Skip to content
Self-Driving DB LabCOMP90050 · G40

Showcase · workflows and screenshots

A guided tour of the lab

Three short recordings of the main workflows, each step captioned on screen, followed by screenshots of every feature. A scripted browser tour made them against this site, so anyone with the repository can re-record them with the same seeds (workload seed 2023, 10 replicates). Measured SQLite timings vary from run to run, so the numbers in a re-recording will differ slightly.

  • 3 walkthroughs
  • captions and transcripts
  • 17 screenshots, light, dark and mobile

Workflow walkthroughs

Watch the main journeys

The videos are silent. Each step is captioned on screen, the player also offers the captions as a text track, and the numbered steps beside each video are its transcript, with the time each step starts.

Walkthrough 1 · 1:07

Levels of autonomy

Scroll the survey's explainer from a manual database to a self-driving one, then filter the taxonomy of every system the report covered.

Open Overview Open Survey map

Steps and transcript

  1. 10:00The survey's argument as a scroll-driven explainer. Level 0: people choose every index.
  2. 20:05Levels 1 and 2: advisors such as AutoAdmin recommend, and a DBA still decides.
  3. 30:11Levels 3 to 5, then the predictor, tuner and organiser of a self-driving system.
  4. 40:25The survey map: Table 1's six levels of autonomy, then a taxonomy of every system covered.
  5. 50:32Filter by component: index selection only. Then by technique: multi-armed bandits.
  6. 60:46Only the systems implemented in this lab's arena, or a search by name. Clear to see all.

Walkthrough 2 · 1:07

Index advisor arena on the Louvre DB

Load the Louvre database from INFO20003, generate a museum workload, race no index against greedy what-if search and the bandit, then repeat it ten times for intervals.

Open Arena Open Benchmark

Steps and transcript

  1. 10:00The arena: index advisors tune a live SQLite database (WebAssembly) in your browser.
  2. 20:04Pick the Louvre database designed in INFO20003: 19 tables of synthetic museum activity.
  3. 30:08A museum workload from seed 2023, in three phases: box office, gallery floor, exhibitions.
  4. 40:18Three contenders: no index, AutoAdmin's greedy what-if search and the C²UCB bandit.
  5. 50:24One measured run: recommendation, index builds and queries, and cumulative time per round.
  6. 60:41One run is an anecdote. The benchmark replays 10 seeded workloads (seeds 2023 to 2032).
  7. 70:50Mean cumulative time with 95% bootstrap intervals, then paired differences against greedy.
  8. 80:57Change, effect size d_z, sign test and the share won; then the bandit's regret with its band.

Walkthrough 3 · 1:07

LLM as advisor (bring your own key)

Open the bring-your-own-key settings, ask the LLM index advisor for a proposal (a mocked reply here), let the validator and a person decide, then benchmark it and read the audit log.

Open Benchmark Open AI audit log

Steps and transcript

  1. 10:00Optional AI: open AI settings. Your own Anthropic or OpenAI key stays in this browser.
  2. 20:06Here a placeholder key is typed and every provider request is intercepted. No model is called.
  3. 30:12Ask the LLM index advisor. It is sent the schema, round 1 of the workload and SQLite's plans.
  4. 40:19A mocked reply for illustration, labelled AI-generated. The validator rejects invalid indexes.
  5. 50:28A person decides: untick one index and accept the edit. The decision is logged first.
  6. 60:35The approved indexes join the benchmark: same seeded workloads, paired against greedy.
  7. 70:46Compare on build + run time: the provider's response time says nothing about the advice.
  8. 80:59The audit log: call, validator verdict, decision and measurement. JSON and CSV export.

Screenshots

Every feature at a glance

Desktop shots are 1440 × 900. Select one to open a larger view; the arrow keys move between them.

On a phone (390 px)

How these were made

  • A Playwright script (pnpm showcase) drives Google Chrome through each journey at a human pace, adds the caption banner and a visible cursor, and records the screen at 1280 × 800. The same script is an end-to-end test: it checks what each journey should show (the Louvre dataset selected, every run finished, ten replicates, the validator's verdicts, the audit record) as it goes.
  • Seeds are the site's defaults, so a re-recording replays the same workloads. Measured SQLite timings are wall-clock times in your browser and differ between runs and machines; the published numbers on the methods page pool five independent sessions.
  • The recordings are trimmed and converted with ffmpeg (H.264 for this page, GIF for the README); the captions above come from the same step list as the on-screen banner.