System Requirements

  • Python version 3.10 or 3.11
  • Mac (ARM or x64), Windows (x64), or Linux (x64)

What’s Vinyl?

Vinyl is Turntable’s framework for developing analytics infrastructure. The only thing you need to know is basic python.

Vinyl is cross-dialect and integrates with your existing data stack. Vinyl can be accessed in places like BI tools, AI agents or data products and works the same if you have 10,000 rows of data or 100M.

Vinyl constructs SQL queries under the hood to avoid pulling all of your data onto your machine and it supports data sources from cloud storage services like S3, databases like Postgres or data warehouses like BigQuery. Vinyl also supports reading from dbt projects. Learn more about our supported connections

Setup your Environment

pip install poetry

First, in a new directory, create a poetry project with Vinyl as a dependency.

poetry init -n --python ">=3.10, <3.12" --dependency vinyl --dev-dependency "vinyl[dev]" && poetry install && poetry shell

If you have an existing poetry project, run poetry add "vinyl[dev]" and make sure you’re in a poetry shell for the next step.

For now, Vinyl projects must also be poetry projects. We’re working on adding support for other package managers.

Create a Project

vinyl login

Now, create a vinyl project with whatever name you’d like. We’ll use my_shop for this example.

vinyl init my_shop

This creates a Vinyl project for you with the following structure:

my_shop/
  |- models/
  |- data/
  |- sources/
  |- resources.py

For the best experience using Vinyl, make sure to change your Python interpreter to your poetry environment’s python. This will provide full autocomplete, syntax highlighting, and renaming capabilities.

Generate Sources

Sources are typed data schemas that help Vinyl understand what data is available. Sources can data warehouses (Snowflake, Bigquery), data lakes (Delta), cloud storage (S3, GCS) or even local files. See all of the supported connections here.

Our demo project contains 3 sample csvs from the Ecuador grocery store dataset under the data/ folder:

  • stores.csv: A list of 54 grocery stores and their locations
  • store_num_transactions.csv: The number of transactions at each store over time (2013-2017)
  • holiday_events.csv: holiday dates in Ecuador (2013-2017)

First, create a Resource. A Resource serves as your connection to locations where you want to access source data.

Inside my_shop/resources.py There’s a local resource created for you to work with our local demo files.

my_shop/resources.py
from vinyl.lib.asset import resource
from vinyl.lib.connect import FileConnector


@resource
def local_filesystem():
    return FileConnector(path = "data")

We can now generate source files for our upstream data. Use this Vinyl CLI command to generate sources:

vinyl sources generate

The result should be 3 files under sources/local_filesystem/ that map to each csv file:

  • stores.py
  • holiday_events.py
  • store_num_transactions.py

Sources help validate breaking changes, propagate metadata and power development ergonomics like column-level lineage and column autocomplete. Federated joins are supported for most connectors (ex: BigQuery data joined with Postgres) for data across resources.

Let’s a take a look at stores.py:

my_store/sources/local_filesystem/stores.py
from vinyl import Field, source # noqa F401
from vinyl import types as t # noqa F401

from my_shop.resources import local_filesystem # noqa F401


@source(resource=local_filesystem)
class Stores:
    _table = "stores"
    _unique_name = "local_filesystem.Stores"
    _path = "data/stores.csv"
    _row_count = 54

    store_nbr: t.Int64(nullable=True)
    city: t.String(nullable=True)
    state: t.String(nullable=True)
    type: t.String(nullable=True)
    cluster: t.Int64(nullable=True)

In just a few lines of python and a @source annotation, we get a typed schema, column-level lineage support, powerful autocomplete and query validation across many different SQL dialects and resource types.

Now let’s do something with this source data.

Create a Model

Models are semantic representations of your data. Models can be used in notebooks for analysis, embedded in BI tools or data products and referenced in other parts of your project. Vinyl provides a wide range of data transforms and complex SQL operations out of the box.

We’ve created some models for you at models/models.py. Let’s take a look at top_stores:

models/models.py
from vinyl import M, T, join, metric, model

@model(deps=[Stores, StoreNumTransactions])
def top_stores(stores: T, txns: T) -> T:
    j = join(store, txns, on = ["store_nbr"])
    j.aggregate(
        cols = {"num_transactions": j.transactions.sum()},
        by = [j.store_nbr, j.city, j.state],
    )
    j.sort(
        by =  -j.num_transactions
    )
    return j

Models help you structure and transform data in a declarative way. Upstream data is passed in as plain python functions or classes as dependencies in the @model annotation. In this example, we pass in the Stores source and StoreNumTransactions sources from earlier so we can aggregate the data to find the stores with the most transactions.

