Skip to content

Building DAGs

Adding nodes

Click + Add Task or + Add Service in the topbar. Each click places a new node on the canvas at an auto-calculated position. Nodes snap to a 20px grid.

A new task node on the canvas with the Configure Node sidebar open

Node configuration

Select a node to open the sidebar. Every field maps directly to a TitanJob property.

Job ID *

Unique identifier for this job within the DAG. Used in dependency wiring, logs, and the visualizer.

  • Must be unique across all nodes in the canvas
  • Alphanumeric, hyphens and underscores allowed
  • Example: preprocess, train-gpu, comp-synthesizer

Script File *

The Python script that runs on the worker. Select from the dropdown (lists all files in perm_files/) or click + Upload to add a new script without leaving the constructor.

Uploaded scripts are saved to perm_files/ and immediately available in the dropdown.

Type

Value Behaviour
run Script runs once and exits. Job completes when process exits 0.
service Process stays alive. Master tracks it as a long-running service.

Requirement

Routes the job to a worker with matching capability. See capability routing.

Args

Command-line arguments passed to the script at launch.

# Script receives: python train.py --epochs 10 --lr 0.001
args: --epochs 10 --lr 0.001

For positional args, order matters and must match the script's sys.argv expectations.

Priority

Integer 1–10. Higher priority jobs are dispatched first when multiple jobs are queued on the same worker. Default: 1.

Delay (seconds)

Wait this many seconds after all dependencies clear before dispatching. Useful for rate limiting or allowing a service to fully start before dependent jobs connect to it. Default: 0.

Sticky Scheduling (affinity)

When enabled, forces this job to run on the same worker that ran its parent job. Useful when jobs share local disk state.

Warning

Do not combine affinity: true with a requirement that differs from the parent job's worker capability. The scheduler needs a worker that satisfies both constraints — if none exists, the job waits indefinitely.

Port (Service type only)

The port the service process listens on. Recorded by the master for service discovery.

Drawing edges (dependencies)

An edge from node A to node B means: B will not start until A has completed successfully.

To draw an edge:

  1. Hover over the source node — a port (●) appears on its right edge
  2. Click and drag from the port
  3. Release over the target node
  4. A blue curved arrow confirms the dependency

To delete an edge: click it (it turns orange/dashed) then press Delete or use the sidebar Delete Edge button.

Orange dashed dependency arrow connecting two nodes

Multi-select

Drag on empty canvas to draw a blue selection rectangle. All nodes within the box are selected.

  • Shift+click a node to add or remove it from the current selection
  • Drag any selected node to move the whole group
  • Delete / Backspace removes all selected nodes and their edges at once

When multiple nodes are selected the sidebar shows a multi-select panel instead of node config. Click a single node to edit it individually.

Duplicate a node

Select a node and press Ctrl+D (or click Copy in the sidebar header). The duplicate appears offset by 40px with -copy appended to the Job ID. All field values are cloned.

Undo / Redo

Action Shortcut
Undo Ctrl+Z
Redo Ctrl+Y or Ctrl+Shift+Z

Undo/Redo tracks structural changes: add node, delete node, add edge, delete edge, duplicate. Drag position changes are not tracked.

The ↩ ↪ buttons in the topbar are disabled (greyed) when there is nothing to undo or redo.

Cycle detection

Before deploying, the constructor runs a DFS cycle check on the edge graph. If a cycle is detected (e.g. A → B → A), deploy is blocked with a "Cycle detected — check your dependencies." error.

Titan's execution model does not support cycles in a DAG. Loops and conditional re-runs are expressed at the orchestrator layer, not inside the DAG structure. See Architecture for the agentic loop pattern.

Node badges

Nodes display small badges for at-a-glance status:

Badge Meaning
HITL (yellow) A human approval gate is configured after this job
svc (blue) Service type node
sticky (orange) Affinity enabled
p:N (blue) Priority above 1

The node border colour reflects its capability requirement (green = GENERAL, purple = GPU, blue = HIGH_MEM, orange = PYTHON).

Live codegen

The Output panel at the bottom of the canvas shows YAML and Python SDK equivalents of the current canvas, updated live.

name: "ml-pipeline"
project: true

jobs:
  - id: "preprocess"
    type: "run"
    file: "preprocess.py"
    args: "dataset-001"

  - id: "train"
    type: "run"
    file: "train.py"
    requirement: "GPU"
    priority: 4
    depends_on: ["preprocess"]

from titan_sdk import TitanClient, TitanJob

client = TitanClient()

preprocess = TitanJob(
    job_id="preprocess",
    filename="preprocess.py",
    args="dataset-001",
)

train = TitanJob(
    job_id="train",
    filename="train.py",
    requirement="GPU",
    priority=4,
    parents=["preprocess"],
)

client.submit_dag("ml-pipeline", [preprocess, train])

Click Copy to copy the current tab's content to clipboard.