Workflows

Workflows are a structured framework used to automate and orchestrate a series of jobs, called activities, in a logical sequence to accomplish a specific goal. They provide a modular and highly configurable mechanism for defining and managing complex operations, such as processing data, leveraging machine learning models, coordinating computational resources, and performing distributed tasks efficiently.

At their core, workflows act as containers for Activities, which represent individual steps in the process. These activities work together by passing inputs, producing outputs, and integrating with various systems, tools, or configurations to create a seamless execution pipeline.

Key Components

1. Activities:
   • An Activity is a foundational building block of a workflow. Each activity represents a discrete operation that performs a specific task, such as training a model, making predictions, or processing analytical results.
   • Activities are extremely versatile and can:
     • Accept inputs (parameters or data) and produce outputs to enable seamless interaction between steps in the workflow.
     • Configure and interact with machine learning models, storage systems, or prompts during execution.
     • Be executed sequentially or in parallel, depending on their execution group assignment.
   • This modularity ensures each activity is designed for a specific purpose, while also enabling the reuse of activities across multiple workflows.
2. Configuration:
   • Both workflows and their activities are built around configurations written in JSON. These configurations define the requirements, operational settings, and execution details for each workflow and activity.
   • Configurations make workflows highly adaptable to different environments and use cases by customizing their behavior without needing to modify the underlying code.
3. Queue Management:
   • Each workflow and its associated activities utilize a queue system to organize execution and distribute tasks across compute resources.
   • This system ensures that workflows can scale efficiently in distributed environments by matching activities to appropriate resources, whether that’s a GPU-accelerated system for computationally heavy tasks or CPU-focused nodes for lighter processing requirements.
4. Scheduling:
   • You can define different repetitive schedules for each workflow (coming soon). You can define RRULEs to define how the workflow should repeat over a given amount of time. For more details, view the data model.

How Workflows Work

1. Executing Activities:
   • Workflows follow a predefined structure where Activities are executed either sequentially or in parallel, as determined by their grouping.
   • Execution is handled by Workflow Runners, which are specialized runtime programs capable of processing workflow jobs. These runners ensure that each activity operates on the right compute resource and under the appropriate conditions:
     • For example, tasks like training a machine learning model may require Python, as it is well-suited for ML workflows due to its vast ecosystem of libraries and tools like TensorFlow or PyTorch. These tasks are often executed on GPU-powered compute resources.
     • On the other hand, tasks like running predictions, inference, or processing outputs could be executed efficiently on CPU resources using languages like Rust, Kotlin, or C#, which are ideal for their performance, reliability, and system-level operations.
   • This execution model allows workflows to fluidly adapt to the strengths of different programming languages and hardware environments, optimizing the performance of each step in the process.
2. Exchanging Data Between Activities:
   • Workflow activities share data using well-defined inputs and outputs. Each activity produces outputs that subsequent activities consume, enabling a seamless flow of information along the workflow pipeline.
   • For workflows that handle large files or require more advanced data management, activities can also integrate with storage systems to read or write shared data efficiently.
3. Models:
   • Activities can interact with machine learning models or large language models (LLMs) through WorkflowActivityModel, enabling workflows to incorporate advanced AI operations seamlessly.
   • These interactions are highly configurable, ensuring that models are used with the right datasets and in the appropriate contexts. For example:
     • One activity could involve training a recommendation model based on historic user data.
     • A subsequent activity might use this trained model to generate personalized recommendations.
4. Prompts:
   • Workflows can include tasks involving large language model (LLM) prompts. These prompts are instructions or queries provided to pre-trained LLMs (e.g., GPT-like models) to perform operations such as natural language processing (NLP), comprehension, or generation tasks.
   • LLM-based activities allow workflows to leverage cutting-edge AI capabilities for operations like summarization, contextual analysis, or content generation.
5. Storage Systems:
   • Workflow activities can read from and write to a variety of storage systems, ensuring efficient data management and integration across distributed environments. This capability enables workflows to handle datasets, models, and results without bottlenecks, whether the storage is local, remote, or cloud-based.
   • Examples of Storage Systems might include Vector Databases to support RAG, or object storage to support management of large files.

Purpose and Advantages of Workflows

1. Modularity:
   • Workflows are highly modular, breaking down processes into distinct, manageable activities. This makes them easier to design, maintain, and test while offering the ability to reuse components in new workflows.
2. Flexibility and Configurability:
   • The JSON-based configuration system enables workflows to be highly adaptable to unique use cases. Users can fine-tune configurations without changing the underlying logic, offering flexibility to support a range of tasks.
3. Scalability:
   • By leveraging queues and Workflow Runners, workflows can operate efficiently in distributed systems, ensuring tasks are executed on the most appropriate compute resources without overwhelming the system.
4. Optimal Resource Utilization:
   • Workflows support a mixed environment where different languages or hardware are utilized for specific tasks:
     • Python and GPUs handle computationally heavy tasks like model training, where libraries and compute acceleration are critical.
     • Other languages on CPUs handle tasks like inference or application logic, where high performance, memory safety, and streamlined resource usage are essential.
   • This ensures workflows maximize the efficiency and capabilities of available resources.
5. Interoperability:
   • Workflows seamlessly integrate with advanced systems, such as machine learning models, large language models, and storage solutions. This interoperability makes workflows highly suitable for scenarios involving automation, data processing, or AI-powered pipelines.
6. Seamless AI and NLP Integration:
   • The inclusion of LLM prompts enables workflows to perform advanced natural language operations easily, further extending their applicability in AI and data-driven environments.

