Streamlit Caching: @st.cache_data vs @st.cache_resource

The problem: full script reruns on every interaction

When a user moves a slider, clicks a button, or changes any widget, Streamlit reruns the entire script from top to bottom. If the script loads a large CSV, trains a model, or opens a database connection on each run, the app feels sluggish.

Choosing the right caching decorator

• @st.cache_data – for pure functions that return data (e.g., filtering, aggregation). The output depends only on the function's arguments.
• @st.cache_resource – for heavy objects that should be created once, such as ML models, database clients, or API sessions.

Worked example: CSV filter with a cached model

import streamlit as st
import pandas as pd
# Dummy model placeholder – replace with your actual model loader
def load_model():
    # Simulate an expensive load
    return {'weights': [0.1, 0.2, 0.3]}

@st.cache_data
def load_data(path: str) -> pd.DataFrame:
    return pd.read_csv(path)

@st.cache_data
def filter_data(df: pd.DataFrame, threshold: float) -> pd.DataFrame:
    return df[df['value'] > threshold]

@st.cache_resource
def get_model():
    return load_model()

def main():
    st.title('Cached Streamlit Demo')
    # File path – adjust to your environment
    data = load_data('sample.csv')
    threshold = st.slider('Value threshold', 0.0, 10.0, 5.0)
    filtered = filter_data(data, threshold)
    model = get_model()
    st.write('Rows after filter:', len(filtered))
    st.write('Model weights:', model['weights'])
    if st.button('Clear cache'):
        st.cache_data.clear()
        st.cache_resource.clear()

if __name__ == '__main__':
    main()

Save the script as app.py and run streamlit run app.py. The first execution loads the CSV and creates the model; subsequent slider moves reuse the cached data and model, so the UI updates instantly.

Trade‑off and practical check

• Cached objects occupy memory for the lifetime of the Streamlit session. Monitor memory with your OS task manager or psutil to ensure the cache does not exceed available RAM.
• If the underlying CSV changes, the cached DataFrame becomes stale. To avoid serving outdated results, either add a version string to the cache key (e.g., @st.cache_data(show_spinner=False) with a custom hash) or manually clear the cache when you know the source updated.
• You can verify the cache is working by adding a temporary st.write('Loading data…') inside load_data; the message appears only on the first run and after you click Clear cache.

Start by annotating the slowest function with @st.cache_data, measure the difference with st.time (or a simple time.time() wrapper), then promote expensive objects to @st.cache_resource. Finally, expose a Clear cache button as shown to let users reset when needed.