Skip to content

Data Access

Principle · data-access

Loads automatically whenever the agent touches code this principle governs. Install the skills · View source

The persistence layer is where clean models meet a database that does not care about them. Most of what goes wrong here is invisible in development — an N+1 that is imperceptible over ten rows, a transaction held open across a network call, a pool sized for a laptop, a migration that only fails when there is data.

All of it surfaces at the same moment: production, under load.

Config Resolution

  1. Read .msskills/config.yaml in the repo root; look up paths.persistence.
  2. A custom document exists at that path → read its frontmatter mode:
  3. override → use it alone; ignore the defaults below.
  4. overlay (default, or no mode key) → read defaults first, then apply the custom document's sections on top, matched by exact heading; new sections append.
  5. A path is configured but no file exists there → say which path is missing, then use the defaults.
  6. No config file or no paths.persistence key → use defaults.
  7. .msskills/stack.md exists → use its datastore, access style, and migration tool. Transactional guarantees, locking, and isolation differ enormously between a relational store, a document store, and a key-value store — name which assumptions the code relies on.

Self-Validation Checklist

STOP after writing or reviewing a repository, a query, a migration, or a cache. Verify every check. Fix failures before presenting.

  1. RETURNS DOMAIN TYPES: Does the repository return domain objects rather than ORM entities, result sets, query builders, or framework page types? A leaked persistence type couples every caller to the database (see clean-architecture).
  2. WHOLE AGGREGATE: Does a repository load and store a complete aggregate, with one repository per root and none for internal entities? (see domain-modelling)
  3. TRANSACTION BOUNDARY EXPLICIT: Is the transaction opened and committed in the application layer, around one use case, and is its extent obvious from reading the code?
  4. NO I/O IN A TRANSACTION: Does any transaction stay open across an HTTP call, a broker publish, or a long computation? → do the remote work outside it; a transaction holding locks while waiting on a network is how a database stalls.
  5. NO N+1: Does any code path issue a query per element of a collection? → fetch in one query, or batch. Check loops, lazy-loaded associations, and mappers that resolve references.
  6. BOUNDED RESULT SET: Does every query have a limit? An unbounded SELECT is a future out-of-memory error with a table that grew.
  7. INDEXED PREDICATE: Is every filter, join, and sort column covered by an index? For a new query on a large table, has the plan been checked rather than assumed?
  8. CONCURRENCY HANDLED: For a read-modify-write, what stops a lost update? Optimistic version check, a conditional update, or an explicit lock — one of them must be present and deliberate.
  9. LOCK ORDER CONSISTENT: Where multiple rows are locked, are they always acquired in the same order? Inconsistent ordering is a deadlock waiting for concurrency.
  10. PARAMETERISED: Is every query parameterised, with no string concatenation of input, including dynamic filters and sort columns? (see secure-service)
  11. KEYSET PAGINATION: Does deep pagination use a cursor rather than a growing offset? OFFSET 100000 scans and discards a hundred thousand rows.
  12. MIGRATION IS COMPATIBLE: Is the schema change compatible with both the currently deployed code and the new code, so rollout and rollback are safe?
  13. BACKFILL IS SAFE: Is any data backfill batched, throttled, resumable, and idempotent — never one transaction over a large table?
  14. CACHE HAS AN INVALIDATION STORY: For anything cached, what invalidates it, what is the maximum staleness, and what happens on a miss storm? A cache without an answer to all three is a bug with better latency.
  15. POOL SIZED: Is the connection pool sized deliberately against the database's limits and the number of instances, rather than left at a default?

All checks pass → state "Data access holds: , transaction , concurrency ."

Active Anti-Pattern Scan

Any box you can check is a defect. Fix it before presenting.

  • [ ] N+1 Query: a query inside a loop, or lazy loading triggered per element → single query or batch fetch.
  • [ ] Leaked ORM Entity: a managed entity, lazy proxy, or query builder returned past the repository → map to a domain type at the boundary.
  • [ ] Repository per Entity: repositories for objects inside an aggregate → roots only.
  • [ ] Fat Repository: a repository with dozens of bespoke finders serving screens → those are queries, not aggregate loading (see cqrs-and-consistency).
  • [ ] Unbounded Query: SELECT with no limit, or a findAll on a growing table.
  • [ ] Offset Pagination at Depth: OFFSET growing with page number on a large table → keyset.
  • [ ] Transaction Around a Network Call: an HTTP request, broker publish, or external API call inside an open transaction.
  • [ ] Transaction per Repository Call: each save its own transaction, so a use case cannot be atomic → own the boundary in the application layer.
  • [ ] Lost Update: read, modify in memory, write back, with no version check or conditional update.
  • [ ] Nested Transaction Confusion: relying on framework propagation nobody can explain → make the boundary explicit.
  • [ ] String-Built Query: any concatenation of input into SQL, a query document, or a sort clause.
  • [ ] Destructive Migration: a column or table dropped, renamed, or narrowed in the same release that stops using it → expand, migrate, contract.
  • [ ] Blocking Index Build: an index created without the concurrent option on a busy table.
  • [ ] Migration at Startup: schema changes run by the application on boot, so a failed migration crash-loops the fleet and concurrent instances race.
  • [ ] Cache Without Invalidation: cached values with no TTL and no invalidation path → stale forever, and impossible to reason about.
  • [ ] Default Pool Size: a connection pool left at the library default while running many instances → the database hits its connection limit before the service saturates.
  • [ ] SELECT *: fetching every column, including large ones nobody reads.

Ambiguity Signals

Route these through collaborative-judgment. Each has two defensible answers.

  • ORM or hand-written SQL. An ORM is productive for aggregate loading and obstructive for reporting queries. Using both — ORM for writes, SQL for reads — is a legitimate and common answer.
  • Optimistic or pessimistic locking. Optimistic is cheaper and pushes retry onto the caller; pessimistic serialises and risks deadlocks and timeouts. Depends on contention you may not have measured.
  • Isolation level. A stronger level removes anomalies and costs throughput. Read committed plus explicit version checks is usually right, and "usually" is doing real work in that sentence.
  • Where to cache. In-process is fastest and inconsistent across instances; a shared cache is consistent and adds a dependency and a failure mode.
  • Denormalising for read performance. Buys query speed, costs a synchronisation obligation forever.
  • Whether to add an index. Every index speeds reads and slows writes, and consumes storage. On a hot write path this is a genuine trade-off.
  • Soft delete or real delete. Soft delete preserves history and quietly poisons every subsequent query that forgets the predicate.