() instead of go, workflow.Channel instead of chan, workflow.Selector instead of select).
- Workflow code should do all logging via the logger provided by the Temporal client library (i.e., workflow.GetLogger()).
- Workflow code should not iterate over maps using range because the order of map iteration is randomized.
Now that we have laid the ground rules, we can take a look at some of the special functions and types used for writing Temporal workflows and how to implement some common patterns.
Special Temporal SDK functions and types
The Temporal client library provides a number of functions and types as alternatives to some native Go functions and types. Usage of these replacement functions/types is necessary in order to ensure that the workflow code execution is deterministic and repeatable within an execution context.
Coroutine related constructs:
- workflow.Go : This is a replacement for the the go statement.
- workflow.Channel : This is a replacement for the native chan type. Temporal provides support for both buffered and unbuffered channels.
- workflow.Selector : This is a replacement for the select statement.
Time related functions:
- workflow.Now() : This is a replacement for time.Now().
- workflow.Sleep() : This is a replacement for time.Sleep().
Failing a workflow
To mark a workflow as failed, all that needs to happen is for the workflow function to return an error via the err return value.
For some client code to be able to invoke a workflow type, the worker process needs to be aware of all the implementations it has access to. A workflow is registered with the following call:
This call essentially creates an in-memory mapping inside the worker process between the fully qualified function name and the implementation. It is safe to call this registration method from an init() function. If the worker receives tasks for a workflow type it does not know, it will fail that task. However, the failure of the task will not cause the entire workflow to fail.