Repository navigation
feat(scheduler): bound the Advisor outage eviction hold and alert past it #1675
Description
When a recommendation collection can't fetch Azure retail prices, the scheduler keeps the previous cycle's Advisor reservation rows instead of evicting them. This lands with the cloud-commitments-go pin bump to 61111428. The hold has no bound.
Problem
If the price lookup keeps failing, for example because the Retail Prices API returns a 4xx after an API change, every collection holds again. The account then keeps rows from the last good collection indefinitely, and nothing alerts anyone.
Proposal
- Count consecutive outage cycles per account.
- After N cycles, stop holding and evict, so the dashboard doesn't show prices that are known to be stale. Pick N against the collection cadence, around a day's worth.
- Alert, or surface it in the freshness signal, once an account passes the bound.
- Match outage skips on a typed skip kind instead of message prefixes, once feat(azure): typed skip kinds on PartialResultError so consumers stop matching skip-reason prefixes cloud-commitments-go#670 (typed
SkipKinds()) lands. Today the platform matches two library message prefixes, and a library rewording would not fail CI.
Activity
cristim commentedon Nov 4, 2026
Also in scope here: incomplete_accounts can overcount by one when an account with a carried pricing hold also fails the same cycle for another reason (from the CodeRabbit review on #1677). An exact count needs per-sweep incomplete account IDs.
cristim commentedon Nov 9, 2026
Claimed by claude-darling: planning (#1715 released: blocked on #669; #1750 is claimed by qa-ui-3). Premise re-checked on main: the Azure pricing hold is carried across cycles in last_collection_partial (config/azure_hold_carry.go) with no bound; the typed SkipKinds bullet is already done in partial_sweep.go (countPricingOutageSkips reads SkipKind), so only the bound and the alert remain. One PR with self-merge on green.