In conclusion, workflows provide a cohesive and efficient way to define, organize, and execute complex processes. By combining modularity, flexibility, and resource optimization, workflows are highly adaptable to a range of use cases, including machine learning pipelines, distributed operations, and AI-driven tasks. Their ability to leverage languages and computational environments for specific purposes makes them an indispensable tool for solving real-world challenges in diverse domains.

Base Data Model

type Workflow {
    activities: [WorkflowActivity!]!
    configuration: JSON!
    description: String!
    id: String!
    name: String!
    queue: String!
}

type WorkflowActivity {
    activityId: String!
    configuration: JSON!
    executionGroup: Int!
    id: Int!
    inputs: [WorkflowActivityParameter!]!
    models: [WorkflowActivityModel!]!
    outputs: [WorkflowActivityParameter!]!
    prompts: [WorkflowActivityPrompt!]!
    queue: String!
    storageSystems: [WorkflowActivityStorageSystem!]!
}

type WorkflowActivityModel {
    configuration: JSON!
    model: Model!
}

type WorkflowActivityParameter {
    name: String!
    value: String!
}

type WorkflowActivityPrompt {
    configuration: JSON
    prompt: Prompt!
}

type WorkflowActivityStorageSystem {
    configuration: JSON
    system: StorageSystem!
}

Traits

Traits are a way to describe content and how they relate to workflows. They help remove the burden of keeping track of which workflows should be assigned to content. Often, Bosca can just automatically pick the right trait. But, there are some types that currently are served by being manually chosen. For instance, Bosca can detect when you upload an MP3 or MP4 and pick the right trait(s) for you. But, if you upload a DBL Bundle, you would need to choose that trait.

Traits Data Model

type Trait {
    description: String!
    id: String!
    name: String!
    contentTypes: [String!]!
    workflowIds: [String!]!
    workflows: [Workflow!]!
}

States & Transitions

Workflows are designed to help manage the state and transitions of both Collections and Metadata. They allow you to define valid transitions, such as moving from draft to published. Additionally, workflows can specify actions to be performed while entering or exiting states.

For a transition to be successful:

1. Entry workflows for the new state must execute successfully.
2. Exit workflows for the prior state must also complete successfully.

Once all conditions are met, the transition completes, and any workflows associated with the final state are executed.

States & Transition Data Models

enum WorkflowStateType {
    APPROVAL
    APPROVED
    DRAFT
    FAILURE
    PENDING
    PROCESSING
    PUBLISHED
}

type WorkflowState {
    description: String!
    configuration: JSON
    entryWorkflowId: String
    exitWorkflowId: String
    id: String!
    name: String!
    type: WorkflowStateType!
    workflowId: String
}

type Transition {
    description: String!
    fromStateId: String!
    toStateId: String!
}

Executions

Workflow execution is the process by which workflows are carried out, leveraging Workflow Execution Plans and Workflow Jobs.

• A Workflow Execution Plan defines the entire workflow, including its structure, associated activities, and execution flow.
• A Workflow Job represents an actionable Activity within the plan, providing detailed instructions for runners.

Key Concepts

1. Workflow Jobs and Children Workflows
   Activities within a workflow (jobs) can initiate child workflows, which then integrate back into the main workflow. A workflow job does not complete until all child workflows have finished execution.
2. Job Status and Reporting
   Runners, which execute jobs, must coordinate with the Bosca Server throughout execution:
   • Notify upon completion (success or failure).
   • Provide periodic updates to indicate active execution.
3. Concurrent Job Execution
   Jobs within a workflow can run concurrently, provided they share the same execution group. Execution groups define the order and concurrency behavior (along with queues) for workflow jobs.

Execution Plan Data Model

type WorkflowExecutionPlan {
    collectionId: String
    complete: [Int!]!
    context: JSON!
    currentExecutionGroup: [Int!]!
    error: String
    failed: [Int!]!
    id: WorkflowExecutionId!
    jobs: [WorkflowJob!]!
    metadata: Metadata
    metadataId: String
    metadataVersion: Int
    next: Int
    parent: WorkflowJobId
    pending: [Int!]!
    running: [Int!]!
    supplementaryId: String
    workflow: Workflow!
}

type WorkflowJob {
    activity: Activity!
    children: [WorkflowExecutionId!]!
    collection: Collection
    collectionId: String
    completedChildren: [WorkflowExecutionId!]!
    context: JSON
    error: String
    failedChildren: [WorkflowExecutionId!]!
    id: WorkflowJobId!
    metadata: Metadata
    metadataVersion: Int
    models: [WorkflowActivityModel!]!
    planId: WorkflowExecutionId!
    prompts: [WorkflowActivityPrompt!]!
    storageSystems: [WorkflowActivityStorageSystem!]!
    supplementaryId: String
    workflow: Workflow!
    workflowActivity: WorkflowActivity!
}