Workflows

Workflows break long-running work into discrete steps. Each step persists its output to a shared state bag in the database. If PHP dies mid-step, the worker retries that step with all prior state intact.

What are workflows?

A workflow is an ordered sequence of steps. Each step is a PHP class that receives the accumulated state from all previous steps and returns new data to merge into the state. The workflow advances one step at a time, with each transition wrapped in an atomic MySQL transaction.

use Queuety\Queuety;

$workflow_id = Queuety::workflow( 'generate_report' )
    ->then( FetchDataHandler::class )
    ->then( CallLLMHandler::class )
    ->then( FormatOutputHandler::class )
    ->dispatch( [ 'user_id' => 42 ] );

State accumulation

The initial payload you pass to dispatch() becomes the starting state. Each step's return value is merged into the state via array_merge:

dispatch(['user_id' => 42])

Step 0: FetchDataHandler   -> returns ['user_name' => 'Alice']
  state: {user_id: 42, user_name: 'Alice'}

Step 1: CallLLMHandler     -> returns ['llm_response' => '...']
  state: {user_id: 42, user_name: 'Alice', llm_response: '...'}

Step 2: FormatOutputHandler -> returns ['report_url' => '/reports/42.pdf']
  state: {user_id: 42, user_name: 'Alice', llm_response: '...', report_url: '/reports/42.pdf'}

Every step sees the full accumulated state. This makes it easy to pass data between steps without any external storage.

Resumability

Workflows are durable because the state is persisted to the database after every step. If the worker crashes during step 1, the workflow restarts from step 1 with step 0's output already in the state.

The step boundary is a single MySQL transaction that atomically:

1. Updates the workflow state with the step's output
2. Marks the current step's job as completed
3. Enqueues the next step's job

This guarantees that no state is ever lost, even in the face of crashes, timeouts, or PHP fatal errors.

Workflow lifecycle

dispatch() -> running -> step 0 -> step 1 -> ... -> step N -> completed
                  |                   |
                  |                   +-> failed -> retry -> running (from failed step)
                  |
                  +-> paused -> resumed -> running
                  |
                  +-> waiting_signal -> signal received -> running
                  |
                  +-> waiting_workflow -> dependency settles -> running
                  |
                  +-> cancelled (via cancel_workflow)

Managing workflows

Check status:

$state = Queuety::workflow_status( $workflow_id );
echo $state->status->value;   // 'running'
echo $state->current_step;    // 1
echo $state->total_steps;     // 3
print_r( $state->waiting_for ); // ['42'] while waiting on another workflow
print_r( $state->wait_details ); // matched/remaining blockers, result key, correlation info

Pause and resume:

Queuety::pause_workflow( $workflow_id );
Queuety::resume_workflow( $workflow_id );

Retry a failed workflow from its failed step:

Queuety::retry_workflow( $workflow_id );

Cancel a workflow and run cleanup handlers:

Queuety::cancel_workflow( $workflow_id );

See Cancellation for details on cleanup handlers and what happens when a workflow is cancelled.

Workflow options

Set queue, priority, and retry behavior for all steps in a workflow:

use Queuety\Enums\Priority;

Queuety::workflow( 'important_pipeline' )
    ->on_queue( 'pipelines' )
    ->with_priority( Priority::High )
    ->max_attempts( 5 )
    ->on_cancel( PipelineCleanupHandler::class )
    ->prune_state_after( 3 )
    ->then( StepA::class )
    ->then( StepB::class )
    ->dispatch( $payload );

The on_cancel() method registers a cleanup handler that runs when the workflow is cancelled. See Cancellation.

The prune_state_after() method enables automatic removal of old step outputs to keep the workflow state bounded. See State Pruning.

Use compensate_with() to attach a per-step compensation handler. Use compensate_on_failure() when you want completed step compensations to run automatically after a workflow failure, not just on explicit cancellation. See Compensation.

Use workflow guardrails when you need safer long-running orchestration: version() tags a definition with an application-level version, Queuety also stores a deterministic definition hash for each run, idempotency_key() makes dispatch durable across retries or duplicate requests, and max_transitions(), max_fan_out_items(), and max_state_bytes() fail runaway runs before they spiral. See Guardrails.

Step types

Queuety supports eleven types of steps within a workflow:

• Sequential chains using ->then() for one-at-a-time execution
• Parallel groups using ->parallel() for concurrent execution
• Dynamic fan-out using ->fan_out() for runtime-discovered branch work and explicit join semantics
• Async handoffs using ->spawn_workflows() or ->spawn_agents() to dispatch independent child workflows from runtime-discovered items
• Conditional branching using named steps and _goto
• Loops using ->repeat_until() and ->repeat_while() for durable back-edges to earlier named steps
• Sub-workflows using ->sub_workflow() for nested composition
• Durable timers using ->sleep() for time-based delays
• Signals using ->wait_for_signal(), ->wait_for_signals(), ->await_approval(), ->await_input(), and ->await_decision() for external event triggers
• Workflow dependencies using ->await_workflow(), ->await_workflows(), and ->await_agents() for cross-workflow coordination
• Streaming steps using the StreamingStep interface for durable chunk-based output

Common patterns

If you are building autonomous or semi-autonomous systems, start with Agent Orchestration. It shows how fan_out(), spawn_agents(), await_agents(), await_workflow(), and await_decision() fit together in real planner/executor and human-review flows.