This kind of data work usually involves a lot of complex SQL. What makes Vinyl different here is our pipelined approach to data modeling, a simple API for performing transform operations, and cross-dialect SQL support. You can take the same model code above, change the source from a local file to Snowflake with the same schema and you’ll get the same result.

Now, let’s preview what this looks like using Vinyl’s built-in data preview tool.

vinyl preview model --name top_stores

preview

Vinyl’s interactive data preview tool works in the terminal and comes in handy when you want to preview changes made to your models.

Let’s take a look at another model in our project.

models/models.py
@model(deps=[StoreNumTransactions, Stores])
def store_txns(txns: T, stores: T) -> T:
    table = join(stores, txns, on = [stores.store_nbr == txns.store_nbr])
    return table

This does a simple join across the two source tables, StoreNumTransactions and Stores. Let’s preview this data using:

vinyl preview model --name store_txns

This model contains the number of transactions per day per store, along with some city and state information for each store.

preview_metric

Time series data is notoriously hard to model. You have to think about time buckets, aggregation windows, missing dates and dimensions. Oftentimes your first guess (weekly sales in the last 30 days) ends up being wrong when someone wants sales data by quarter.

Luckily, Vinyl has a powerful abstraction for handling time series data called Metrics.

Create a Metric

Metrics are a powerful feature in Vinyl that allows for tracking time series data. Metrics auto-generate complicated timeseries SQL for you and can be queried across dimensions and time buckets dynamically.

Create a metric by passing in a MetricStore to your function and adding a @metric annotation. We’ll reference the same store_txns model from above.

@metric(deps=[store_txns, MetricStore])
def sales_metrics(table: T, metric_store: M) -> M:
    metric_store.metric(
        cols = {
            "total_txns": table.transactions.sum(),
            "average_txns": table.transactions.mean(),
        },
        ts = table.date.cast("timestamp"),
        by = [table.store_nbr, table.city,table.state],
    )
    return metric_store

Let’s preview our metric using the CLI command:

vinyl preview metric \
    --name sales_metrics \
    --grain days=7

preview_metric

That’s pretty neat! In just a few lines of python, you’ve generated a complicated SQL query to show weekly transaction data that can be sliced across many dimensions and time grains at query time.

Let’s try another time grain. Instead of weekly transactions, let’s look at monthly transactions:

vinyl preview metric \
     --name sales_metrics \
     --grain months=1

preview_metric

This show us the monthly transactions for each store. To learn more about all of the different time grains and dimensions you can use with metrics, check out the Metrics documentation.

Deployment

While Vinyl is a great data modeling and analysis tool on it’s own, it’s also designed to be accessed in places like BI tools, AI agents or data products.

You can orchestrate and deploy your Vinyl project using the CLI. By default, Vinyl runs workloads and serves data locally. For production use caes, learn more about our deployment options.

Let’s get our project ready to be served by running the deploy command:

vinyl project deploy

Vinyl uses the latest query engines and intelligent caching to support sub second queries on 10s of millions of rows of data. Try serving your project using the serve command:

vinyl project serve

serve

Once the server starts up, you can serve models and metrics via a Postgres connector or HTTP API call. You can query your deployed Vinyl project in numerous BI and data tools (learn more about destinations here).

Querying your Project using Postgres

Vinyl supports fast sub second queries via its built-in postgres proxy. Connect to your deployed project using a Postgres client like psql.

psql -h 0.0.0.0  -p 5433 \
    -c "select * from top_stores limit 10";

Integrate your Project with an HTTP API

You can also integrate your Vinyl project with other services or data products. Vinyl supports querying models and metrics via an HTTP API.

curl -H "Content-Type: application/json" \
'http://localhost:8000/metrics/sales_metrics?grain=months=1&limit=10'

Next Steps

Want to learn more about Vinyl? Check out our concepts and reference documentation.

Have a question for the team or want to learn more about running Vinyl in production? Reach out to us at team@turntable.so or join our Slack community.