Skip to content

[P1] Clean up the state #1223

Open @tcharchian opened 2026-05-21 22:34 UTC 2 comments Updated 2026-09-17 06:52 UTC
Priority: Medium

Review what’s currently stored, identify any leftovers, and remove them.


💬 Comments (2)

@Ryanchen911 commented 2026-05-28 02:09 UTC

hi @tcharchian ,I will take this one, thank you!

@Ryanchen911 commented 2026-09-17 06:52 UTC

Developer-stats scope is now resolved: the pre-devshard DeveloperStatsByTime, DeveloperStatsByEpoch, embedded InferenceStats, and supporting stats/... indexes do not need to be retained. They must be removed gradually rather than from an upgrade handler.

I opened #1793 for that cleanup. It sweeps all four prefixes (~10.4 GiB / ~34.7M keys in the measured snapshot) with a shared 1,000-key per-block budget, while limiting the large DeveloperStatsByEpoch aggregate values to one deletion per block. It also reclaims stale by-model entries that cannot be reached through the other indexes.

Current split and merge order: - #1773: generic pruner correctness fixes - #1499: gradual InferenceValidationDetails cleanup (~4 GiB), stacked on #1773 - #1793: gradual legacy developer-stats cleanup (~10.4 GiB), stacked on #1499 - #1287: independent v0.2.16 bounded legacy cleanup + state analysis tooling

Relevant keeper/types/upgrade tests pass locally.


🔄 Auto-synced from Issue #1223 every hour.