Rendered capacity profiles¶
Generated by python scripts/render_capacity_profiles.py --write from the chart.
The production/experiment render uses the synthetic CI provider-domain fixture solely to satisfy required schema validation; operators must supply real quotas and model IDs.
No credentials or live dependency limits are verified by this table.
| Profile | API pods | API processes/pod | Ingress active | Shared preflight global/org | DB pools control/foreground/telemetry/worker | Redis critical/cache/bulk | CPU request/limit | Memory request/limit | Production contract |
|---|---|---|---|---|---|---|---|---|---|
| Base | 2 | 1 | off | off | 20/8/5/5 | 64/16/16 | 250m/1000m | 512Mi/1Gi | no |
| Evaluation | 1 | 1 | off | off | 20/8/5/5 | 64/16/16 | 250m/1000m | 512Mi/1Gi | no |
| Production | 3–12 HPA | 1 | 100 | 300/100 | 20/8/5/5 | 64/16/16 | 500m/2000m | 1Gi/2Gi | yes |
| Experiment | 2 | 1 | 100 | 100/50 | 20/8/5/5 | 64/16/16 | 1/2 | 2Gi/4Gi | yes |
| Native evaluation | 1 | 1 | off | off | 20/8/5/5 | 64/16/16 | 250m/1000m | 512Mi/1Gi | no |
| Native production | 3–12 HPA | 1 | 100 | 300/100 | 20/8/5/5 | 64/16/16 | 500m/2000m | 1Gi/2Gi | yes |
Peak dependency budgets¶
The totals include API, batch, accounting roles, rollout overlap, migration, and reserved connections. The native request role owns two PostgreSQL connections. The native projection role owns eight connections. Both native roles have no Redis or provider connections.
For these two native roles, accounting_hot_path_db_pool_size owns the only
database pool. Helm uses this allocation for db_pool_size and the
DELTALLM_DB_POOL_SIZE startup setting. The legacy worker's control-pool default
does not set the native role's limit. API and legacy worker limits do not change.
| Profile | Peak API/request/projection processes | PostgreSQL used/maximum | Redis used/maximum |
|---|---|---|---|
| Base | 5/0/0 | 160/1000 | 544/2000 |
| Evaluation | 3/0/0 | 100/100 | 352/2000 |
| Production | 37/0/0 | 1508/2000 | 3680/5000 |
| Experiment | 7/0/0 | 368/2000 | 800/5000 |
| Native evaluation | 3/3/3 | 160/160 | 352/2000 |
| Native production | 37/4/4 | 1548/2000 | 3680/5000 |
Production and experiment mount the typed report.json capacity contract and reject a mismatch at startup. These allocation, ingress, preflight and transport settings require a restart; a durable configuration update cannot silently change them. The base and evaluation profiles do not enable the production contract.
The table does not show provider credentials, the DB/Redis URLs, or physical headroom. Read the capacity and autoscaling runbook before using a profile.