Two owners-only pages that answer "can we trust this?": the Audit log (/admin/audit) says who did what and when; Data health (/admin/data) says whether the stored data holds together and is up to date.
Audit log (/admin/audit)
The Audit log is one timeline of the security-relevant things the platform already stores, newest first, 50 events a page. It never shows a key or a token. Nothing is copied into it: each entry is read from where the platform keeps it (sign-in events, the catalog's change log, projects, signatures and share purchases).
What the audit log records
Each event belongs to one kind, which is also the Kind filter:
| Kind | Events |
|---|---|
| Sign-ins and membership | Signed in, Sign-in refused, Signed out, GitHub membership ended, Role changed on GitHub, GitHub membership could not be confirmed |
| Sessions | Session ended, Sessions ended (an owner ended them), Session expired |
| Groups and keys | Group added with its key, Group key replaced, Group key checked, Group removed with its key |
| Games | Game registered, Game state changed, Game removed, Game API key created, Game API key rotated, Game API key revoked |
| Projects | Project registered, with the project's name |
| Contracts | Contract submitted for signatures, Contract signed, Contract approved |
| Shares | Share bought |
A contract changed after it was submitted is not its own entry: the project keeps only its latest submission, and signatures cleared by a change are deleted. Catalog entries for games say what happened and who did it, but do not name the game.
The audit log's columns
- When
How long ago, in your browser's time; hover for the exact date and time.
- Event
The sentence, with its kind as a badge.
- Person
Who did it, or whose access it was, linked to their member page. "Not recorded" for the system or a deleted account.
- Detail
Where it came from or why, such as an address and browser, or a refusal's reason. Hidden on a phone (shown from 768 pixels wide).
How to filter the audit log
Pick a person
Person: anyone in the log, or Anyone.
Pick a kind
Kind: one of the kinds above, or Every kind.
Pick the days
From and To are days, inclusive, in UTC (the times in the table are your browser's).
Apply
Choose Apply. Clear removes them. A filter that does not make sense in the address is dropped, not an error.
For example, to see who signed in last week, pick the kind Sign-ins and membership and set From and To to that week.
Under the table, "Events 1 to 50" with Previous and Next pages through older events. A member's page on Members & access links here with the person filter already set.
How long audit entries are kept
Sign-ins, sessions and group-key changes are kept for the access log's retention, 180 days by default (AUTH_AUDIT_RETENTION_DAYS); the API removes older entries once an hour. Projects, contracts and shares are kept for as long as their rows exist.
Data health (/admin/data)
The Data health page says whether the stored data can be trusted: the database's invariant checks from their latest daily run, how fresh each source is, and what the dead-game cleanup last found. It only reads; nothing on it changes data.
Data health tiles
- Checks failingchecks
The accent tile: checks that found broken rows or did not run, "of" all checks.
- Last checked
When the checks last ran, or Never. The worker runs them once a day.
- Stale sourcessources
Sources whose newest data is older than their schedule allows, "of" all sources.
- Cleanup would deletegames
How many games the dead-game cleanup's last run counted, or "Not known" until it has run since the worker started.
Source freshness on Data health
The Freshness card has one row per source: Source, Last stored (how long ago, or Never) and State, Fresh or Stale, with "Within" the age allowed and, for a game, why its read is not working.
| Source | Stale when its newest data is older than | Its schedule |
|---|---|---|
| Roblox snapshots | 15 minutes | The hottest games are read every 5 minutes |
| GitHub organization | 1 hour | Collected every 30 minutes |
| Analytics: each game | 36 hours | Each studio game's owned analytics, once a day |
The limits leave room for one slow or missed run, so a single late run is not an alarm. Roblox snapshots are counted from the newest snapshot of the last week; a game whose analytics never read is stale with Never. What the numbers mean elsewhere: Freshness and unknowns.
Invariant checks on Data health
The Invariant checks card lists every rule the database must hold (for example, no row pointing at something that does not exist), with its latest result, the ones that did not run or found broken rows first. Columns:
- Check
Its name, the sentence saying what a counted row is, and why it did not run if it failed.
- Result
Holds (no broken rows), Broken (rows a rule says cannot exist), Watched (counted and reported, not a fault) or Did not run.
- Rowsrows
How many rows it counted, or a dash.
- Fixed by
The card that brings a known fault to zero, while one is open. Hidden below 1,024 pixels.
The checks count rows; they never change them. A developer runs the same checks against any database with pnpm db:check.
The dead-game cleanup on Data health
The Dead-game cleanup card reports the collector's daily cleanup: games dead for at least a week that no one at the studio owns or watches. A game is dead when Roblox leaves it out of three readings in a row; the collector then reads it once a week, and one more miss, with its last reading over seven days old, makes it eligible (see Collector page). Studio games and watched games are never eligible.
- Mode
Dry run: counts only (the default: nothing is deleted) or Deletes.
- Ran
How long ago.
- Dead for a week or moregames
Games eligible.
- Keptgames
Dead games kept because they are studio games or someone watches them.
- Would delete this rungames
Or Deleted this run: games, up to 1,000 a run.
- Rows per tablerows
A table of the rows per database table those games hold.
The report lives in the worker's memory, so it is empty ("No report yet") until the cleanup has run since the worker last started. Collector opens the Collector page.
Audit log and Data health on a phone
On the Audit log, the four filters stack full width and the Detail column is hidden. On Data health, the tiles sit two to a row, the cards stack one under another, and the Fixed by column is hidden.
Related: Admin pages, Collector and data freshness.
Updated