Transaction Isolation Levels

RU, RC, RR, S.

Overview

Transaction isolation levels balance consistency guarantees against throughput. Serializable everywhere is the easy default; right isolation per workload is the discipline that prevents both correctness bugs and unnecessary throughput tax.

• RU, RC, RR, S. Read uncommitted, read committed, repeatable read, serializable; the four standard levels.
• Read committed. Default for most engines; matches general transactional workloads; the safe baseline.
• Repeatable read. Postgres default in many configurations; matches read-heavy workloads with consistency needs.
• Serializable plus anomaly awareness. Strongest guarantee, lowest throughput; matches financial; know the dirty-read, non-repeatable-read, phantom anomalies.

The approach

The practical approach: pick the level that matches the workload, tune per transaction where needed, document the rationale, test for anomalies. The team’s discipline produces matched isolation, not blanket serializable.

• Read committed for general. Default for most workloads; the right baseline; matches reality of most application code.
• Serializable for financial. Strongest guarantee; supports correctness where money is involved.
• Per-transaction tuning. Some transactions need higher; raise per transaction, not for the whole connection.
• Document the choice. Per-table rationale committed to the repo; supports operational reviews.
• Test for anomalies. Per-isolation level tests; catches bugs that production would not surface for months.

Why this compounds

Isolation level discipline compounds across services. Each correct choice produces ongoing correctness; the team’s database expertise grows; new tables ship with the right isolation on the first try.

• Better correctness. Right isolation matches workload; the consistency anomalies the workload cannot tolerate are prevented at the engine.
• Better throughput. Right isolation balances cost; the throughput tax of over-strict isolation is avoided.
• Better debugging. Per-transaction documentation; the next debugger knows what guarantees the transaction had.
• Institutional knowledge. Each choice teaches consistency patterns; the team’s database engineering muscle grows.

Transaction isolation discipline is an operational discipline that pays off across years. Nova AI Ops integrates with database telemetry, surfaces patterns, and supports the team’s database discipline.