Streamlit in Snowflake: Advanced Caching

Happy Valentine’s Day! Is there anything more romantic than a dashboard that loads instantly? I didn’t think so.

Streamlit in Snowflake (SiS) has revolutionized how we build internal data apps. But a common complaint is performance. “It’s slow when I change the filter.” This is almost always a caching issue.

Understanding the Execution Model#

Streamlit runs the entire python script from top to bottom every time a user interacts with a widget. If you have a query session.sql("SELECT * FROM BILLION_ROW_TABLE").collect() at the top of your script, you are re-scanning that table every time someone clicks a button.

The Solution: `st.cache_data`#

In 2025, st.cache_data is the standard decorator for data fetching.

import streamlit as st
from snowflake.snowpark.context import get_active_session

session = get_active_session()

@st.cache_data(ttl=3600)  # Cache for 1 hour
def get_sales_data(region):
    # This string generation happens fast
    query = f"SELECT * FROM sales WHERE region = '{region}'"
    # This expensive network/compute call happens once per hour per region
    return session.sql(query).to_pandas()

selected_region = st.selectbox("Region", ["NA", "EMEA", "APAC"])
df = get_sales_data(selected_region)
st.dataframe(df)

python

Key Parameters#

ttl: Time To Live. Set this! Default is forever (until app restart).
show_spinner: Toggle the “Running…” spinner.

Caching vs. Session State#

Caching is global (shared across users in some contexts, or persistent across re-runs). Session State is per-user-session.

Use st.session_state to store user inputs or intermediate calculation results that don’t need to be re-computed but are specific to this user’s workflow.

if 'counter' not in st.session_state:
    st.session_state.counter = 0

if st.button('Count'):
    st.session_state.counter += 1

python

Large Dataframes: Mutability Warning#

Remember that st.cache_data stores a copy of the data. If your function returns a massive DataFrame, Streamlit has to pickle/unpickle it.

For massive datasets, consider pushing the aggregation down to Snowflake SQL and only caching the results. Don’t pull raw data into the Streamlit app unless necessary.

Conclusion#

A well-cached Streamlit app feels like a native compiled application. A poorly cached one feels like a sluggish web page from 1999. Use st.cache_data aggressively for all SQL interaction.