MySQL Migration Safety Review

Quick Reference

§1 Scope

Verified — 5.7, 8.0, 8.4 (transcribed from each manual, 2026-08-06). Assumed — 8.1–8.3, 9.x: 8.4's rules applied but never confirmed there; 9.x shares 8.4's online-DDL matrix byte for byte, yet 9.1.0 raised the INSTANT row-version ceiling 64→255 without touching it, so matrix identity is not rule identity — re-check numeric thresholds against the target's manual. Unverified — 5.6 and older, past 9.x: say so in §9.9 rather than answering as if covered. MM028 reports both; --fail-on warning makes either a hard stop.

In scope — schema migration safety for InnoDB:

• ALTER TABLE (add/drop/modify column, add/drop index, rename, convert charset); CREATE / DROP INDEX
• Data backfill and transformation migrations
• Table restructuring (partitioning, splitting, merging), foreign key changes
• Migration files as SQL: raw .sql/.ddl, Flyway, golang-migrate. Liquibase only via its SQL changelogs or liquibase updateSQL output — XML/YAML/JSON changelogs and programmatic (Go/Java/Python) migrations are not parsed; review the generated SQL and say that is what you did
• Rollback planning, verification, and replication impact assessment

Out of scope — delegate to dedicated skills:

• Query optimization, connection pooling, buffer tuning → mysql-best-practise
• Application code changes → go-code-reviewer or language-specific reviewer
• Security hardening, privilege management → security-review

§2 Mandatory Gates

Execute gates sequentially. Each gate has a STOP condition — if triggered, do not proceed until resolved.

Gate 1: Context Collection

Collect before giving migration advice:

If database access is available, run:

SELECT VERSION();
SHOW TABLE STATUS WHERE Name = '<table>' \G
SHOW CREATE TABLE <table> \G

STOP: Cannot determine whether the target is MySQL at all (e.g., migration file has no MySQL-identifiable syntax). Redirect to appropriate skill.

PROCEED: At least MySQL version and table name are known or conservatively assumed. Record all assumptions.

Gate 2: Scope Classification

Classify the request mode:

STOP: Request is not migration-related (e.g., query optimization, tuning). Redirect to mysql-best-practise.

PROCEED: Migration intent confirmed. Continue with depth selection.

Gate 3: Risk Classification

Risk is blast radius × duration × reversibility, not row count alone: a 1M-row table at 20k QPS is more dangerous than a 50M-row archive nobody writes to. Score each DDL statement across five axes and take the highest band any axis reaches.

Row count enters via axis B only, as a proxy for rebuild duration — prefer table size in bytes or a timed run on a replica. Any axis UNSAFE → UNSAFE (no averaging); two axes WARN → UNSAFE unless you name the one you mitigated; axis C or D unknown counts as WARN, not SAFE — an unmeasured hot table is not a cold one.

STOP: Any UNSAFE item has no mitigation plan. Must provide tool-based alternative or phased approach before proceeding.

PROCEED: Every DDL statement has a per-axis score, an overall band, and a corresponding mitigation.

Gate 4: Output Completeness

Before delivering output, verify all §9 Output Contract sections are present. If any section is missing, add it (even if "N/A — [reason]"). §9.9 Uncovered Risks must never be empty.

§3 Depth Selection

Force Standard or higher when any signal appears: column type change, NOT NULL addition, PK modification, FK add/remove, charset change, data backfill, partition change, column rename/removal.

§4 Degradation Modes

When context is incomplete, degrade gracefully — never fabricate information.

Hard rule: Never claim "SAFE" without evidence. In Degraded/Minimal mode, mark items as "SAFE (assumed — verify against production)" and list all assumptions in §9.9 Uncovered Risks.

§5 DDL Safety Checklist

Execute every item for each DDL statement. Mark SAFE / WARN / UNSAFE with evidence.

5.1 Algorithm & Lock Assessment

1. Algorithm selection — determine ALGORITHM=INSTANT, ALGORITHM=INPLACE, or ALGORITHM=COPY for each ALTER TABLE against the exact server version, then state it explicitly. Never rely on server default. Load references/ddl-algorithm-matrix.md whenever the operation is not a plain ADD COLUMN or ADD INDEX.

