Skip to content

✋ Human-in-the-Loop & Approvals

What you'll learn

How to pause pipelines for manual review, safety checks, or compliance approvals. Not everything should be automated — deployment to production often needs a human "thumbs up."

Inject human intelligence into automated workflows with approval gates that pause execution and wait for human decisions.

Why Human Review Matters 🤔

Use Case Example
Safety "Does this model output look safe?"
Compliance "Has Legal approved this dataset?"
Quality Assurance "Is the generated image high quality?"
Cost Control "Approve spending $500 on this training run?"
Deployment "Ready to deploy to production?"

✋ Approval Steps

Insert an approval step anywhere in your pipeline to pause execution:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
from flowyml import Pipeline, step, approval

pipeline = Pipeline("deployment_pipeline")

@step
def train():
    return model

# 🛑 Pause here! Wait for human approval
approve_deploy = approval(
    name="approve_deploy",
    approver="lead-data-scientist",
    timeout_seconds=86400,  # 24 hours to approve
)

@step
def deploy(model):
    # Only runs if approved
    production.deploy(model)

pipeline.add_step(train)
pipeline.add_step(approve_deploy)
pipeline.add_step(deploy)

How It Works

sequenceDiagram
    participant Pipeline
    participant Approval as approval step
    participant Human
    participant Deploy as deploy step

    Pipeline->>Approval: Execute
    Approval-->>Human: 🔔 Notification (Slack/Email)
    Note over Approval: ⏸️ Pipeline paused
    Human->>Approval: ✅ Approve / ❌ Reject
    alt Approved
        Approval->>Deploy: Continue pipeline
    else Rejected
        Approval->>Pipeline: Pipeline stopped
    end

Interactive CLI Approval 💻

When running locally, the pipeline pauses and prompts in the terminal:

1
2
3
4
5
6
7
8
$ flowyml run deployment_pipeline

...
[INFO] Step 'train' completed.
[WARN] ✋ Step 'approve_deploy' requires approval.
       Waiting for approval from: lead-data-scientist
       Timeout: 86400s
       Approve execution? [y/N]:

Auto-Approval Logic 🤖

Define conditions for automatic approval — useful for CI/CD environments or non-production stages:

1
2
3
4
5
6
7
import os

approve_deploy = approval(
    name="approve_deploy",
    approver="ml-team",
    auto_approve_if=lambda: os.getenv("ENVIRONMENT") == "staging",
)
Environment Behavior
staging Auto-approved, no human needed
production Pauses and waits for human approval

Approval Parameters ⚙️

Parameter Type Default Description
name str required Step name
approver str required Team or person to approve
timeout_seconds int 3600 Max wait time (default: 1 hour)
auto_approve_if Callable None Lambda for conditional auto-approval

Real-World Examples 🌍

Multi-Stage Approval

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
pipeline = Pipeline("regulated_deployment")
pipeline.add_step(train_model)

# Stage 1: ML team reviews quality
pipeline.add_step(approval(
    name="ml_review",
    approver="ml-team",
    timeout_seconds=86400,
))

# Stage 2: Compliance reviews fairness
pipeline.add_step(approval(
    name="compliance_review",
    approver="compliance-team",
    timeout_seconds=172800,  # 48 hours
))

pipeline.add_step(deploy_to_production)

Cost Approval Gate

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
@step(outputs=["estimated_cost"])
def estimate_training_cost(config):
    return calculate_gpu_cost(config)

pipeline.add_step(estimate_training_cost)
pipeline.add_step(approval(
    name="cost_approval",
    approver="engineering-manager",
    auto_approve_if=lambda: ctx["estimated_cost"] < 100,  # Auto-approve under $100
))
pipeline.add_step(train_expensive_model)

📧 Notifications

Approvers can be notified via configured channels when their attention is required. See Notifications for setup.

Combine with notifications

Configure Slack or email notifications so approvers get pinged immediately — don't make them check a dashboard.