[amanpruthi]

Summary

• Adds aws.cost.undiscounted — list-price cost per service/account/region, filtered to RECORD_TYPE IN (Usage, SavingsPlanCoveredUsage) only.
• Fixes a real blind spot: an account under a 100% promotional/committed AWS credit shows Credit == -Usage every day in Cost Explorer, so UnblendedCost/AmortizedCost read ~$0 while real infra consumption continues invisibly.
• Positive filter (include only the two record types that carry on-demand consumption value), not an exclude-list of every known discount/credit RECORD_TYPE name — so a new AWS discount type Cost Explorer adds later is automatically excluded instead of silently leaking into the total. Ported from the validated KEPT_RECORD_TYPES set in pde_daily_glassbox_cost_reporter_clone/l9cost/aws_cost.py (a 37-day, six-account measured audit).
• Separate GetCostAndUsage call (RECORD_TYPE can't be a third GroupBy dimension alongside SERVICE+REGION — Cost Explorer caps GroupBy at 2), merged into the same OTLP payload as an independent metric rather than joined onto the existing rows (a service can have Usage without net UnblendedCost after a credit, or vice versa).
• Adds CAPTURE_OFFSET_DAYS (default 3): Cost Explorer's daily cost is an estimate that keeps revising for ~2-3 days after the day passes. Since each (tenant, service, region, cost_date) combination is written to Last9's TSDB exactly once, and the TSDB accepts neither backfilled nor overwritten samples for the same labels+timestamp, a day captured too early (at the old DAYS_BACK=1/no-offset default) is permanently understated. CAPTURE_OFFSET_DAYS shifts the whole fetch window back so each day is only captured once its estimate has settled.

Verified live

• Deployed to a credit-covered test account: undiscounted rows export correctly, summing to a real per-day cost instead of the near-zero noise from UnblendedCost/AmortizedCost.
• Independently cross-checked exported values against raw aws ce get-cost-and-usage CLI output (same filters/GroupBy) — matched to within floating-point precision once CAPTURE_OFFSET_DAYS was applied; confirmed the pre-fix (no offset) values were measurably low (one line item was ~47% understated one day after capture, matching exactly once re-checked 3+ days later).

Test plan

• python test_parse.py — 10/10 pass, including tests for fetch_undiscounted_costs (asserts RECORD_TYPE is a Filter, never a third GroupBy dimension, and that zero-amount groups are skipped) and for CAPTURE_OFFSET_DAYS's window-shifting behavior.
• Live aws lambda invoke against a redeployed account, confirmed via direct Prometheus query that aws_cost_undiscounted_USD landed with correct per-service values matching raw Cost Explorer output.

🤖 Generated with Claude Code

[nishantmodak]

Request changes at 4fbd9646d94863ae0ac9e69ebc0195a982d7adb9 for two reproduced P2 defects: CloudFormation ignores the new capture-offset override, and a collection crossing UTC midnight queries different billing windows for its metric families. Each inline thread includes the required fix and a complete executable regression.

Reviewed the complete six-file diff, collector/export paths, both deployment adapters, configuration, tests, CI and discussions. Local Python 3.13 validation: all 10 existing head tests and five added export/pagination/error controls pass, including 64 seeded amount/label cases. Both defect assertions fail twice on head and pass with local corrections. The same rollover assertion passes merge-base b59a2672f38a08a56d22c8395e2d8c7304d7a486; the capture option is new and has a passing direct-Lambda control.

Limits: external AWS/HTTP/package commands were mocked; no live billing, deployment or TSDB verification. Broader list-price/RI contracts and account-discovery pagination remain pending. GitHub tests pass, but both lint checks fail. No approval pending these fixes and remaining checks.

[nishantmodak]

Approved at c45fc7475ecc463daeb15b7c77d3d510ba9d9b49. Both reported defects are fixed: CloudFormation receives the capture offset, and poll/Lambda compute one billing period for all metric families. The original attached regressions, recovered from these threads, pass unchanged. Finished the previously outstanding review scope: complete seven-file diff, collector/exporter and callers, configuration/deployment adapters, tests, account/filter semantics and failure paths. No remaining introduced blocker found.

Local Python 3.13.15: 13 repository tests and 9 regression/control/probe tests pass, including 64 seeded amount/label cases, independent export with empty net rows, cost-result pagination, Lambda window sharing and fetch/export errors. Current-head GitHub lint/tests pass, including the CUR fixture change. AWS, HTTP and deployment commands were synthetic boundaries; no cloud deployment, live billing or TSDB overwrite/backfill verification. The metric covers the documented Usage/SavingsPlanCoveredUsage filter; this review does not establish universal public-list-price equivalence for RI-covered or negotiated usage.