[herosjourney]

Problem

Running heroku-to-aws against a large Heroku Enterprise account (hundreds of apps, multi-terabyte Postgres, dozens of Redis instances, Kafka in Private Spaces, a long tail of third-party add-ons) left important parts of the estate as "Deferred":

• the largest Postgres databases — plans private-8, private-9, private-xl-9 have no sizing row, so the biggest data stores in the estate were unsized
• Redis instances on plan mini
• Kafka add-ons on private-standard-*
• every scheduler add-on — Heroku Scheduler is reported by the CLI as slug scheduler, but the fast-path row was keyed heroku scheduler, so each one fell through to the specialist gate; advanced-scheduler had no row at all. Together these were the large majority of the unmapped add-ons.
• rediscloud, auth0, herokuconnect, jawsdb, heroku-inference had no fast-path row

The same run also multiple-counted one app's data volume (a leader plus followers each reported at full size) and had no way to say how much of the estate is production. Separately, the pricing cache said Private and Shield Postgres plans are priced equal to Premium; they are not, and the Redis sizing table carried plan memory values roughly 10× what Heroku publishes today.

Fix

Sizing tables (knowledge/design/)

• postgres-rds-sizing.json: standard-8/9/10, and premium/private/shield -8, -9, -10, -l-6, -xl-6, -l-9, -xl-9. Instance classes follow the table's existing ladder; every class already has a rate row in the vendored pricing file.
• redis-elasticache-sizing.json: mini, private-9/10/12/14, shield-3..14. Also corrects memory_limit on premium-5..14 and private-3..7 to the plan sizes Heroku publishes (which the skill's own heroku-pricing-cache.md already carried). private-7 was 100 GB; it is 7 GB. The old values pushed ElastiCache two to three node classes too high.
• kafka-msk-sizing.json: private-standard-0/1/2, shield-standard-*, shield-extended-*.
• fast-path-addons.json: row keyed scheduler with a heroku scheduler alias; new rows advanced scheduler, rediscloud, jawsdb, jawsdb maria, auth0, herokuconnect, inference. Targets with no rate row say so in their notes; the Estimate's existing unpriced handling (warn + exclude from total) applies — no rates were invented.

Discovery / design

• discover-live.md: parse Following: / Forked From: from pg:info into config.pg_role / config.pg_leader (Classic path; Advanced-tier add-ons already carry this in advanced_pools[]).
• design-mapping.md 2B: followers collapse onto the leader (no separate instance; warning + metadata.postgres_followers_collapsed[]), forks get a confirm-with-customer warning, Terraform-only runs get a "cannot tell leaders from followers" warning per multi-DB app.
• discover-assemble.md rule 5a + schema: environment_tier per app (pipeline stage first, else name-token heuristic), counts in metadata.apps_by_environment_tier.
• design-mapping.md slug table: new slugs plus a note that deferring scheduler as "Not found in Fast_Path_Table" is a normalization bug.

Pricing cache: full Postgres Classic plan lists (Standard/Premium/Private/Shield) with current public prices from elements.heroku.com and devcenter. The Advanced-tier section is untouched. tests/test_heroku_postgres_advanced_pricing.py's Classic-row guard now compares cell values rather than column padding and expects the corrected private-0/shield-0 prices.

Ported from awslabs/startups#364, which is closed in favour of this PR. Independent of the companion estate-workbook PR; either can land first.

Scope

plugins/aws-startup-advisor/skills/heroku-to-aws/ only: four knowledge/design/*.json tables, references/phases/design/design-mapping.md, references/phases/discover/discover-live.md, references/phases/discover/discover-assemble.md, references/shared/schema-discover-heroku.md, references/shared/heroku-pricing-cache.md; plus one guard test in plugins/aws-startup-advisor/tests/. No root or .github changes.

Known limitations: rate rows for RDS MySQL/MariaDB, Cognito, AppFlow and Bedrock are not added to skills/shared/pricing/aws-infra-pricing.json (shared with azure-to-aws), so those new fast-path targets render as unpriced lines until a separate change adds them. Plan specs and prices were cross-checked against devcenter.heroku.com/articles/heroku-postgres-plans, …/heroku-postgres-and-private-spaces and elements.heroku.com/addons/heroku-redis (Oct 2026); no live run against a customer account is included.

