菜单
Database backup and recovery strategy selection for production applications. Use when: designing backup schedules, implementing WAL or point-in-time recovery (PITR), choosing between full/differential/incremental backups, defining RTO and RPO targets, or recovering from data loss events.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/cardinalconseils/claude-starter --skill database-recovery命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | database-recovery |
| description | Database backup and recovery strategy selection for production applications. Use when: designing backup schedules, implementing WAL or point-in-time recovery (PITR), choosing between full/differential/incremental backups, defining RTO and RPO targets, or recovering from data loss events. |
| allowed-tools | Read, Grep, Glob, Bash |
Every production database needs a backup strategy before it has users. The right strategy depends on one question: what is the business cost of losing one hour of data? One day? One week? Match the strategy to the answer.
Recovery Point Objective (RPO) — How much data loss is acceptable? "RPO = 1 hour" means losing up to one hour of writes is tolerable.
Recovery Time Objective (RTO) — How long can the system be unavailable during recovery? "RTO = 30 minutes" means the service must be restored within 30 minutes of a failure.
Define both before choosing a strategy. If you cannot answer these, ask the business stakeholders — they own the risk.
The entire database is dumped to object storage once per week. The dump is replicated to a second geographic region.
pg_dump commandA full "base" backup is taken once per day. Every hour, a differential backup captures only the changes since the last base backup. Recovery combines the latest base with the most recent differential.
Every write to the database is first written to a sequential log (the Write-Ahead Log). Only after the log entry is persisted does the actual database write occur. Periodic base backups provide a starting point; the WAL provides the path from there to any point in time.
| Level | Backup Frequency | RPO (Data Loss) | RTO | Storage Impact | Complexity |
|---|---|---|---|---|---|
| Noob | Weekly full | Up to 1 week | Hours | High | Low |
| Pro | Daily base + hourly diff | Up to 1 hour | Medium | Moderate | Medium |
| Hacker | Continuous WAL + periodic base | Near zero (minutes) | Variable | Optimal | High |
WAL works by enforcing log-first writes:
On failure, recovery loads the last base backup and replays WAL entries in sequence up to the desired restore point. This enables restoring to any timestamp between the last base backup and the failure.
Managed WAL in practice:
pg_basebackup + WAL archiving to S3/GCS with archive_command| Context | Strategy |
|---|---|
| Prototype, internal tool, no real users | Noob (weekly full) |
| Real users, no financial data | Pro (daily base + hourly diff) |
| Payments, user-generated content, regulated data | Hacker (WAL / PITR) |
| AI billing or metered usage | Hacker — usage data loss = revenue loss |
Untested backups are not backups. A backup that has never been restored is an assumption.
| Rationalization | Reality |
|---|---|
| "We'll add proper backups before launch" | Data loss happens before launches too. Implement the strategy before first real user. |
| "We have automated backups enabled" | Automated backups exist. Have you ever restored from one? Test it. |
| "WAL is too complex for our stage" | Supabase and RDS enable WAL/PITR with a checkbox. Complexity argument no longer holds. |
| "We can recover from app logs if needed" | App logs are not a recovery mechanism. They are missing schema, foreign key state, and transaction ordering. |
| "Weekly backups are fine, we don't change much" | The week you change the least is often the week you lose the most. |