Interleaving-First Concurrency Control
Interleaving. Name the durable invariant and the schedule of reads and writes that can break it. Choose the least complex sufficient mechanism under the real store's isolation and failure model. Simultaneous activity alone is not a defect.
Choose the requested result
For audit or diagnosis, inspect state, writers and evidence read-only; explain the supported schedule and smallest closing mechanism. A plausible schedule is not a claim that it was reproduced. For design, specify arbitration, conflict and recovery at the requested fidelity. For build or fix, change the affected owner and run focused local checks, repairing failures introduced by the change.
Releasing production locks, forcing leadership, repairing raced data and running live load/fault tests require exact action, target and bounds authorization. Preflight and freshly verify authorized operations. Host permissions still apply; do not infer authority from credentials or a proposed test. Preserve unrelated work and settled decisions outside the requested change.
Identify the contested unit
Identify the row, key or cross-row invariant, and the writers that can affect it: requests, retries, replicas, jobs, consumers, services and operators where relevant. Reads feeding a write decision are part of the path. A missing writer is a coverage gap, not evidence of a single-writer guarantee. Start from the affected state rather than inventorying every writer in the entire system.
Construct the smallest counterexample: both actors read before either writes; both pass a preflight check; they update disjoint rows after reading one predicate; or a lease holder resumes after a successor acquires ownership. State the invalid final state and the observation that distinguishes it from correct behavior. If the invariant survives because effects commute, are idempotent or have a verified single owner, retain that result rather than adding a lock by default.
Select arbitration and conflict behavior
Read mechanism choices when choosing or reviewing conditional writes, constraints, isolation, locks or fencing. It is a comparison menu, not a requirement to reject every alternative one by one. Preserve an existing sufficient mechanism unless the requested change exposes a concrete gap.
Check the actual isolation level, participating transaction/connection and row-count or error result. All paths must use the arbiter. A preflight query, process-local mutex or exclusive-looking lease name cannot establish durable cross-process exclusion by itself.
Give losers an explicit disposition: surface conflict for a new decision, merge when the domain permits, or retry from fresh state within time/attempt budgets. Do not replay stale computations or multiply external effects. Preserve stable operation identity, jitter retries, and consider starvation under the relevant contention pattern. Cancellation or a timeout is not proof of rollback or absence of a remote effect.
Cover the selected failure windows
For a lease or leader, allow for pause, expiry, failover and return of a stale holder. Reject stale durable effects with the authority's generation/fencing check, or use effect idempotency and reconciliation that actually protect the invariant. Give each scheduled occurrence its own stable identity.
For locks and transactions, identify relevant connection/crash cleanup, deadlocks, acquisition ordering and bounded conflict recovery. A transaction-scoped row lock is not a wall-clock lease. If the arbiter is unavailable, state fail-closed or safe degraded behavior and the signal that reveals it. A hot-key retry storm may need admission or serialization rather than more retries.
Prove and finish
Force the implicated schedule with controlled coordination instead of relying only on many randomly concurrent workers. Check the business invariant, conflict result and committed state; the test should distinguish a missing mechanism. For fencing, let the old holder resume after the new generation and verify its effect cannot overwrite the successor. For deadlock handling, observe bounded termination and the intended victim policy, not just a timeout in the test harness.
Use the real store when isolation, constraints or locks are the claim; in-memory fakes can check orchestration, not those guarantees. Add bounded stress or load work when contention, starvation or capacity is claimed or project policy requires it, not as a gate for every diagnosis or narrow correctness change. Reuse existing fixtures and valid results for the same revision/environment.
An analysis can finish with a supported schedule or precise missing evidence. A requested runnable proof includes source, setup and assertions; label it unrun when its dependency is absent. A local fix finishes with the requested change and applicable checks, or the specific blocker. Do not invent a store cluster or claim production safety from a weaker test. Report schedule, mechanism, conflict path, observed evidence and residual windows without restating the whole ladder.
Adjacent job leases, broker ordering, provider idempotency, schema constraints and
PostgreSQL contention may use durable-background-jobs, reliable-messaging,
external-api-integration, postgres-schema-design or postgres-performance.
Load one only for a distinct unresolved decision; none is a prerequisite. Runtime
memory races belong to language-specific tooling, not this durable-state skill.