Recovery selection, drills, and acceptance¶
A successful backup does not establish recoverability. Define the failure, acceptable data loss (RPO), and recovery time (RTO), select a method, and drill on an independent target.
Choose by failure¶
Scenario |
Candidate |
Boundary |
|---|---|---|
Mistaken update/delete with intact storage |
SQL Snapshot/PITR |
Depends on retained objects/history in the same system; not an independent DR copy |
Lost/unavailable object storage or site |
Verified backup on independent media |
Must not depend on the failed original bucket/prefix; target/tool/format compatibility required |
Migration/selective restoration |
Logical export/import |
Verify objects, privileges, types, encoding, sequences, external dependencies; file exports are not full-instance backups |
Physical full/incremental restoration |
Compatible tool and complete chain |
Retain base, ordered increments, checksums/keys, and tool version; follow actual CLI |
SQL Snapshot/PITR, object-store version history, application backups, and mo_br physical backups differ. Seeing the database in SHOW DATABASES verifies its name, not restored data, privileges, or business behavior.
Independent-target drill¶
Record failure time/timezone, source version/configuration, backup point/chain, checksums, key references, and scope.
Prepare an independent namespace/volume/bucket prefix. Stop concurrent target writes, check tool version and
--help, and follow its reference.Restore on the target and preserve errors; clearing production directories or shared buckets is not a diagnosis step.
Check definitions, counts, key ranges, critical aggregates/watermarks, samples, application privileges, and external references. For CDC/Tasks, assess replay, duplicates, and recovery point.
Validate business behavior from actual clients before switching. Measure lost/duplicate data and duration; planned backup frequency is not measured RPO.
Isolated SQL snapshot experiment¶
Run only on a test instance with unused names. This experiment restores its own temporary database to an earlier state. It verifies recovery from a mistaken write, not object-storage loss or physical-backup DR.
CREATE DATABASE mo_recovery_lab;
USE mo_recovery_lab;
CREATE TABLE ledger(id INT PRIMARY KEY, amount INT);
INSERT INTO ledger VALUES(1,42);
CREATE SNAPSHOT mo_recovery_lab_snap FOR DATABASE mo_recovery_lab;
UPDATE ledger SET amount=99 WHERE id=1;
-- Expected-Value: 99
SELECT amount FROM ledger WHERE id=1;
RESTORE DATABASE mo_recovery_lab {SNAPSHOT='mo_recovery_lab_snap'};
-- Expected-Value: 42
SELECT amount FROM mo_recovery_lab.ledger WHERE id=1;
DROP SNAPSHOT mo_recovery_lab_snap;
DROP DATABASE mo_recovery_lab;
The v4.2.5 probe restored using {SNAPSHOT='...'}. RESTORE DATABASE ... FROM SNAPSHOT ... returns 1064 in that version; do not copy another version’s compatibility syntax blindly. Check creation with SHOW SNAPSHOTS before modifying data.
Physical backups and retention¶
Use the mo_br reference and tool release guidance for availability, versions, commands, and compatibility. Historical tool names/logs do not establish current availability. This review did not obtain and execute the complete physical full/incremental toolchain and does not claim physical-backup verification.
Clean only expired backups managed by the tool after checking dependencies from retained bases/increments/snapshots. Do not manually delete shared objects, Raft/WAL, checkpoint files, or snapshot-referenced files. Isolate recovery targets from production and ensure backups do not disappear with primary storage.
Acceptance record¶
Save the recovery point/checks, target version, steps/errors, business validation, measured loss interval/duration, cutover/backout plan, owner, and follow-ups. Explicitly record unverified areas such as multi-CN restoration, privileges, external references, and cross-version recovery.