Superior Database

Make data changes correct for the actual datastore, workload, and rollout environment. Preserve established migration tools and ownership boundaries; do not import generic database conventions without verifying engine behavior and repository practice.

Establish the data contract

Before changing a schema, query, or migration, determine:

• datastore engine, version, topology, and relevant extensions;
• authoritative schema and migration history;
• entity ownership, identifiers, relationships, and tenant boundaries;
• invariants that must hold at write time and read time;
• expected cardinality, access patterns, ordering, and retention;
• compatibility window across application versions, workers, and replicas;
• data volume, traffic shape, lock sensitivity, and recovery expectations.

Inspect effective production constraints only through authorized, non-sensitive evidence. Do not assume local fixtures represent production volume, skew, nullability, or legacy states.

Design schema and queries together

Read references/schema-query-and-index-design.md when adding or changing tables, collections, fields, relationships, indexes, query shapes, pagination, or retention behavior.

• Encode durable invariants in the datastore when the engine and rollout permit it.
• Keep application validation for user feedback, but do not rely on it as the sole protection against concurrent or alternate writers.
• Select indexes from real predicates, joins, ordering, uniqueness, and cardinality rather than indexing every filtered field.
• Verify query semantics for nulls, collation, timezones, ordering ties, and empty sets.
• Bound result sets and write batches; avoid hidden full scans, unbounded fan-out, and per-row queries.

Do not denormalize, partition, shard, or introduce a cache without demonstrated workload pressure and a clear consistency owner.

Preserve transactional truth

Read references/transactions-and-consistency.md when work spans related writes, concurrent actors, replicas, queues, caches, idempotency, or cross-service workflows.

• Define the invariant and the exact boundary that must be atomic.
• Use datastore guarantees deliberately; do not infer serialization from application request order.
• Keep transactions short and avoid remote calls while holding locks unless the established architecture proves it safe.
• Make retryable operations idempotent and distinguish transaction retry from business-operation replay.
• Treat cache invalidation, replica lag, and event publication as explicit consistency contracts.

Plan migrations as deployments

Read references/safe-migrations-and-backfills.md for production-facing schema changes, constraint introduction, type changes, table rewrites, backfills, data repair, or removal of old representations.

• Prefer compatibility-first expand, migrate, verify, and contract phases when multiple application versions may coexist.
• Identify locks, rewrites, validation cost, replication impact, and failure recovery before execution.
• Make backfills bounded, observable, resumable, and idempotent where practical.
• Define verification and rollback or forward-recovery behavior before destructive or irreversible steps.
• Never run a production migration, destructive query, broad data correction, or shared-environment load test without explicit authorization and the intended environment.

A migration file compiling locally is not evidence that its production execution is safe.

Verify the result

Use the repository's existing database and test tooling. Verify proportionately:

• migration applies from the supported previous state and reaches the intended schema;
• rollback works when it is promised and semantically possible;
• old and new application versions behave according to the compatibility plan;
• constraints reject invalid states and preserve valid legacy data;
• representative queries return correct results with stable ordering;
• query plans and runtime behavior are checked against representative shape and scale when performance is in scope;
• backfill counts, checkpoints, error handling, and reruns preserve invariants.

Report commands actually run, datastore/version used, data scale represented, and conditions not exercised. Never claim zero downtime without evidence for the relevant engine, topology, workload, and deployment sequence.

Combine skills narrowly

• Use superior-logic for domain behavior, API contracts, and application orchestration around persistence.
• Use superior-testing when migration, integration, compatibility, or concurrency test design is a primary concern.
• Use superior-debugging when a data symptom exists but its causal mechanism is uncertain.
• Use superior-performance when representative query, capacity, resource, or latency measurement is a primary concern.
• Use superior-reliability when replication, availability, failover, backup objectives, or recovery readiness is primary.
• Use superior-delivery when migrations must be coordinated with artifact promotion, rollout, or rollback.
• Use superior-api-security for tenant isolation, unsafe query construction, sensitive-data exposure, or abuse-driven resource controls.
• Use superior-docs after verified schema, migration, recovery, or operational procedure changes.

Do not broaden a focused database change into a datastore replacement, ORM rewrite, or infrastructure migration without explicit scope.