Feat/ha phase3 - #9

Merged

milk333445 merged 16 commits into

mainfrom

Jun 30, 2026

Commits on Jun 28, 2026

feat(ha): publish live instance addresses to the shared store (Phase 3a)

milk333445
and
claude
committed
f02a833
fix(ha): persist overlay to DB before router reload so rollback/import sticks

milk333445
and
claude
committed
5b85bd2
feat(ha): node registry + assignments foundation (Phase 3b-1)

milk333445
and
claude
committed
622950b
feat(ha): gate actuation on node ownership (Phase 3b-2)

milk333445
and
claude
committed
309366f
feat(ha): backfill full observed state to the shared store (Phase 3d-1)

milk333445
and
claude
committed
6fe1c52
feat(ha): follower reads the fleet from the shared store (Phase 3d-2)

milk333445
and
claude
committed
8515895
feat(ha): cross-node placement scheduler (Phase 3c)

milk333445
and
claude
committed
f476df4
feat(ha): router /ready reports routable instances (Phase 3e)

milk333445
and
claude
committed
ee1a11c
docs: mark HA Phase 3a-3e done; note the one remaining netns step

milk333445
and
claude
committed
b9ef064

Commits on Jun 29, 2026

feat(router): make gunicorn worker count configurable (ROUTER_WORKERS)

milk333445
and
claude
committed
ea7b731
feat(ha): optional routable vLLM bind for multi-node inference (Phase 3b split)

milk333445
and
claude
committed
1585183
docs: mark vLLM routable-bind (netns decoupling) done; only multi-GPU host + 3f remain

milk333445
and
claude
committed
69ebe19
feat(deploy): Helm chart for k8s (HA Phase 3f)

milk333445
and
claude
committed
85a25a4
docs: per-node actuation design (the final multi-GPU-node step)

milk333445
and
claude
committed
f5810b3
feat(deploy): omit GPU limit when backend.gpu=0 (CPU/dev installs)

milk333445
and
claude
committed
569d56e
fix(store): serialize Postgres schema init with an advisory lock

milk333445
and
claude
committed
8b90909