Spice.ai SMART demo

The SMART demo shows getting started with Spice in 5 mins.

Machine Setup

# Clone this repository
git clone [email protected]:lukekim/smart-demo.git
cd smart-demo

# Setup Python Virtual Environment
python -m venv venv

# Install the SpicePy Python package
pip install git+https://github.com/spiceai/[email protected]

Edit demo.py and set the API key on line 6.

E.g.

SPICEAI_API_KEY = '3232|f42bb127dceb48cab83437264897b06a'

Data demo

Step 1. Show cloud data warehouse (CWD) performance in the Spice.ai Playground

Navigate to the Spice.ai Playground. E.g. spice.ai/lukekim/demo/playground.

Execute the query:

SELECT * from eth.recent_blocks LIMIT 10;

Note that it's very fast... subsecond.

Execute a more compute intensive query. A join across two large datasets:

SELECT *
FROM eth.recent_traces trace
JOIN eth.recent_transactions trans ON trace.transaction_hash = trans.hash
ORDER BY trans.block_number DESC

Note that while still fast it takes 5-7 seconds to complete because it's fetching ~130K rows over HTTP.

Step 2. Show the same query using the Spice Python SDK to simulate app or Notebook usage

Install the Spice CLI:

curl https://install.spiceai.org | /bin/bash

Open demo.py in VS Code or another editor.

Show that it's the same query as above, but this time we'll run it from the Python script simulating an app or Notebook.

python demo.py

Note that it takes about 2-3 seconds to complete. An improvement over the Playground because it's using Apache Arrow Flight not HTTP/JSON.

Step 2. Show getting started with Spice OSS and performance of local queries

# Initialize the Spice app and show the generated spicepod.yaml
spice init

# Set the name to "smart"

# Log in to the Spice.ai platform
spice login

# Start the Spice runtime
spice run

Note, the runtime starts HTTP, Arrow Flight, and OTEL endpoints.

Open another terminal window so the runtime can continue to run.

# Add the lukekim/demo spicepod
spice add lukekim/demo

Note the four loaded datasets in the runtime logs.

Open /spicepods/lukekim/demo/spicepod.yaml and show the dataset definitions.

# Start the Spice SQL REPL
spice sql

# Execute the show tables query
sql> show tables;
+---------------+--------------------+--------------------------+------------+
| table_catalog | table_schema       | table_name               | table_type |
+---------------+--------------------+--------------------------+------------+
| datafusion    | public             | eth_recent_transactions  | BASE TABLE |
| datafusion    | public             | eth_recent_traces        | BASE TABLE |
| datafusion    | public             | eth_recent_blocks_duckdb | BASE TABLE |
| datafusion    | public             | eth_recent_blocks        | BASE TABLE |
| datafusion    | information_schema | tables                   | VIEW       |
| datafusion    | information_schema | views                    | VIEW       |
| datafusion    | information_schema | columns                  | VIEW       |
| datafusion    | information_schema | df_settings              | VIEW       |
+---------------+--------------------+--------------------------+------------+

Query took: 0.006861292 seconds

Note the four datasets from the Spicepod are now loaded.

# Execute a query on the recent blocks table
sql> select * from eth_recent_blocks;

# Exit the REPL
sql> exit

Open demo.py again, comment out the first code block. Note line 28 client = Client(SPICEAI_API_KEY, 'grpc://127.0.0.1:50051') configuring the SDK to use the local Spice runtime.

Model Dseployment and Machine Learning Inference

Start telegraf to collect local disk data which will be used for ML inference.

telegraf --config telegraf.conf

Show the collected data using spice sql.

Run:

curl localhost:3000/v1/models/drive_stats/predict | jq

Explain what just happened in the prediction result and performance. prediction means the life percent the disk is at now.

Model Refinement and Lifecycle

In the Spice.ai playground spice.ai/lukekim/smart/playground, show there is no data in last 30 minutes:

SELECT *
FROM lukekim.smart.drive_stats
WHERE time_unix_nano / 1e9 > (UNIX_TIMESTAMP() - 1800) -- show data in last half hour
ORDER BY time_unix_nano DESC

Explain how to use mode: read_write to replicate data to the Spice.ai platform.

Go back to the Spice.ai Playground show data is now being replicated to the cloud.

SELECT *
FROM lukekim.smart.drive_stats
WHERE time_unix_nano / 1e9 > (UNIX_TIMESTAMP() - 1800) -- show data in last half hour
ORDER BY time_unix_nano DESC

Go to Models, edit on GitHub, to change the training query to use the view ** FROM lukekim.smart.drive_stats_with_local which will combine the cloud dataset with local replicated data.

Click Train to start a new training run with the combined data.

While the model is training, copy the training run ID of the previous model version and create a new model entry in the spicepod with the new version.

Edit spicepod yaml, duplicate the existing model entry to add a non-latest version.

datasets:
    from: spice.ai/lukekim/smart/models/drive_stats:latest
    name: drive_stats_v2

    from: spice.ai/lukekim/smart/models/drive_stats:60cb80a2-d59b-45c4-9b68-0946303bdcaf` # Previous training run ID
    name: drive_stats_v1

Restart the runtime to show both models are now loaded.

Execute a new inference on both models.

curl --request POST \
  --url http://localhost:3000/v1/predict \
  --header 'Content-Type: application/json' \
  --header 'User-Agent: insomnia/8.6.0' \
  --data '{
    "predictions": [
        {
            "model_name": "drive_stats_v1"
        },
        {
            "model_name": "drive_stats_v2"
        }
    ]
}' | jq -r '.predictions.[] | [.model_name, .status, .duration_ms, .prediction[0]] | @csv' | column -s, -t

Explain why this is valuable (versioning, A/B testing, etc.) and the model lifecycle.