tree 474dd9eca841432ce8fe47b7a038d2ded85f114a
parent e58f8dd9577523f636704a96b907137f36fd4b5b
author Xinhao Yuan <xinhaoyuan@google.com> 1773149905 -0700
committer Copybara-Service <copybara-worker@google.com> 1773149933 -0700

Use reservoir sampling in auto-dictionary tracing and collection.

This is to avoid biases that favors later entries that always overwrite the earilers.

As side-effects, auto-dictionary tracing tables are filled randomly only after they become full of entries, and it saves the memory copying when it randomly decides to skip an entry.
Also the table cleanup is now lazy, and it cleans only the previously used entries. It should be much faster than the previous method of cleaning all the entries on small tests.

PiperOrigin-RevId: 881397367