Benchmark

Measures the overhead added by the audit extension for create, update, and delete operations.

What We Measure

Benchmark	Description
A) create — no audit	Baseline Prisma create() without extension
B) create — with audit	create() + audit log INSERT (after-only changes)
C) update — with audit + diff	update() + before/after diff calculation + audit log INSERT
D) delete — with audit	delete() + before-only changes + audit log INSERT

The update benchmark (C) is the most expensive because the extension must:

1. Fetch the existing record (before state)
2. Execute the update
3. Compute the diff between before and after
4. INSERT the audit log entry

Test Setup

• Database: PostgreSQL 16 (Docker, port 5433)
• Data: Fresh rows per benchmark (300 iterations each)
• Warmup: 30 iterations (discarded)
• Tracked model: User with password as sensitive field

Running Locally

# Start PostgreSQL
docker compose -f test/e2e/docker-compose.yml up -d

# Run benchmark
DATABASE_URL=postgresql://test:test@localhost:5433/audit_test \
  npx ts-node benchmarks/audit-overhead.ts

Results

Measured on Apple M-series, PostgreSQL 16, local Docker. Your results will vary.

Benchmark	Avg	P50	P95	P99
A) create — no audit (baseline)	0.40ms	0.40ms	0.52ms	0.57ms
B) create — with audit	1.44ms	1.37ms	1.84ms	3.11ms
C) update — with audit + diff	2.06ms	2.01ms	2.54ms	2.85ms
D) delete — with audit	1.71ms	1.57ms	2.09ms	3.91ms

Create overhead: +1.04ms (+260%) Update is the slowest at 2.06ms due to the additional findFirst (before state) + diff computation.

Interpretation

The audit extension adds ~1ms per write operation. This is the cost of the additional INSERT INTO audit_logs plus (for updates) a findFirst to capture the before state and compute the diff.

In absolute terms, even the slowest operation (update with diff) completes in 2ms — well under the threshold for any API endpoint. For most CRUD APIs where writes are infrequent compared to reads, this is negligible.

For bulk operations (createMany with 100+ records), the extension logs each record individually, so consider using noAudit context for batch imports and adding a single manual audit log entry instead.

Methodology

• performance.now() for millisecond-precision timing
• AuditContext.run() wraps each operation with actor context (matches real usage)
• Append-only rules are temporarily dropped for cleanup between benchmarks
• Sensitive field masking (password → [REDACTED]) is active during measurement