feat: persist and rehydrate scheduling state across root restarts#64

Merged

kaiitunnz merged 31 commits into

mainfrom

feat/rolling-node-restart

Jun 13, 2026

Commits on Jun 11, 2026

feat(sdk): add drain_workers() to stack node client

kaiitunnz
and
claude
committed
feat(cli): recreate a single stack service in place via `stack restart [SERVICE]`

kaiitunnz
and
claude
committed
feat(server): persist scheduler state and rehydrate it on restart

kaiitunnz
and
claude
committed
feat(server): make task events replayable across a restart

kaiitunnz
and
claude
committed
feat(server): rehydrate dispatch state during lifespan startup

kaiitunnz
and
claude
committed
docs+test: document rolling updates and cover restart/rehydration

kaiitunnz
and
claude
committed
fix(server): stop flushing Redis on shutdown so restarts keep state

kaiitunnz
and
claude
committed
fix(server): guard terminal tasks against replayed dispatch/start/update
kaiitunnz
committed
fix(server): advance task-event cursor per entry and document replay contract
kaiitunnz
committed
fix(server): give rehydrated in-flight tasks a heartbeat grace after restart
kaiitunnz
committed
fix(server): detect task-event stream trim on consumer resume
kaiitunnz
committed
feat(worker): add echo delay_sec knob to hold a task in flight
kaiitunnz
committed
fix(server): keep the task-event consumer alive when a handler raises
kaiitunnz
committed
refactor(server): address review nits
kaiitunnz
committed
docs: note the no-worker grace restarts on a root restart
kaiitunnz
committed
refactor(server): rework durable scheduler-state persistence
kaiitunnz
committed
refactor(server): tidy the task-event consumer and watchdog
kaiitunnz
committed
feat(cli): restart multiple compose services in one call
kaiitunnz
committed
refactor(sdk): drop the redundant NodeClient.drain_workers
kaiitunnz
committed
chore(worker): remove the echo delay_sec knob
kaiitunnz
committed
docs: reframe rolling-update docs as generic service restarts
kaiitunnz
committed
refactor(cli): drop rolling-specific framing from the restart docstring
kaiitunnz
committed
docs: note SSH in-flight work is cancelled, not requeued, on drain
kaiitunnz
committed
chore: ignore new torch and vllm pip-audit advisories
kaiitunnz
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat: persist and rehydrate scheduling state across root restarts#64

feat: persist and rehydrate scheduling state across root restarts#64
kaiitunnz merged 31 commits into
mainfrom
feat/rolling-node-restart

Commits on Jun 11, 2026

feat(sdk): add drain_workers() to stack node client

feat(cli): recreate a single stack service in place via `stack restart [SERVICE]`

feat(server): persist scheduler state and rehydrate it on restart

feat(server): make task events replayable across a restart

feat(server): rehydrate dispatch state during lifespan startup

docs+test: document rolling updates and cover restart/rehydration

fix(server): stop flushing Redis on shutdown so restarts keep state

fix(server): guard terminal tasks against replayed dispatch/start/update

fix(server): advance task-event cursor per entry and document replay contract

fix(server): give rehydrated in-flight tasks a heartbeat grace after restart

fix(server): detect task-event stream trim on consumer resume

feat(worker): add echo delay_sec knob to hold a task in flight

fix(server): keep the task-event consumer alive when a handler raises

refactor(server): address review nits

docs: note the no-worker grace restarts on a root restart

refactor(server): rework durable scheduler-state persistence

refactor(server): tidy the task-event consumer and watchdog

feat(cli): restart multiple compose services in one call

refactor(sdk): drop the redundant NodeClient.drain_workers

chore(worker): remove the echo delay_sec knob

docs: reframe rolling-update docs as generic service restarts

refactor(cli): drop rolling-specific framing from the restart docstring

docs: note SSH in-flight work is cancelled, not requeued, on drain

chore: ignore new torch and vllm pip-audit advisories

Commits on Jun 12, 2026

fix(server): guard terminal task transitions against replay

fix(server): recover the task-event consumer from handler failures

test: tidy review nits in cli/sdk/server tests

fix(server): persist terminal transitions last, heal on replay

Commits on Jun 13, 2026

fix(server): persist cancelled tasks atomically in cancel_workflow

perf(server): batch terminal-set writes per workflow and status

perf(server): batch rehydrate task-state loads with mget