[agent_runs.py] Unbounded in-memory _RUNS dict with no size limit #682

New issue

Closed

opened 2026-06-02 23:43:47 +02:00 by sleepy · 1 comment

sleepy commented

2026-06-02 23:43:47 +02:00

Owner

agent_runs.py uses unbounded in-memory dict with no size limit

File: `src/agent_runs.py`

_RUNS: Dict[str, _Run] = {}

This dict holds run buffers indefinitely. While there is a grace-period eviction (3 minutes after last subscriber disconnects on a finished run), there is no overall size cap:

A burst of 1000 sessions each streaming large responses could consume gigabytes
If runs never get subscribers (started but never subscribed to), the eviction timer still fires — but if they stay "running" indefinitely (wedged background task), they never get evicted
No metric or logging of _RUNS size for monitoring

Additionally

_Run.buffer is a list that grows unbounded during a run — a very long agent loop with many rounds accumulates all SSE events
No max-buffer-size protection

Suggested fix

Add a MAX_RUNS cap (e.g. 100) with oldest-first eviction
Add periodic logging of _RUNS size
Cap _Run.buffer at a max number of events or bytes
Consider using WeakValueDictionary or explicit TTL-based eviction

## agent_runs.py uses unbounded in-memory dict with no size limit ### File: `src/agent_runs.py` ```python _RUNS: Dict[str, _Run] = {} ``` This dict holds run buffers indefinitely. While there is a grace-period eviction (3 minutes after last subscriber disconnects on a finished run), there is no overall size cap: - A burst of 1000 sessions each streaming large responses could consume gigabytes - If runs never get subscribers (started but never subscribed to), the eviction timer still fires — but if they stay "running" indefinitely (wedged background task), they never get evicted - No metric or logging of `_RUNS` size for monitoring ### Additionally - `_Run.buffer` is a `list` that grows unbounded during a run — a very long agent loop with many rounds accumulates all SSE events - No max-buffer-size protection ### Suggested fix - Add a `MAX_RUNS` cap (e.g. 100) with oldest-first eviction - Add periodic logging of `_RUNS` size - Cap `_Run.buffer` at a max number of events or bytes - Consider using `WeakValueDictionary` or explicit TTL-based eviction

sleepy added the

bug

enhancement

labels

2026-06-02 23:43:47 +02:00

sleepy referenced this issue

2026-06-03 00:10:52 +02:00

[chat/streaming] _active_streams dict has no eviction, TTL, or size limit #736

sleepy referenced this issue from a commit

2026-06-04 00:56:25 +02:00

fix(agent): add size limit to unbounded _RUNS dict

sleepy closed this issue