[huangxiaobai-ydmy]

Description

Fixes #4856.

Standalone (non-HA) mooncake_master previously returned from main() after SIGTERM/SIGINT, so destructors walked metadata_shards_ object-by-object. At 12M–25M objects that took ~20–48 s after the RPC server had already stopped, delaying Kubernetes rollouts; the index is not reused across a restart after the legacy snapshot removal.

This change matches the HA shutdown path since #4745:

1. Install the shared shutdown signal handlers and watchdog on the non-HA path.
2. On a requested shutdown: stop the RPC server, admin server, RPC worker pool, and HTTP metadata server.
3. Call ExitAfterShutdown(0) so the object index is reclaimed by the kernel without a user-space walk.

Standalone has no etcd lease to revoke; the HTTP metadata server is stopped explicitly before _exit.

Suggested Merge Priority

• P0 — Urgent
• P1 — High
• P2 — Normal
• P3 — Low

Module

• Mooncake Store (mooncake-store)

Type of Change

• Bug fix
• New feature
• Refactor
• Breaking change
• Documentation update
• Performance improvement
• Other

How Has This Been Tested?

Test commands:

# Rely on CI Build & Test (Linux). Full multi-million-object SIGTERM timing
# needs the reporter's load setup from #4856 and was not re-run locally
# (Windows host without Mooncake toolchain).

Test results:

• Unit tests pass
• Integration tests pass (if applicable)
• Manual testing done (code review against HA ExitAfterShutdown path; destructor audit for non-HA external effects)

Checklist

• I have performed a self-review of my own code
• I have formatted my code using ./scripts/code_format.sh
• I have run pre-commit on the files changed in this PR and all hooks pass
• I have updated the documentation (if applicable)
• I have added tests to prove my changes are effective
• For changes >500 LOC: I have filed an RFC issue

AI Assistance Disclosure

• No AI tools were used
• AI tools were used (specify below)

Cursor AI assisted with issue analysis against #4856 / HA #4745, implementing the non-HA ExitAfterShutdown path, and PR drafting. The submitter reviewed the changes.

[huangxiaobai-ydmy]

Previous build-wheel failure was an apt Failed to update package lists flake (not packaging). Pushed an empty commit to retrigger CI.

[staryxchen]

Thanks for picking up the test-wheel-ubuntu failure as well.
The _structured_ranges.py staging belongs in a dedicated wheel PR so this one can stay scoped to the standalone-master SIGTERM path (#4856). We opened that as #5030: it only adds the module to MIGRATED_PYTHON_MODULES and the packaging asserts, without touching master.cpp or CUDA EP JIT packaging.
The current scripts/build_wheel.sh diff on this branch also drops CUDA_EP_JIT_DIR staging relative to main (#4581). After #5030 merges, please rebase onto main and drop the wheel/packaging commits (_structured_ranges staging plus the packaging-test edits) so this PR only contains the mooncake-store/src/master.cpp shutdown change.

[huangxiaobai-ydmy]

Thanks @staryxchen — agreed. Will wait for #5030 to land, then rebase onto main and drop the wheel/packaging commits so this PR only keeps the master.cpp shutdown change.

[staryxchen]

#5030 is now on main (3fbd075). Please rebase this branch onto current main and drop the wheel/packaging commits (scripts/build_wheel.sh, python/tests/packaging/test_project.py, python/tests/packaging/test_installed_wheel.py) so the remaining diff is only the standalone-master SIGTERM change in mooncake-store/src/master.cpp.

[huangxiaobai-ydmy]

Thanks @staryxchen — rebased onto main (3fbd075) and dropped the wheel/packaging commits. The remaining diff is only mooncake-store/src/master.cpp.

[huangxiaobai-ydmy]

Friendly Monday ping for review when you have a moment — already rebased onto main after #5030 and trimmed to only the master.cpp shutdown change; CI is green.

@ykwd @stmatengss @YiXR @Aionw @XucSh — thanks!