Partitioning and Capacity Operations Guide

This guide summarizes the operational policy for the four tables that grow without bound during operation — transactions, audit_logs, customer_score_history, and alerts — based on the currently implemented table structure and their real-world growth characteristics.

For the DDL templates themselves, see:

• transactions / audit_logs: ADR-0010, docs/compliance/data-retention.md
• customer_score_history / alerts: ADR-0011, docs/compliance/data-retention.md

This repository's migrations (migrations/*.sql) never retrofit partitioning onto tables that are already live in production. What follows is guidance for a new deployment building its environment from scratch, plus the general procedure a deployment already running non-partitioned tables would follow if it later migrates to partitioning.

1. Migrating an already-live table to partitioning (reference)

To migrate a non-partitioned table to declarative partitioning (PARTITION BY RANGE), the following steps are typical, since PostgreSQL cannot partition an existing table in place. This repository does not provide a migration that automates this procedure — it is disruptive work that requires downtime or a dual-write period, so the decision of whether and when to run it is left to the deploying organization.

1. Create the new table: create a new partitioned parent table under a name such as <table>_partitioned, and pre-create the monthly partitions you need.
2. Write in parallel: for a period, replicate writes to the existing table into the new table as well, using triggers or dual writes in the application.
3. Migrate the data: run INSERT INTO <table>_partitioned SELECT * FROM <table> in batches (for example, split by primary-key range or date range) to avoid lock contention.
4. Verify consistency: confirm the old and new tables match, using row counts, checksums, or similar checks.
5. Cut over: within a maintenance window, rename the tables to swap them, e.g. ALTER TABLE <table> RENAME TO <table>_old; ALTER TABLE <table>_partitioned RENAME TO <table>;. Create foreign keys, indexes, and grants on the new table ahead of time.
6. Keep the old table: retain the old table for a verification period (for example, one to two weeks) after cutover, and only drop it once you've confirmed there are no problems.

2. GIN index operator classes

For JSONB columns such as attributes / metadata / definition, prefer the jsonb_path_ops operator class (it produces a smaller index than the default GIN operator class and suits equality-heavy queries).

CREATE INDEX idx_customers_attributes_path_ops
    ON customers USING GIN (attributes jsonb_path_ops);

For a high-write table like transactions, the update cost of a GIN index is not negligible, so rather than indexing the whole JSONB column, index only the specific fields that need frequent lookups, using expression indexes.

CREATE INDEX idx_transactions_counterparty_country
    ON transactions ((metadata->>'counterparty_country'));

3. Capacity planning

As a sustained throughput target, use 100 transaction ingestions and evaluations per second on a standard configuration as the baseline (consistent with the PERF-001 latency target). If you expect load beyond that, plan a migration to a horizontally scalable setup such as Kubernetes.

4. Read-replica routing policy

When running a DB replica configuration (AVAIL-002), separate routing by query type as follows:

Query type	Routed to	Reason
Read-only queries: reporting, detection analytics, backtesting, etc.	Read replica	Spreads load off the primary
Real-time TM/CDD/screening evaluation	Primary	Avoids inconsistent evaluation results caused by replication lag

5. Recommended autovacuum tuning

For high-write tables (transactions, alerts, customer_score_history), we recommend self-hosted deployments tune autovacuum_vacuum_scale_factor below the default (0.2), to 0.02–0.05, to keep bloat from accumulating.

ALTER TABLE transactions SET (autovacuum_vacuum_scale_factor = 0.02);
ALTER TABLE alerts SET (autovacuum_vacuum_scale_factor = 0.02);
ALTER TABLE customer_score_history SET (autovacuum_vacuum_scale_factor = 0.02);

6. Handling archived partitions

• customer_score_history rows are only physically deleted once they've satisfied both the deploying organization's configured retention period and the 30-day grace period. If you detach/drop partitions as part of operations, confirm ahead of time that doing so does not shorten any row's individual retention deadline.
• alerts underpin case management and audit evidence, so keep detached partitions as an archive rather than dropping them.
• For how transactions and audit_logs are handled once their retention period elapses, see docs/compliance/data-retention.md and RET-002/003.