The engine’s database — runbook

Sizing worksheet

• Instances/month: ______ × avg activities/instance: ______
• History level (9.02): ______ → rows/instance ≈ activities + tasks + variables
• Keep window: ______ months
• Steady-state history rows ≈ rate × rows/instance × keep window = ______
• Runtime rows ≈ live instances × (tokens + open tasks + variables). This is small; verify it stays small (a growing runtime table is a stuck-work signal, 9.04).

Indexes

• Engine-shipped runtime indexes: leave alone; never add to ACT_RU_* without a measured engine-level regression justifying it (every index taxes token writes).
• Add for your query load on history, typically:
  • ACT_HI_PROCINST (END_TIME_) — cleanup + completed-work dashboards
  • ACT_HI_PROCINST (BUSINESS_KEY_) — case lookup by application id
  • ACT_HI_ACTINST (PROC_INST_ID_, START_TIME_) — timelines
  • ACT_HI_VARINST (NAME_, TEXT_) only if you must query by variable value — and consider a reporting replica instead.

Cleanup-job guardrails (retention deletes, version pruning)

• [ ] Bounded batches (e.g. 10k rows), commit per batch, sleep between batches.
• [ ] Off-peak schedule; abort switch checked between batches.
• [ ] Archive verified (counts reconcile) before the delete of any batch.
• [ ] Replication lag watched during the run; pause threshold defined.
• [ ] END_TIME_-based only; open instances untouchable (9.02).

Contention triage

Restore drill (quarterly)

1. Snapshot the production schema (one snapshot — runtime + history together).
2. Restore to a scratch engine; point a capstone driver at it.
3. Verify: parked instances resume; history timelines are complete; probe (9.04) shows no dead letters created by the restore.
4. Record time-to-restore; that number is your real RPO/RTO, not the one in the DR document.