Question about st.cache_data TTL and Memory Cleanup

Using Streamlit
,
1

Hi Streamlit Team,

I have a question about how st.cache_data works with ttl, specifically around memory cleanup.

For example:

@st.cache_data(ttl=3600)
def load_data():
    ...

Scenario

  • At 10:00 AM, User 1 calls the function.

  • The function caches a DataFrame of about 1 GB.

  • The cache TTL is 1 hour.

No one calls this function again until 4:00 PM.

What I observed

While monitoring my application’s memory usage, I noticed that even though the TTL had expired several hours earlier, the process memory did not decrease. For example, after the cached DataFrame increased the application’s memory to around 1 GB, the memory usage stayed close to 1 GB instead of dropping after the TTL expired.

This made me wonder:

  • Is this the expected behavior?

  • Does Streamlit have any background process that automatically removes expired cache entries from memory, or are expired entries only cleaned up when the cached function is called again?

If a cache entry has expired, is it removed from Streamlit’s cache while the underlying Python memory is still retained? If so, is there a recommended way to release or reclaim that memory?

I looked into the available cache management options, including:

  • max_entries

  • st.cache_data.clear()

  • function_name.clear()

However, none of these seem ideal for my use case.

Using max_entries across many cached functions would increase cache evictions, leading to more cache misses and recomputation, which impacts page performance.

Similarly, st.cache_data.clear() or function_name.clear() removes all cached entries (globally or for a function), which isn’t practical in a production application where other users may still be relying on those cached results.

Is there a recommended approach for releasing memory used by expired cache entries without clearing the entire cache or aggressively limiting the cache size?

I also tested the behavior with a second user.

  • At 4:00 PM, User 2 called the same cached function with different input parameters.

  • The function created a new cached result of approximately 800 MB.

I expected the old cache entry to be released and the application’s memory usage to settle around 800 MB. Instead, the process memory increased from about 1 GB to approximately 1.8 GB, which suggests that the earlier cached data was still occupying memory while the new cache entry was added.

Is this also the expected behavior? If so, when is the memory used by an expired cache entry actually released?

I’d appreciate any clarification on how TTL expiration and memory management are intended to work internally.

Thank you!

2

Welcome to the community and thanks for your detailed question! :blush: What you’re seeing is expected behavior in Streamlit and Python. Expired cache entries (from @st.cache_data(ttl=...)) are only removed from memory when the cached function is called again, not automatically in the background. TTL expiration is “lazy”: the cache marks entries as expired, but actual cleanup (and memory release) only happens on the next access or when you explicitly clear the cache. There is no background process that proactively frees memory after TTL expiry—this is by design and matches Python’s memory management model. See Streamlit caching docs and architecture notes.

Even after an expired entry is removed from the cache, Python’s memory allocator may not immediately return memory to the OS; it keeps it reserved for reuse by the process. This is standard Python/OS behavior and not a Streamlit bug, as confirmed in Streamlit GitHub Issue #12120 and Issue #6354. The best way to ensure memory is reclaimed is to trigger cache cleanup by calling the cached function or using .clear(), but there is no built-in way to force memory release for only expired entries without affecting other cache contents.

Sources:

3
Behavior Expected? When memory is actually released
TTL expires but RSS stays high Yes Only on next access to that key, explicit clear, or max_entries eviction
New call with different args adds more memory Yes Old expired entry is not touched
Background cleanup of expired entries No There is none