   Gates that are wrong more often than not — the ALGORITHM=INSTANT clause itself does not exist before 8.0.12, so 5.7 and 8.0.0–8.0.11 reject it for every operation, SET DEFAULT included; within 8.0.12+, positional ADD COLUMN needs 8.0.29+, RENAME COLUMN 8.0.28+, DROP COLUMN 8.0.29+; DROP COLUMN is INPLACE+LOCK=NONE on 5.7 and pre-8.0.29, never COPY; extending VARCHAR is never INSTANT and is COPY across the 255/256-byte boundary; ADD FOREIGN KEY is INPLACE only while foreign_key_checks=0, else COPY. Exception — partition clauses: ADD/DROP/REORGANIZE/COALESCE/REBUILD PARTITION accept only ALGORITHM=DEFAULT, LOCK=DEFAULT on 5.7, where naming INPLACE makes the statement fail. Use matrix §4 for those.

2. Lock level — the rule below applies to INPLACE and COPY only. With ALGORITHM=INSTANT, omit the LOCK clause or write LOCK=DEFAULT; NONE, SHARED and EXCLUSIVE are rejected — "Only LOCK = DEFAULT is permitted for operations that use ALGORITHM=INSTANT." ALGORITHM=INSTANT, LOCK=NONE is a failed statement, not a stronger guarantee.

   For INPLACE/COPY: specify LOCK=NONE when the matrix says concurrent DML is permitted. Where it is not (ADD FULLTEXT/SPATIAL INDEX, CONVERT TO CHARACTER SET, DROP PRIMARY KEY alone, adding an AUTO_INCREMENT column, and 8.0 REORGANIZE/COALESCE/REBUILD PARTITION), the best available is LOCK=SHARED — state it explicitly and budget the write outage, or escalate to gh-ost/pt-osc. Omitting the clause turns a planned decision into an unplanned one. Note that INPLACE, LOCK=NONE still rebuilds the table for ADD PRIMARY KEY, MODIFY … NULL/NOT NULL, and ROW_FORMAT changes: online is not free — budget the I/O, ~2× disk, and axis-D lag.

3. MDL contention — including INSTANT — all three algorithms can take an exclusive metadata lock. ALTER TABLE reference: "INSTANT: … An exclusive metadata lock on the table may be taken briefly during the execution phase of the operation." A long-running transaction holds it, the DDL queues, every later query queues behind the DDL — the usual cause of a "safe" change taking a site down, and INSTANT is not exempt. Guard every DDL, INSTANT included, with SET SESSION lock_wait_timeout = 3; before the statement, and pre-check SELECT * FROM information_schema.innodb_trx WHERE trx_started < NOW() - INTERVAL 30 SECOND;. (The What Is New page's "no metadata locks are taken" is a looser summary than the reference; not licence to skip the guard.)
4. Replication impact — MySQL 5.7 replicas apply DDL single-threaded; COPY causes severe lag. Will DDL duration exceed the replica lag SLA?

5.2 Data Integrity

5. NOT NULL + DEFAULT safety — adding NOT NULL to column with existing NULLs → ALTER fails. Use phased approach.
6. Type change truncation — narrowing VARCHAR, reducing DECIMAL precision → silent data loss. Widening may change algorithm.
7. FK cascade and algorithm risk — ON DELETE CASCADE on a large parent → uncontrolled write amplification. And ADD FOREIGN KEY is COPY unless foreign_key_checks=0: with checks on the server validates every child row and rebuilds the table; with checks off you get INPLACE but an unvalidated constraint. There is no online-and-validated option — see references/migration-anti-examples.md AE-13.
8. Index write amplification — each new index costs every INSERT/UPDATE. Check for redundant indexes (prefix of existing composite).

5.3 Backward Compatibility

9. Deployment ordering — column add → schema first, then app; column remove → app first, then schema; column rename → two-phase with dual-write.

10. Reversal path — MySQL DDL cannot be rolled back. Every DDL issues an implicit COMMIT before and after itself, so ROLLBACK after a completed ALTER does nothing. 8.0's atomic DDL is crash-safety for the dictionary + storage change, not user-visible undo.

    "Rollback" therefore means one of five concrete things. Name which one applies per phase:

    Path	Applies to
    Abort before cut-over	gh-ost/pt-osc runs and phased plans not yet at the switch — free
    Compensating DDL	Additive changes: drop the column or index you added
    Application revert	Anything behind dual-write or a feature flag; schema stays
    Roll forward	Bad backfill or wrong default — fix with another migration
    Restore / PITR	Irreversible loss: DROP COLUMN/TABLE, narrowed type, destructive UPDATE

    Classify each phase reversible (compensating DDL or app revert), reversible-with-loss (roll forward; rows written to the dropped structure are gone), or irreversible (restore only). Never emit a rollback: block that cannot restore state — an ADD COLUMN offered as the rollback for a DROP COLUMN recreates an empty column and reads as if recovery happened.

5.4 Operational Safety

11. Session guards — every migration session MUST set lock_wait_timeout and innodb_lock_wait_timeout before DDL.
12. Disk space — COPY needs ~2× table size. gh-ost needs ghost table + binlog backlog.
13. Idempotency — can the migration re-run after partial failure? MySQL ALTER TABLE has no IF NOT EXISTS / IF EXISTS for columns or indexes (that is MariaDB); writing it is a parse error. Only CREATE TABLE/DROP TABLE/CREATE DATABASE/DROP DATABASE accept it. Achieve idempotency by: the framework's history table (Flyway flyway_schema_history, golang-migrate schema_migrations); a pre-flight information_schema.COLUMNS/STATISTICS probe that decides whether to emit the DDL; one DDL per file so a partial failure has an unambiguous resume point; and a recorded checkpoint for batched backfills.
14. Statement granularity — one DDL per migration file, so a failure leaves an unambiguous state and the compensating DDL is obvious (there is no transactional rollback — see item 10). Exception: independent ADD COLUMNs should be grouped, which also costs one INSTANT row version instead of several.

§6 Execution Plan (Standard + Deep)

For non-trivial migrations, decompose into the standard phased pattern:

1. Phase 1 — Additive schema: add nullable columns, new indexes (online DDL)
2. Phase 2 — Backfill: populate from existing data in PK-ordered batches (see references/large-table-migration.md §4)
3. Phase 3 — App deploy: deploy code writing to both old and new schema
4. Phase 4 — Constraints: add NOT NULL, UNIQUE, or FK after backfill verified
5. Phase 5 — Cleanup (separate release): drop old columns, remove dual-write

Each phase requires: Pre-condition → SQL (with session guards) → Validation → Rollback → Go/No-go criteria.

For tables >10M rows requiring COPY, use gh-ost (default) or pt-osc (if inbound FKs). Details in references/large-table-migration.md.

§7 Anti-Examples

AE-1: Implicit algorithm — trusting server default

-- WRONG: server may silently choose COPY → outage on large table
ALTER TABLE users ADD COLUMN age INT;
-- RIGHT on 8.0.12+:
ALTER TABLE users ADD COLUMN age INT DEFAULT NULL, ALGORITHM=INSTANT;
-- RIGHT on 5.7 (INSTANT does not exist there; the line above would error out):
ALTER TABLE users ADD COLUMN age INT DEFAULT NULL, ALGORITHM=INPLACE, LOCK=NONE;

Explicit only helps when it is also correct for the version.

AE-2: NOT NULL on populated column without phased approach

-- WRONG: fails if any row has NULL
ALTER TABLE orders ADD COLUMN status VARCHAR(20) NOT NULL;
-- RIGHT: add nullable → backfill → enforce NOT NULL (see §6)

AE-3: DDL without session guards

-- WRONG: blocks indefinitely if long transaction holds MDL
ALTER TABLE large_table ADD INDEX idx_date (created_at);
-- RIGHT:
SET SESSION lock_wait_timeout = 3;
ALTER TABLE large_table ADD INDEX idx_date (created_at), ALGORITHM=INPLACE, LOCK=NONE;

AE-4: DROP COLUMN without data backup

-- WRONG: data gone forever
ALTER TABLE users DROP COLUMN legacy_field;
-- RIGHT: backup → wait one release cycle → drop

AE-5: Native COPY on 100M-row table

-- WRONG: hours of exclusive lock
ALTER TABLE events MODIFY COLUMN payload MEDIUMTEXT;
-- RIGHT: use gh-ost (see references/large-table-migration.md)

AE-6: Style nitpick reported as migration risk

-- WRONG: "WARN — column name 'usr_nm' violates naming convention"
-- RIGHT: only flag naming if it causes functional problems

AE-7 through AE-17 in references/migration-anti-examples.md: partition-clause algorithm rejection (AE-14), VARCHAR/INSTANT (AE-15), gh-ost mode confusion (AE-16), INSTANT row-version exhaustion (AE-17).

