doc: add doc best-practice/deploy-mooncake-store-with-rbg

[diw-zw]

No description provided.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[cheyang]

/copilot review

[cheyang]

@gemini-code-assist /review

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[cheyang]

Review: doc: add doc best-practice/deploy-mooncake-store-with-rbg

Overall this is a well-structured documentation set covering Mooncake Store deployment with RBG. However, the markdown-check CI is currently failing with 16 lint errors that need to be resolved before merge.

Blocking Issues

1. MD024 (duplicate headings) -- Both guide docs reuse identical heading text ("Expected Behavior", "Verification") across multiple operation sections. Qualify each heading with its parent section context to make them unique.

2. MD051 (invalid link fragments) -- The "Related Documentation" sections in both concept docs use (#filename.md) syntax instead of (filename.md). The # prefix makes markdownlint interpret these as in-page anchors.

Non-blocking Issues

3. Broken cross-reference -- The guide docs link to ../10. 通过 RBG 部署 Mooncake Store 实现 KV Cache 复用.md which does not exist. The concept doc lives at ./10-deploy-mooncake-store-with-rbg.md.

4. Inconsistent replica count -- Guide Operation 1 states "3 replicas x 10 GB = 30 GB" but the YAML has replicas: 1 and MOONCAKE_GLOBAL_SEGMENT_SIZE: "100gb", matching the verification output of 100 GB.

Please fix the CI failures and the broken link, then this should be good to go.

[cheyang]

This text says "Total L3 cache capacity = 3 replicas x 10 GB = 30 GB", but the YAML above specifies replicas: 1 and MOONCAKE_GLOBAL_SEGMENT_SIZE: "100gb". The verification output below also confirms a 100 GB pool. Please reconcile this -- either update the YAML to 3 replicas/10gb, or update this text to match the actual 1-replica/100gb configuration.

[cheyang]

This cross-reference links to ../10. 通过 RBG 部署 Mooncake Store 实现 KV Cache 复用.md which does not correspond to any file in the repository. The concept document appears to be at ./10-deploy-mooncake-store-with-rbg.md (same directory). Please update this link. Same issue in the ZH guide.

[cheyang]

/copilot review

[cheyang]

@gemini-code-assist /review

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[cheyang]

The cross-reference ../10.%20通过%20RBG%20部署%20Mooncake%20Store%20实现%20KV%20Cache%20复用.md does not match the actual concept doc filename 10-deploy-mooncake-store-with-rbg.md. Since both files are in the same en/ directory, this should be ./10-deploy-mooncake-store-with-rbg.md.

[cheyang]

Item 4 states "Total L3 cache capacity = 3 replicas x 10 GB = 30 GB" but the YAML above deploys replicas: 1 with MOONCAKE_GLOBAL_SEGMENT_SIZE: "100gb". The actual verification output correctly shows 100.00 GB. Consider updating this line to match the actual deployment config.

[cheyang]

The inference service snippet uses lmsysorg/sglang:v0.5.5 while the Mooncake Store YAML and the prerequisites note both reference v0.5.9. Consider aligning to a single version for clarity.

[cheyang]

Reading this from a user's perspective, several things will trip people up:

Two things that will likely break kubectl apply:

1. YAML anchors: the manifests use image: &image lmsysorg/sglang:v0.5.9 and image: *image inside a cat <<'EOF' | kubectl apply -f - heredoc. Anchors are resolved by the YAML parser, and the client-side handling in kubectl for streamed heredocs is inconsistent — some versions serialize the resolved value, some pass the raw ref. Safer to inline the image value everywhere so users don't hit a "field required" error.
2. The Operation 4 port-forward uses 8000:8000, but Operation 3 already set up a port-forward on 8000. A user running the guide top-to-bottom will hit "address already in use".

The Summary table promises operations that don't exist:

The Summary section lists "Benchmark" and "Lossless Update" as operations, but there are no corresponding step-by-step sections in the doc. Users will scroll looking for these and be confused. Either add the operations or remove them from the summary.

Data doesn't match the config:

In the concept doc's "Expected Behavior": "Total L3 cache capacity = 3 replicas × 10 GB = 30 GB". But the actual YAML has replicas: 1 for the store and MOONCAKE_GLOBAL_SEGMENT_SIZE=100gb. The numbers should either match the manifests, or the manifests need to be updated to match the narrative.

Hardware assumption:

All examples assume RDMA (rdma/hca: "1", MOONCAKE_PROTOCOL=rdma, hostNetwork: true, privileged: true). Most users evaluating this will start on a plain TCP cluster. Please add a complete TCP-mode variant (or at least one full working manifest) so users without RDMA infrastructure can try it.

Minor:

• The reference at the top of the guide points to ../10. 通过 RBG 部署 Mooncake Store 实现 KV Cache 复用.md — Chinese characters and spaces in file paths are fragile across doc renderers. The actual file appears to be 10-deploy-mooncake-store-with-rbg.md.