{{ message }}
Repository navigation
Should history manager cache results? #15388
Description
JupyterLab has inline completion which searches the history based on what user types. Since user usually types extra characters, it could be faster if instead of redoing the computation we filtered within previous results. In ipykernel this is exposed in https://github.com/ipython/ipykernel/blob/213cbf668219fe4e62afc97ee875bb9ead893899/ipykernel/ipkernel.py#L626-L641
Same optimization might be useful for IPython history completions too.
Activity
krassowski commentedon Sep 9, 2026
It looks like caching results is not useful in itself - see LLM benchmarking report in https://gist.github.com/krassowski/e8a11492e760d54f72b419e860f6b656.
Adding a covering index isntead might give us a significant boost (worst case scenario goes down from 183 ms to 37 ms at 1M history rows) at a cost of 50% increase disk usage.
Make inline history provider debounce (fix flaky tests), use ipykerne…