Performance & high traffic

Measured overhead is 0.4–0.6 ms per request (median, 100 models loaded, without persisting), and memory stays flat for a command that updates 10,000 records.

How it stays cheap

• Nothing is written in the middle of a request. Entries are persisted in the terminating phase, after the response was sent.
• Model changes are buffered and flushed every buffer_size changes, so long imports don't grow in memory.
• Retrieved models are stored as class => ids with at most max_ids_per_model ids.
• State is scoped per lifecycle, so it does not leak between requests on Octane or between jobs on a worker.

Tuning for high traffic

1. Queue persistence — AUDITOR_QUEUE=true (and a dedicated queue with AUDITOR_QUEUE_NAME). The request only dispatches a small job.
2. Sampling — http.sample_rate = 0.1 keeps 10% of plain requests; entries with changes or denied abilities are always kept.
3. A separate database connection — AUDITOR_DB_CONNECTION.
4. Retention — schedule auditor:prune with a sensible AUDITOR_KEEP_DAYS.
5. Turn off what you don't need — models.track_retrieved, listeners.*, jobs.enabled, console.enabled.
6. Partitioning — on very large tables, partition auditor_entries and auditor_model_changes by created_at range in your own migration.

The entries list uses cursor pagination, so it stays fast on millions of rows.