Skip to content

Grafana: fold to query result, decision and cost; label decision rows; prerequisite check in the one-click script - #258

Merged
GordonYuanyc merged 4 commits into
mainfrom
grafana-drop-disagreement-marks
Nov 17, 2026
Merged

GordonYuanyc merged 4 commits into
mainfrom
grafana-drop-disagreement-marks

Conversation

@GordonYuanyc

@GordonYuanyc GordonYuanyc commented Nov 16, 2026 •

Copy link
Copy Markdown
Collaborator

Cleans up the use case dashboards after review questions, and makes the one-click script check its prerequisites up front.

Dashboards

  • Removed the vertical approx ≠ exact marks; shaded problem periods that cover the whole run are skipped.
  • Only the rows with the query result (exact vs approx, error), the application decision and query cost start expanded; every other row starts collapsed. Use case 3 has no query result row, so its rule rows stay open.
  • State strips show a legend of their colors; each dashboard has a line saying the light blue shading marks the injected problem periods.
  • Use case 4's "query result and decision" panel labels its two rows of dots "exact decisions" and "approx decisions" on the right axis, with a key line above it.
  • The scenario figures are embedded in the dashboards instead of fetched by the browser from localhost:8001, so they show when Grafana runs on a remote machine with only port 3000 forwarded; run.sh no longer starts the figure server.
  • Removed use case 4's canary p99 / stable p99 ratio panel (it always showed rule ours, whatever rule was selected), its Fidelity over rollouts row, and the Stage 1 / Stage 2 rows of every use case. The runners still write that data; the dashboards no longer read it.

One-click script

Checked

  • A fresh public clone of main, with an empty cargo cache and a bare Python venv, ran run_all_use_cases.sh --no-grafana once cmake was on PATH: 6.5 min build, 92 min for all five use cases, exit 0, 64 runs. The only unsupported query is use case 1's entropy_10s in the error-budgeted arms (no accuracy guarantee).
  • The regenerated dashboards render on a separate Grafana for both that fresh data and the existing data; the use case unit tests (115) pass.

🤖 Generated with Claude Code

GordonYuanyc and others added 4 commits November 16, 2026 15:32
…shaded areas

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… Grafana before the build

ASAPPlanner builds the HiGHS solver from source, so a machine without cmake
failed four minutes into the build. Every missing tool is now listed at once
before anything starts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… rows; drop uc4 ratio, fidelity and stage rows

- Rows other than Query result, Application and Query cost (uc3: its rule rows
  and Query cost) start collapsed.
- State strips show a legend of their states; each dashboard says what the blue
  shading is.
- The uc4 decision dots carry exact/approx labels on the right axis, with a key
  line above the panel.
- Removed: uc4's canary-over-stable ratio panel (it showed rule ours whatever
  rule was picked), Fidelity over rollouts, and the Stage 1 / Stage 2 rows of
  every use case.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er serves them on port 8001

The browser fetched each figure from localhost:8001, so it broke when Grafana
ran on a remote machine with only port 3000 forwarded, or when 8001 was taken.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@GordonYuanyc
GordonYuanyc merged commit f1dc204 into main Nov 17, 2026
2 checks passed
@GordonYuanyc
GordonYuanyc deleted the grafana-drop-disagreement-marks branch November 17, 2026 04:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant