SQL Safety Reference

SQL Safety Reference

dbward processes every SQL statement through three safety layers before execution:

1. Classification — determines the operation type
2. Review — checks for risky patterns
3. Risk scoring — calculates overall risk level for auto-approve decisions

---

SQL Classification

Every statement is classified into one of three categories:

ExecuteSelect (read-only)

• SELECT
• SHOW
• EXPLAIN (without ANALYZE)
• EXPLAIN ANALYZE on SELECT statements
• SET (safe session variables only)

ExecuteDml (write operations)

• INSERT, UPDATE, DELETE, MERGE
• COPY
• CALL (stored procedures)
• CREATE TABLE, CREATE VIEW, CREATE INDEX
• ALTER TABLE
• SELECT with writable CTE (WITH ... INSERT/UPDATE/DELETE)
• SELECT with dangerous functions (24 known functions)
• EXPLAIN ANALYZE on DML statements (actually executes the inner statement)
• SELECT INTO

DestructiveDdl (subcategory of ExecuteDml)

These statements are classified as ExecuteDml but flagged as DestructiveDdl. They pass through the classifier and are controlled by sql_review rules:

• DROP TABLE ¹
• DROP VIEW ¹
• DROP INDEX ¹
• DROP SEQUENCE ¹
• TRUNCATE ¹
• CREATE SEQUENCE ¹

¹ Requires Permission::RequestDdl (request.ddl). Controlled by sql_review rules (default: block for drop_table/truncate, warn for others). --allow-ddl bypasses sql_review blocks. See Break-Glass.

Rejected (blocked by default)

• DROP SCHEMA/DATABASE/FUNCTION/ROLE
• CREATE FUNCTION/PROCEDURE/TRIGGER/ROLE/DATABASE
• GRANT, REVOKE
• BEGIN, COMMIT, ROLLBACK, SAVEPOINT
• LOCK TABLE
• LOAD DATA
• SET (unsafe variables)

Special rules

Condition	Result
Parse failure	Classified as ExecuteDml (fail-closed: requires approval)
Unknown statement type	Classified as ExecuteDml
Input > 1 MB	Rejected
> 100 statements	Rejected
Multiple SELECT statements	Rejected (use single SELECT)
SET + SELECT combo	Allowed
NULL bytes in input	Rejected

---

SQL Review (14 rules)

Each rule has a configurable severity: warn, block, or off.

• block (default for no_where_delete, no_where_update, drop_table, truncate) — rejects the request regardless of workflow (DDL rules ² can be bypassed with --allow-ddl)
• warn (default for drop_column, not_null_without_default, create_index_not_concurrently, alter_column_type, mixed_ddl_dml, large_in_list, drop_index, drop_view, drop_sequence) — adds a finding to the risk assessment
• off (default for create_sequence) — rule is disabled

[[sql_review]]
database = "*"
environment = "development"
no_where_delete = "warn"
no_where_update = "warn"
drop_table = "warn"
drop_column = "warn"
not_null_without_default = "warn"
create_index_not_concurrently = "warn"
alter_column_type = "warn"
truncate = "warn"
mixed_ddl_dml = "warn"
large_in_list = "warn"
drop_index = "warn"
drop_view = "warn"
drop_sequence = "warn"
create_sequence = "off"

Rule descriptions

Rule	Fires when	Risk
no_where_delete	DELETE without WHERE clause	Entire table deletion
no_where_update	UPDATE without WHERE clause	Entire table overwrite
drop_table	DROP TABLE detected	Permanent data loss
drop_column	ALTER TABLE DROP COLUMN	Column data loss
not_null_without_default	ALTER TABLE ADD COLUMN NOT NULL without DEFAULT	Fails on existing rows
create_index_not_concurrently	CREATE INDEX without CONCURRENTLY (PostgreSQL)	Table lock during build
alter_column_type	ALTER COLUMN ... TYPE	Table rewrite, potential data loss
truncate	TRUNCATE TABLE	All data removed
mixed_ddl_dml	DDL and DML in same request	Complex rollback
large_in_list	IN (...) with > 100 values	Performance concern
drop_index	DROP INDEX detected	Index removal, query performance impact
drop_view	DROP VIEW detected	View removal, dependent queries break
drop_sequence	DROP SEQUENCE detected	Sequence removal, ID generation impact
create_sequence	CREATE SEQUENCE detected	New sequence creation

² DDL rules (drop_table, drop_column, truncate, create_index_not_concurrently, alter_column_type, not_null_without_default, drop_index, drop_view, drop_sequence, create_sequence) can be bypassed with --allow-ddl. DML safety rules (no_where_delete, no_where_update, large_in_list) and mixed_ddl_dml are never bypassable.

---

Risk Scoring

Risk scoring determines whether a request qualifies for auto-approve.

Levels

Level	Numeric	Meaning
Low	1	Safe (SELECT, safe DDL, small tables)
Medium	2	Moderate concern
High	3	Significant risk
Critical	4	Reserved
Unknown	5	Cannot assess (schema not synced)
Unavailable	6	Parse failure

Scoring rules

Condition	Level
SELECT + allow_read_only	Low
Safe DDL (CREATE TABLE/VIEW/INDEX) + allow_safe_ddl	Low
Schema not synced	Unknown
Multi-statement DML (>1 DML)	High
DROP or TRUNCATE detected	High
≥ 3 review warnings	High
CASCADE FK + large table (> max_estimated_rows)	High
CASCADE FK + small table	Medium
Large table without cascade	Medium
1-2 review warnings	Medium
None of the above	Low

Safe DDL

These CREATE/ALTER statements are considered safe regardless of table size:

• CREATE TABLE (new table, no existing data)
• CREATE VIEW
• CREATE INDEX CONCURRENTLY (PostgreSQL only)
• ALTER TABLE ... ADD COLUMN (PostgreSQL only, no lock on existing rows)

---

Dangerous functions (24)

Functions that promote a SELECT to ExecuteDml:

dblink, dblink_exec, dblink_connect, lo_export, lo_import, lo_unlink, pg_read_file, pg_read_binary_file, pg_ls_dir, pg_execute_server_program, copy_to, copy_from, set_config, pg_cancel_backend, pg_terminate_backend, pg_sleep, pg_advisory_lock, pg_advisory_xact_lock, pg_notify, sys_exec, sys_eval, load_file, sleep, benchmark

---

See also

• Auto-Approve — how risk scoring drives auto-approve
• Policies Overview — fail-closed principle
• Configuration Reference — full config fields