Find the moment
Find exactly when rows were deleted or changed, or a table was emptied or dropped.
Customers are gone, orders all say "closed", a table is empty. Before you rewind, you need the moment. Find the moment lists each big delete, update, TRUNCATE and DROP, with its exact second, table, row count and transaction ID.
It only reads. The agent reads the change log in your storage, on your server. Only names, times and counts reach Rowsafe.
In the dashboard
- Open the database, then Rewind → Find the moment.
- Pick the tables (or All tables), how far back (an hour up to 7 days, or a range), and what happened: deleted rows, changed rows, emptied tables, dropped tables.
- Click Find the moment. A day of changes usually takes under a minute.
You get totals, a timeline (click a bar to filter) and the changes, newest first. Each transaction is one row. A table changed the same way again and again (an hourly cleanup, say) is folded into one row, so the unusual change stands out.
Rewind to just before this opens Rewind at the moment just before that transaction committed. From there, restore a copy, compare, and bring the rows back.
Any member can search. Rewinding needs the admin or owner role.
From a terminal
rowsafe rewind find app --table customers --since 6hSearched 2026-09-24 08:00:00 CEST (06:00:00 UTC) to 2026-09-24 14:00:00 CEST (12:00:00 UTC).
WHEN (YOUR TIME ZONE) WHAT TRANSACTION
2026-09-24 11:47:03 CEST 1,204 rows deleted from customers 8841207
Found 1 transaction with matching changes between 06:00 and 12:00 UTC on Sep 24. The biggest: 1,204 rows deleted from customers at 09:47:03 UTC.
Rewind to just before the biggest one (a copy next to production; it never touches it):
rowsafe rewind copy app --at 2026-09-24T09:47:03.521880Z| Option | What it does |
|---|---|
--table T | Only this table: schema.table or table. Repeat or comma-separate for several. |
--since 24h | How far back: 30m, 6h, 2d, 1w (default 24 hours, at most 7 days). |
--from TIME, --to TIME | An exact range instead, in your time zone ("2026-09-24 14:00", "14:00", "2h ago") or RFC 3339. |
--kind K | delete, update, truncate, drop (comma-separated). |
--db DB | Only this PostgreSQL database, when the server has several. |
--min-rows N | Leave out transactions that changed fewer rows. |
--json | The full result as JSON. |
For AI agents
The MCP tool find_moment does the same read-only search, for every key, read-only ones too. An assistant can answer "when were rows deleted from applications?" with the exact point. It rewinds only if you ask, as the person who connected it. See Guard.
By database
| Reads | Differences | |
|---|---|---|
| PostgreSQL | the WAL in your bucket | to the microsecond, with transaction IDs |
| MySQL, MariaDB | the binary logs in your bucket, with the server's own mysqlbinlog | to the second; with binlog_format = STATEMENT, only TRUNCATEs and DROPs are found |
| MongoDB | the oplog in your bucket | to the second; outside a multi-document transaction, deletes and updates are grouped by collection and second |
| Redis, Valkey | the stream of changes in your bucket; you type keys or patterns (user:*), not tables | to the second; writes grouped by logical database, pattern and second, a MULTI/EXEC as one; results name only the patterns you typed, never the keys they matched; expired keys show as deleted |
| ClickHouse | the record of the parts ClickHouse wrote, in your bucket | to the microsecond; a mutation is one change at its first part; UPDATE rows are those of the parts it rewrote ("up to"); TRUNCATE and DROP PARTITION show as emptied tables |
| SQLite | the WAL pages in your bucket, replayed on a private copy on your server | to the millisecond Rowsafe copied each change; a DELETE without WHERE shows as every row deleted (SQLite has no TRUNCATE); needs WAL mode (details) |
Row contents never leave your server. To the second means the rewind goes back to the second before the change, so earlier changes in that same second aren't in the copy. The notes below are about PostgreSQL.
What it shows, and what it can't
- Times are commit times, to the microsecond. Rewinding to just before leaves the change out and keeps everything before.
- Rows are counted exactly for deletes and updates. Emptied or dropped tables show PostgreSQL's last size estimate, marked "about".
- Only committed changes. Savepoint changes count with their transaction. Prepared (two-phase) transactions aren't listed.
- Not who did it. The change log doesn't record the user or app. Check your app logs or
log_statement. - Names of tables that are gone. The agent notes table names every 15 minutes and keeps them 8 days. A table created and dropped within minutes may show as "a table Rowsafe couldn't name".
- The newest minute. If your range ends now, the agent first uploads the latest changes, waiting up to a minute.
- Limits. One search: at most 7 days, the 100 biggest transactions, about 8 GB of change log (past that, the most recent part, and it says so). It can't look back past your oldest kept backup.
VACUUM FULL,CLUSTERand table rewrites aren't reported: they don't lose rows. - ClickHouse. Reads the parts record Rowsafe keeps for restores to any second: quick, no extra disk.
DELETE(lightweight orALTER ... DELETE) is counted exactly.UPDATEshows the rows of the parts it rewrote ("up to").TRUNCATEandDROP PARTITIONshow as an emptied table. Times are when the first part was written. - Cost on the server. Low CPU and IO priority. The change log downloads in 16 MB pieces, deleted once read. Needs about 1 GB free under
/var/lib/rowsafe.
SQLite
SQLite's change log holds pages, not statements, so the search works differently:
- It replays, then compares. The agent restores the file as of the start of your range into a private scratch file, then replays one transaction at a time, comparing changed pages by primary key (or row number). Gone rows were deleted, changed rows updated, a missing table dropped. Values are compared as hashes on your server, and the scratch file is deleted after.
- Times are when Rowsafe copied each change, within a second of commit, to the millisecond. Rewind to just before this goes back one millisecond; if other changes were copied in that same millisecond, the result says so, since the copy leaves them out too.
- What it lists. Deleted rows, changed rows and dropped tables, counted exactly. A
DELETEwithoutWHEREshows as every row deleted. Added rows aren't listed.VACUUMisn't listed (rows don't change); a table rebuilt by a migration is, with a note. A transaction that rewrote more than about 512 MB (like aVACUUMof a large file) is listed without counts. - Needs WAL mode. Rollback-journal mode keeps no record of changes: Pulse offers to turn WAL on.
- Cost on the server. About the file's size free under
/var/lib/rowsafe. It stops after 10 minutes of replay: split a busy week into shorter ranges. The result says how far it got. - One database per file, so the database filter doesn't apply.