§8 Migration Scorecard

Critical — any FAIL means overall FAIL

• Algorithm explicitly specified for every ALTER TABLE (ALGORITHM=INSTANT|INPLACE|COPY), or the statement is a partition clause where the matrix says only DEFAULT is accepted
• Session guards set before every DDL (lock_wait_timeout, innodb_lock_wait_timeout)
• Every phase names its reversal path from the §5.3-10 table (abort / compensating DDL / app revert / roll forward / restore), with SQL where SQL can actually restore state and an explicit backup + retention plan where it cannot

Standard — 4 of 5 must pass

• DDL algorithm and lock verified against the exact server version, not the major version (INSTANT gates at 8.0.12/8.0.28/8.0.29; no INSTANT on 5.7; partition clauses per matrix §4)
• Replication impact assessed for each COPY/INPLACE operation
• Backward-compatible deployment order (additive before app, removal after app)
• Backfill uses PK-range batching, not LIMIT/OFFSET
• Validation SQL provided for each phase

Hygiene — 3 of 4 must pass

• Disk space impact estimated for COPY/gh-ost operations
• Re-runnable after partial failure — via the framework's history table or an information_schema pre-check, not IF [NOT] EXISTS, which ALTER TABLE rejects
• Post-deploy monitoring checks specified (replication lag, error rate)
• One DDL per migration file (or grouped ADD COLUMN justified)

Verdict: X/12; Critical: Y/3; Standard: Z/5; Hygiene: W/4. PASS requires: Critical 3/3 AND Standard ≥4/5 AND Hygiene ≥3/4.

§9 Output Contract

Every migration review MUST produce these sections. Omit none — write "N/A — [reason]" if inapplicable.

### 9.1 Context Gate
| Item | Value | Source |
| MySQL Version | 8.0.32 | SELECT VERSION() |
| ... | ... | ... |

### 9.2 Depth & Mode
[Lite/Standard/Deep] × [review/generate/plan] — [rationale]

### 9.3 Risk Assessment Table
| # | DDL Statement | Algorithm | Lock | Risk | Notes |

### 9.4 Execution Plan (Standard/Deep; "N/A — Lite" for Lite)

### 9.5 Migration SQL (with session guards, explicit algorithms)

### 9.6 Validation SQL

### 9.7 Rollback Plan (per-phase)

### 9.8 Post-Deploy Checks

### 9.9 Uncovered Risks (MANDATORY — never empty)
| Area | Reason | Impact | Follow-up |

Volume rules:

• UNSAFE: always fully detailed with mitigation
• WARN: up to 10; overflow to §9.9
• SAFE: summary row only
• §9.9 minimum: document all conservative assumptions made

Scorecard summary (append after §9.9):

Scorecard: X/12 — Critical Y/3, Standard Z/5, Hygiene W/4 — PASS/FAIL
Data basis: [full context | degraded | minimal | planning]

§10 Reference Loading Guide

§11 Deterministic Checker

scripts/lint_migration.py decides version-gated algorithm/lock questions mechanically. Run it on the migration under review, then reason about what it cannot see.

python3 scripts/lint_migration.py --mysql-version 8.0.29 path/to/migration.sql   # or a dir
python3 scripts/lint_migration.py --list-checks   # add --format json for machine output

29 checks: INSTANT version gates and its LOCK=DEFAULT-only rule (MM029), never-INSTANT operations, partition-clause algorithm support, LOCK=NONE on write-blocking operations, ADD FOREIGN KEY + foreign_key_checks, VARCHAR byte-boundary crossing, IF [NOT] EXISTS on ALTER, stored-program-only loops, sql_log_bin, gh-ost/pt-osc flag misuse, version-correct replication and lock-inspection statements, an unverified target version (MM028), and unread migration carriers (MM030).

Directory mode reads .sql, .ddl, .mysql, .md, .sh, .bash. A file named explicitly is scanned as SQL only if its extension is unknown; a known-unparseable carrier (Liquibase XML/YAML/JSON, Go/Java/Python) becomes an MM030 finding whether named explicitly or found in a directory — naming it does not make it parseable, and scanning it as SQL reports "clean" about DDL masked inside string values. --fail-on warning therefore refuses any input whose DDL nobody read.

Use it as evidence, not as the review. It reads statements, not your database, so it cannot decide whether a MODIFY changes the type or only nullability (compare SHOW CREATE TABLE), the VARCHAR band without both widths, or risk axes B–D. A clean run means "nothing here is rejected outright" — not that the migration is safe.

Three companions, all opt-in: verify_against_server.sh runs representative ALTERs against a real (disposable) server and reports any matrix claim it contradicts; mutation_sweep.py reintroduces each historical defect and requires the suite to catch it; run_model_eval.py grades with-skill vs without-skill responses on a deterministic rubric.