Type of Change

Select at least one.

• New skill
• Bug fix
• Documentation update
• CI/CD change
• Other

Startup Advisor review checklist

By submitting this PR, I confirm that:

• I have reviewed the full diff and understand all changes I am submitting.
• I have tested my changes and confirmed they work as intended.
• I have run mise -C plugins/aws-startup-advisor run ci and all checks pass.
• I have disclosed any known issues, limitations, or untested scenarios.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[leon1418]

[🤖 AI review 🤖]

Reviewed b16463067f13af25e680e8f07c91be07f244211f against 00840a095dcc53242dcb19402e7140fa896d2d46 with OCR-assisted repository review and one independent general reviewer. Five corrections are needed:

• Preserve follower read capacity across Design, Estimate and Generate, including unresolved leaders.
• Size the nine affected Redis rows using memory available after the default reserve.
• Give the 13 new large Aurora mappings adequate per-instance capacity or an explicit deferral.
• Load and populate the Redis Cloud sizing/configuration contract.
• Make all documented Heroku Connect aliases survive normalization.

Validation: all ten changed files and relevant unchanged consumers reviewed; 22 original-source tests passed; core rate coverage, vendor plan prices, alias and capacity checks completed. Current-head CI reports success. Static contract probes were used; no live account migration or AWS deployment was run. Details and bounded acceptance criteria are in the inline comments.

[herosjourney]

Reviewed: head b164630 (base 00840a0), 10 files (+609/−81). Only heroku-to-aws is touched, in the single plugin tree plugins/aws-startup-advisor (this repo has no migrate/ twin).

Ran:

• pytest plugins/aws-startup-advisor/tests -q: 1475 passed, 225 subtests passed.
• Changed JSON parses: 4/4 OK.
• tools/sync-plugin-skills.py --check: in sync.
• sync-vendored.py --check: PASS, 62 files.
• check-skill-refs.py --check: PASS, 1498 refs, 0 missing.
• validate-artifacts.py --fixtures: PASS, 104 artifacts.
• A script cross-checking every sizing row against heroku-pricing-cache.md and against instance memory.

GitHub CI (startup-advisor, CodeQL, author checklist) is green at head and not restated.

Primary sources (read 2026-11-10):

• Heroku Dev Center: heroku-postgres-plans, heroku-postgres-and-private-spaces, heroku-postgres-follower-databases, heroku-inference-api-model-cards.
• Heroku Elements: heroku-postgresql, heroku-redis, heroku-kafka.
• AWS: AuroraUserGuide Concepts.DBInstanceClass.SupportAurora; RDS USER_ReadRepl and Concepts.MultiAZ; AWS Price List API (AmazonRDS, us-east-1).

Did not run: live AWS calls, an end-to-end skill run, terraform plan, or mise run ci locally (CI ran it). ElastiCache node memory figures were not re-fetched; the node-type names were confirmed.

Cross-plugin parity: PASS. There is a single tree, and vendored parity is unchanged.

Merge state: merges cleanly into main. It has a textual conflict with #610 in references/shared/schema-discover-heroku.md, and shares files with #609, #611 and #614 without textual conflict. #609 reads environment_tier / apps_by_environment_tier, which only this PR produces, so landing #608 → #609 is the natural order.

Change narrative: A large Enterprise run deferred its biggest databases, Redis mini, Private/Shield Kafka and Scheduler, because lookup rows were missing or mis-keyed. The PR:

• adds sizing rows;
• re-keys Scheduler to its CLI slug;
• adds 7 fast-path add-ons;
• corrects Private/Shield Postgres prices and Redis memory sizes;
• adds follower/fork and environment-tier discovery.

Body vs diff: it mostly matches. The body says it adds "premium … -8, -9", but premium-8/premium-9 already exist at base.

Verified correct:

• All Standard/Premium/Private/Shield Postgres plan names, RAM, storage and $/month in the cache match Heroku Elements and Dev Center (56 rows). Private and Shield really are priced above Premium.
• Redis plan sizes and prices match Elements (25 plans, including mini at $3).
• Kafka private-standard-* / shield-* names and prices match Elements.
• The plan sets in the sizing JSONs and the cache are identical.
• Every RDS class meets or exceeds the plan RAM.
• MSK storage covers the Heroku capacity on all 19 Kafka rows.
• Rate-card RDS and Aurora rates match the Price List.
• The Scheduler re-key fixes a real miss.

Axis	Stance	One line
Parsimony	advisory	Mostly data-only and in idiom. Three new producer fields have no reader, and runtime strings carry revision history.
Completeness	finding	Follower collapse is false on the RDS branch and for cross-app followers. The RedisCloud sizing instruction can't be reached. No regression test covers the headline fix.
Security	clear	No new command (pg:info was already captured), no IaC, no credentials.
Alternatives	advisory	Two below.
Contradictions	finding	The RedisCloud row says "ElastiCache Redis" vs the Valkey default. It sizes from memory_limit, which the sizing file declares provenance-only. The Heroku Inference note contradicts Heroku's model cards.
Overlap	advisory	Plan RAM is restated in the cache and in the JSONs; a parity test would keep them from drifting.
Cross-plugin parity	clear	Single tree; vendored and skill parity PASS.
Propagation & data	finding	SSOT: heroku-pricing-cache.md + knowledge/design/*.json. Consumers swept: design-mapping, estimate-cost-engine, generate-terraform, generate-report, clarify-interview. The Aurora 768 GB/1 TB rows are undersized.

Blocking

1. [Completeness / Contradictions] Follower collapse removes databases and read capacity the AWS design never replaces — references/phases/design/design-mapping.md:261

A Heroku follower is a read replica of a leader on the same app; on AWS that role is served by the leader's Aurora reader or RDS Multi-AZ standby / read replica.

Both premises contradict the primary sources:

• Heroku: "Followers don't need to be tied to the same application … you can provision the leader and follower databases on separate applications" (heroku-postgres-follower-databases). The follower shows its leader by attachment name ("Following: DATABASE"), so pg_leader does not identify an inventory resource.
• AWS: "A Multi-AZ DB instance deployment has one standby DB instance that provides failover support, but doesn't serve read traffic" (Concepts.MultiAZ).
• The RDS branch, the default path, emits a single aws_db_instance (generate-terraform.md:1747) and no read replica.

Consequence: on the default path, an app that reads from HEROKU_POSTGRESQL_<COLOR>_URL loses that endpoint, while the design tells the user reads are covered. A follower whose leader lives on another app gets no database designed at all.

Fix: collapse only when pg_leader resolves to a heroku-postgresql resource on the same app in this inventory; otherwise design it normally and warn. Make the claim branch-accurate:

"On Aurora, reads go to the cluster reader endpoint. On RDS, a Multi-AZ standby does not serve reads — add a read replica if the app routes reads to this follower."

Also drop "excluded from data-volume totals": no data-volume total exists in Design, Estimate or Generate. Regression: a worked-case table in 2B covering:

• same-app follower × RDS;
• same-app follower × Aurora;
• cross-app follower;
• fork.

Should-fix

1. [Propagation & data] The Aurora class is half the plan RAM on every 768 GB / 1 TB row — knowledge/design/postgres-rds-sizing.json:141 (same on all -9/-l-9/-xl-9/-10 rows; pre-existing at premium-9)

"aurora_instance_class": "db.r6g.16xlarge"

db.r6g.16xlarge is 512 GiB. Aurora PostgreSQL supports db.x2g.16xlarge (1024 GiB). The cross-check flagged 14 undersized Aurora rows and 0 RDS rows. The estimate also understates cost by (12.064 − 8.306) × 730 × 2 = $5,486.68/month per database (Price List, us-east-1).

Fix: set db.x2g.16xlarge on those rows. Add "db.x2g.16xlarge": 12.064 under aurora_postgresql.instances in skills/shared/pricing/aws-infra-pricing.json, then run tools/sync-vendored.py.

2. [Contradictions / Completeness] The RedisCloud row has the wrong engine label and a sizing instruction Design can't reach — knowledge/design/fast-path-addons.json:58

• (a) redis-elasticache-sizing.json loads only when "inventory has a heroku-redis addon" (design.md:13), so a RedisCloud-only app never has it in context.
• (b) That file says memory_limit is "PROVENANCE ONLY … do not recompute node_type from it".
• (c) "aws_services": ["ElastiCache Redis"] (:57) is the Redis OSS exception label. The default engine is Valkey (design-mapping.md:351/362).

