Skip to content

Re-run the baseline drift/step scan at 60-90 days of store history (early September) #2019

Description

@erikdarlingdata

The dated follow-up from #1743's phase-2 verdict. The drift/step measurement on 2026-08-03 ran against ~16 days of baseline history and concluded alert machinery wasn't warranted yet: batch drift median 0.116σ/day with 47/52 servers in a razor band (secular fleet growth, not per-server anomalies), and steps nearly absent (1 of 104 server-metric series >2σ). The honest disposition was to re-measure once the store carries enough history for step detection to have statistical teeth.

When: early September 2026 (60–90 days of baseline-relevant history).

How: the residual-analysis scripts are preserved in the session scratchpad lineage as phase2_r1.sql / phase2_r2.sql (residual = sample − hour×dow bucket median; drift = per-server σ/day regression on residuals; steps = level-shift scan per server-metric series). Re-run both against the production store via the SSM+psql pattern, compare against the August baseline numbers, and decide between: (a) changepoint-triggered baseline-window truncation, (b) a cohort-relative drift weekly digest, or (c) park permanently.

Full measurement detail: #1743 (comment)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions