Small-scale MLOps library#

No need for a platform#

What you get without tracking server, cloud or complex setup

• Local UI

• Experiment tracking

• Configuration management

• Data lineage and validation

• Experiment results querying

Local UI#

pip install cascade-ui

Just do cascade ui to get a nice dashboard for your experiments

See more docs on Cascade UI

Experiment tracking#

Configuration management#

Dataset versioning#

Comparison with other MLOps tools#

Cascade is designed for local-first ML development and small teams. Here is how it compares to other popular MLOps tools.

Quickstart#

Here is a simple example of how you can use Cascade to track an sklearn classifier.

pip install cascade-ml

import random
from sklearn.linear_model import LogisticRegression
from cascade.lines import ModelLine
from cascade.utils.sklearn import SkModel

model = SkModel(
    blocks = [
        LogisticRegression()
    ]
)
model.tag("training")
model.describe("Regression model for index page demo")
model.add_metric('acc', random.random())

line = ModelLine("index_demo_line")
line.save(model)

[
    {
        "name": "cascade.utils.sklearn.sk_model.SkModel",
        "description": "Regression model for index page demo",
        "tags": [
            "training"
        ],
        "comments": [],
        "links": [],
        "type": "model",
        "created_at": "2026-09-04T13:54:28.273927+00:00",
        "metrics": [
            {
                "name": "acc",
                "value": 0.20003771067823384,
                "created_at": "2026-09-04T13:58:28.868433+00:00"
            }
        ],
        "params": {},
        "pipeline": "Pipeline(steps=[('0', LogisticRegression())])",
        "path": "/home/ilia/work/cascade/line/00000",
        "slug": "dramatic_gibbon_of_swiftness",
        "saved_at": "2026-09-04T13:59:05.717339+00:00",
        "python_version": "3.12.3 (main, Mar 23 2026, 19:04:32) [GCC 13.3.0]",
        "user": "ilia",
        "host": "your-pc-name",
        "cwd": "/home/ilia/cascade",
        "git_commit": "ddcb2f700d64592c3e433209dff5c6c5d544906f",
        "git_uncommitted_changes": [
            "M cascade/docs/source/index.rst\n?? line/"
        ]
    }
]

Migrating to Cascade#

Cascade can be introduced into an existing ML project without requiring a complete rewrite of the training code. The following examples show how common experiment-tracking frameworks can be replaced by Cascade.

- import aim
+ from cascade.lines import ModelLine
+ from cascade.models import BasicModel

- run = aim.Run()
+ line = ModelLine("line")

- run.add_tag("demo")
- run["hparams"] = {
-     "learning_rate": 1e-5,
-     "batch_size": 32,
-     "epochs": 10,
- }

for epoch in range(10):
+     model = BasicModel()
+     model.params.update({
+            "learning_rate": 1e-5,
+            "batch_size": 32,
+            "epochs": 10,
+        })
+
-     run.track(0.90, name="acc", epoch=epoch)
+     model.add_metric("acc", 0.90)
+     model.tag("demo")
+
+     line.save(model)

- import mlflow
+ from cascade.lines import ModelLine
+ from cascade.models import BasicModel

- with mlflow.start_run():
-     mlflow.log_params({
-         "learning_rate": 1e-5,
-         "batch_size": 32,
-         "epochs": 10,
-     })
-
-     mlflow.set_tag("example", "demo")
-
+ line = ModelLine("line")
for epoch in range(10):
+     model = BasicModel()
+     model.params.update({
+            "learning_rate": 1e-5,
+            "batch_size": 32,
+            "epochs": 10,
+        })
+
-     mlflow.log_metric("acc", 0.90, step=epoch)
+     model.add_metric("acc", 0.90)
+     model.tag("demo")
+
+     line.save(model)

- import wandb
+ from cascade.lines import ModelLine
+ from cascade.models import BasicModel

config = {
    "learning_rate": 1e-5,
    "batch_size": 32,
    "epochs": 10,
}

- with wandb.init(
-     project="project",
-     config=config,
-     tags=["demo"],
- ) as run:

+ line = ModelLine("line")
for epoch in range(10):
+     model = BasicModel()
+     model.params.update(config)
+
-     run.log({"acc": 0.90, "epoch": epoch})
+     model.add_metric("acc", 0.90)
+     model.tag("demo")
+
+     line.save(model)

- from clearml import Task
+ from cascade.lines import ModelLine
+ from cascade.models import BasicModel

config = {
    "learning_rate": 1e-5,
    "batch_size": 32,
    "epochs": 10,
}

- task = Task.init(
-     project_name="project",
-     task_name="training",
- )
- task.connect(config)
- task.add_tags(["demo"])
- logger = task.get_logger()

+ line = ModelLine("line")
for epoch in range(10):
+     model = BasicModel()
+     model.params.update(config)

-     logger.report_scalar(
-         title="metrics",
-         series="acc",
-         value=0.90,
-         iteration=epoch,
-     )

+     model.add_metric("acc", 0.90)
+     model.tag("demo")
+
+     line.save(model)

How to do a thing with Cascade#

• Pipeline building

• Model training

• Track a file

• Track Dataset Errors

• Track experiment logs with Cascade

• Track scikit-learn experiment

Documentation#

Key Principles#

• Elegance - ML code should be about ML with minimum meta-code

• Agility - it should be easy to build prototypes and integrate existing project with Cascade

• Reusability - code should have an ability to be reused in similar projects

• Traceability - everything should have meta data

If you have any questions#

Any contributions are welcome!

Write an issue

Join GitHub discussions

Cascade on X