Fix: use "ElastiCache Valkey". Then either extend the _knowledge _when to include rediscloud and add an explicit 2C rule, or state that the node type is chosen in Design with a warning.

3. [Propagation & data] The Heroku Inference note contradicts Heroku's model cards — knowledge/design/fast-path-addons.json:83

"Map to Bedrock directly; model IDs carry over."

Heroku Inference IDs are Heroku-specific (claude-sonnet-5-5, nova-pro, gpt-oss-120b) and sit on an OpenAI-style /v1/chat/completions endpoint. Bedrock uses different model / inference-profile IDs and a different API, so copied INFERENCE_MODEL_ID values fail. Fix: say the IDs and API must be remapped, and route the code rewrite to the llm-to-bedrock skill.

4. [Completeness] Three new fields have no reader — discover-assemble.md:53, design-mapping.md:264. A sweep of the skill tree, tests/ and fixtures/ for apps_by_environment_tier, environment_tier and postgres_followers_collapsed finds only the producer and schema lines. (#609 adds readers for the first two; nothing reads postgres_followers_collapsed.) Fix: render them in the generate-report.md estate summary, or drop them.

5. [Propagation & data] No regression test for the headline fixes, and the edited test contradicts its own name — tests/test_heroku_postgres_advanced_pricing.py:216. test_pricing_cache_classic_tier_rows_untouched now asserts changed values. Fix: rename it to …_match_published_prices, and add a test asserting:

• the scheduler key and alias;
• every Classic cache plan has a sizing row with matching RAM / memory_limit;
• class GiB ≥ ram for both class columns. This last check would have caught Should-fix 1.

Nits

1. design-mapping.md:405 and kafka-msk-sizing.json:10 say "3 brokers" for standard/extended/private; add Shield, which the PR adds at kafka-msk-sizing.json:8.
2. design-mapping.md:486-487: the heroku-connect alias can't fire, because step 2 strips the heroku- prefix first. jawsdb-maria already normalizes without the alias.
3. redis-elasticache-sizing.json:264 and heroku-pricing-cache.md:106-107 carry revision history in runtime-read text, and the "~10x … two to three instance classes" figure is off. The actual ratios are 1.4–2.5× on premium and 10–33× on private.
4. PR body: premium-8 and premium-9 existed at base.

Heads-up (out of diff, pre-existing — not blocking)

1. Heroku Kafka Extended plans are "8 Kafka Brokers" with 400 / 800 / 2400 GB capacity (Elements). _broker_count_and_azs maps extended to 3 brokers, and the cache lists 1.5 / 3 / 6 TB. The new shield-extended-* rows inherit this.
2. Kafka basic-1 / basic-2 are priced in the cache but have no sizing row.
3. Redis legacy premium-8 (15 GB) maps to cache.m6g.xlarge, which has about 12.9 GiB (not re-fetched this session).

Alternatives (max 2)

1. Model a same-app follower as an RDS read replica on the leader instead of collapsing it. This preserves read capacity on the default RDS branch. Recommendation: do the guarded collapse now and the replica as a named follow-up.
2. Defer 768 GB+ plans on the Aurora path instead of adding an x2g rate. That would defer the biggest databases again, which undercuts the PR's goal. Recommendation: switch the class to x2g as in Should-fix 1.

Verdict

Comment. One Blocking finding (follower collapse) and five Should-fix items are open. The data additions themselves check out against Heroku and AWS primary sources.
Commenting rather than approving: author == reviewer identity.

Stale: findings addressed and all threads resolved; branch has since been updated with main (84393b4). Re-review requested.

[herosjourney]

All review findings are addressed and every thread is resolved. The branch is now merged up to current main (head 84393b4e), with conflicts resolved by keeping both sides. Full CI passed on the conflict-resolution commit. Requesting a fresh review of 84393b4e. The older CHANGES_REQUESTED review was against an earlier head and is dismissed as stale. Merge still needs one code-owner approval from someone who did not push.