For the complete documentation index, see llms.txt. This page is also available as Markdown.

Limits & performance

Size, schedule, and tune PostgreSQL workloads for reliable production operations.

Set practical resource and concurrency limits before production demand grows. These guardrails protect database stability as workloads, API traffic, and BI usage increase.

Operational boundaries

Compute sizing

A 1 to 2 vCPU, 2 to 4 GiB instance is usually enough for dev. For first backfills, temporarily raise work_mem / maintenance_work_mem and ensure disk headroom. Index the incremental time_column / unique_key and size IOPS to write volume. For the semantic API and concurrent BI, add a pooler or read replica and watch the connection count.

Concurrency and API replicas

Start with concurrent_tasks: 2 for daily incrementals; raise it only if plan/run duration is the bottleneck and the instance has connection and I/O headroom. The API is stateless, but each replica consumes PostgreSQL connections, so pair replica increases with PgBouncer or higher connection capacity. Size max_connections for Vulcan workers, API replicas, and BI sessions together.

Scheduling and latency floor

Schedule after upstream lands, set timezone: UTC, set endOn 1 to 2 years out, and use concurrencyPolicy: Forbid (prevents overlapping windows and connection pile-ups). The latency floor: typically no cold start on provisioned PostgreSQL (serverless or auto-paused deployments may have wake-up latency); connection establishment is sub-second when warm; queries on unindexed filters pay sequential-scan cost.

Performance reference

Measure active connections versus max_connections, plan/run duration per model, semantic query wall time, rows touched per incremental window, and the sequential-scan versus index-scan ratio (EXPLAIN). When concurrency degrades, fix in order: remove cross-join-shaped queries, add missing indexes on hot filter/join columns, add a connection pooler such as PgBouncer, add a read replica for query traffic, then raise instance size last.

Measured results from a 50M-row local run (2026-06-12) against a managed PostgreSQL 15.17 instance show the shape clearly: a point lookup by primary key returns in about 49 ms (index scan), while a 3-way join aggregate over 50.1M orders runs about 35 seconds (sequential scan ×3), and a COUNT(DISTINCT customer_id) runs about 28 seconds. statement_timeout enforcement is confirmed. A Vulcan-orchestrated full-refresh backfill of the same data (three FULL models in parallel: plan, snapshot, CTAS, audits) completed in about 2 minutes 28 seconds wall clock. Reproduce on your own instance before quoting these.

Next step

When behavior needs investigation, use Troubleshooting.

Last updated

Was this helpful?