sql-server-optimizer

Capture an actual execution plan plus SET STATISTICS IO, TIME ON with representative parameters, identify the dominant bottleneck such as a non-sargable predicate or cardinality error, change one lever, then re-measure logical reads, CPU, and elapsed time under the same conditions.

Treat parameter sensitivity as a hypothesis to test, not a diagnosis. Compare actual plans and metrics across representative fast and slow parameter values, check data distribution and statistics, and avoid reflexive RECOMPILE hints or local-variable masking without measured evidence.

No. The skill is scoped strictly to SQL Server T-SQL and explicitly excludes PostgreSQL, MySQL, and Oracle. Its rules for execution plans, indexes, and isolation are SQL Server-specific and should not be applied to other database engines.

NOLOCK and READ UNCOMMITTED permit dirty and internally inconsistent reads, trading correctness for weaker observations without solving the underlying access-path or blocking design. The skill rejects blanket dirty reads and requires measured blocking evidence plus a correctness-preserving alternative.

Replace inline <%Variable%> tokens with native @parameter placeholders and configure the action's typed parameter mapping for each one. Wrapping an inline token in a DECLARE statement is still text substitution and does not remove injection or typing risk.

An index proposal needs exact key order with rationale, included columns, optional filter semantics, overlapping existing indexes, expected read benefit, write and maintenance cost, and a before/after measurement plan. Selectivity must come from data distribution, not column names.