designing-real-world-ai-agents-workshop: Skill for Claude Code

Streamlit performance

Performance is the biggest win. Without caching and fragments, your app reruns everything on every interaction.

Caching

@st.cache_data for data

Use for any function that loads or computes data.

# BAD: Recomputes on every rerun
def load_data(path):
    return pd.read_csv(path)

# GOOD: Cached
@st.cache_data
def load_data(path):
    return pd.read_csv(path)

@st.cache_resource for connections

Use for connections, API clients, ML models—objects that can't be serialized.

@st.cache_resource
def get_connection():
    return st.connection("snowflake")

@st.cache_resource
def load_model():
    return torch.load("model.pt")

Critical warning: Never mutate @st.cache_resource returns—changes affect all users:

# BAD: Mutating shared resource
@st.cache_resource
def get_config():
    return {"setting": "default"}

config = get_config()
config["setting"] = "custom"  # Affects ALL users!

# GOOD: Copy before modifying
config = get_config().copy()
config["setting"] = "custom"

Cleanup with on_release: Clean up resources when evicted from cache:

def cleanup_connection(conn):
    conn.close()

@st.cache_resource(on_release=cleanup_connection)
def get_database():
    return create_connection()

TTL for fresh data

@st.cache_data(ttl="5m")  # 5 minutes
def get_metrics():
    return api.fetch()

@st.cache_data(ttl="1h")  # 1 hour
def load_reference_data():
    return pd.read_csv("large_reference.csv")

Guidelines:

• Real-time dashboards → ttl="1m" or less
• Metrics/reports → ttl="5m" to ttl="15m"
• Reference data → ttl="1h" or more
• Static data → No TTL

Prevent unbounded cache growth

Important: Caches without ttl or max_entries can grow indefinitely and cause memory issues. For any cached function that stores changing objects (user-specific data, parameterized queries), set limits:

# BAD: Unbounded cache - memory will grow indefinitely
@st.cache_data
def get_user_data(user_id):
    return fetch_user(user_id)

# GOOD: Bounded cache with TTL
@st.cache_data(ttl="1h")
def get_user_data(user_id):
    return fetch_user(user_id)

# GOOD: Bounded cache with max entries
@st.cache_data(max_entries=100)
def get_user_data(user_id):
    return fetch_user(user_id)