Skip to content

Diskless implementation presets: inkless-3az, kafscale-3b, automq-3az and DB/ET codes (v2.3) - #8

Merged
mmatloka merged 1 commit into
mainfrom
ai/diskless-impl-presets
Nov 13, 2026
Merged

mmatloka merged 1 commit into
mainfrom
ai/diskless-impl-presets

Conversation

@mmatloka

Copy link
Copy Markdown
Contributor

Summary

The simulator's v2.3 free play has three diskless implementation presets, and the website's sim-link MCP docs (reference/mcp-tools.mdx, in a "from v2.3" block) already list them as build_playground_url cluster values, with DB:0/DB:1 for Inkless. sim-link refused all three with "not a free-play topology". This PR brings it to parity with the playground's own codec.

  • Shape: cluster now takes these three presets:

    • inkless-3az (default 6x3x3): orders classic, clicks diskless
    • kafscale-3b (default 3x3x1)
    • automq-3az (default 3x6x1)

    All three use the shared multi-DC caps of 9 brokers and 6 partitions. Each rule below mirrors freePlayTopologies.ts:

    • kafscale-3b and automq-3az: every topic is diskless at RF 1.
      • A topic with RF above 1 is refused.
      • A compacted topic is refused. The simulator does not model compaction on either preset.
      • An extra topic gets the ~1 field even when the caller did not set it, because the playground writes it.
    • kafscale-3b: disklessTiming.commitIntervalMs is refused, because KafScale has no commit interval.
    • The primary diskless flag d is never written on these presets, since the playground drops it.
    • inkless-3az: disklessEnable on orders is refused.
    • A caller topic named clicks on diskless-3az or inkless-3az is refused. Before, the playground silently renamed it.
  • actionlog: adds DB:0/DB:1 (stop and start the Inkless PostgreSQL) and ET:0/ET:1 (stop and start the KafScale etcd). A payload other than 0 or 1 is an error.

  • MCP actions: adds stop_coordinator_db, start_coordinator_db, stop_etcd and start_etcd. They are encoded into the URL and echoed back.

  • The AutoMQ zone router still has no shape input, as the docs already say.

Files changed

  • mcpserver/shape.go: the three topologies, impl/fixedTopics, and the RF, compaction, timing and primary-flag rules
  • mcpserver/actions.go, mcpserver/mcpserver.go: the four action kinds and the tool and schema descriptions
  • actionlog/actions.go, actionlog/decode.go, actionlog/actionlog.go: the DB and ET codes
  • mapping/testdata/ts-fixtures.txt: 16 SIZE-IMPL lines, made with decodeFreePlaySize/encodeFreePlaySize per preset from website main e9bf8e22
  • Tests: mcpserver/shape_test.go, mcpserver/mcpserver_test.go, actionlog/actionlog_test.go, actionlog/property_test.go
  • docs/mapping.md, README.md

Commands run

  • gofmt -l on the changed files: clean
  • go vet ./... and go vet -tags e2e ./e2e/: clean
  • go test ./...: all packages pass, including the new TestShapeImplMatchesTypeScriptFixtures, TestShapeImplPresets, TestShapeImplValidation, TestBuildURLDisklessControlPlanes and TestDisklessControlPlaneWire
  • e2e (Docker) and govulncheck were not run locally: no dependency changed, and the e2e path (connect) is untouched. CI runs both.

Known risks

  • Compaction on automq-3az is refused, although AutoMQ supports it. The playground would accept the slug, but the simulator's engine does not model compacted topics there, so refusing it is the conservative choice. The error message says that the gap is in the simulator.
  • clicks on diskless-3az is now refused. Before, the URL loaded with the topic renamed to clicks-2. This is a small behavior change for an existing preset.
  • On a playground deployment older than v2.3, these cluster values fall back to single-dc and the DB/ET codes are dropped. This is the same behavior as for other new codes.

Follow-up

  • The website docs (mcp-tools.mdx) mention only the DB actions. stop_etcd/start_etcd could be added there next to kafscale-3b.
  • A sim-link release before or with simulator v2.3, and the website version bump that goes with it.

🤖 Generated with Claude Code

… and the DB/ET control-plane codes (v2.3)

The simulator's v2.3 free play adds three diskless implementation presets,
and the website's sim-link docs already describe them as build_playground_url
cluster values. sim-link refused all three ("not a free-play topology").
This brings it to parity with the playground's own codec:

- mcpserver shape: cluster takes inkless-3az (6x3x3, orders classic, clicks
  diskless), kafscale-3b (3x3x1) and automq-3az (3x6x1), with the shared
  multi-DC caps. On kafscale-3b and automq-3az every topic is diskless at
  RF 1: an RF above 1 or a compacted topic is refused, and an extra topic
  gets the "~1" the playground writes for it. kafscale-3b refuses
  disklessTiming.commitIntervalMs (KafScale has no commit interval). The
  primary diskless flag is never written on these presets, and
  disklessEnable on inkless-3az's orders is refused, since the playground
  drops both. A caller topic named like a preset's own (clicks on
  diskless-3az and inkless-3az) is refused instead of silently renamed.
- actionlog: StopCoordinatorDB/StartCoordinatorDB (DB:0/DB:1, Inkless
  PostgreSQL) and StopEtcd/StartEtcd (ET:0/ET:1, KafScale etcd) encode and
  decode with the playground's wire form; any other payload is an error.
- mcpserver actions: build_playground_url accepts stop_coordinator_db,
  start_coordinator_db, stop_etcd and start_etcd and echoes them back.
- SIZE-IMPL fixtures dumped from the real TypeScript codec
  (decodeFreePlaySize/encodeFreePlaySize per preset) pin the slugs;
  docs/mapping.md and the README describe the presets and codes.

The AutoMQ zone router has no shape input, as the docs already say.
@mmatloka
mmatloka merged commit c1da09c into main Nov 13, 2026
10 checks passed
@mmatloka
mmatloka deleted the ai/diskless-impl-presets branch November 13, 2026 14:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant