merge: testing → main (reconcile 2-week divergence)

This commit is contained in:
2026-04-28 18:36:00 -04:00
parent 499836c9e4
commit 862e4dbb31
1235 changed files with 160255 additions and 7996 deletions

View File

@@ -14,22 +14,44 @@ from decnet.fleet import all_service_names
from decnet.composer import write_compose
from decnet.config import DeckyConfig, DecnetConfig
from decnet.engine import _compose_with_retry
from decnet.logging import get_logger
from decnet.telemetry import traced as _traced
from pathlib import Path
import anyio
import asyncio
import contextlib
from decnet.bus import topics as _topics
from decnet.bus.base import BaseBus
from decnet.bus.factory import get_bus
from decnet.bus.publish import (
publish_safely as _publish_safely,
run_control_listener_signal as _run_control_listener_signal,
run_health_heartbeat as _run_health_heartbeat,
)
from decnet.mutator.events import MutationTrigger, emit_decky_mutated
from decnet.web.db.repository import BaseRepository
log = get_logger("mutator")
console = Console()
async def mutate_decky(decky_name: str, repo: BaseRepository) -> bool:
@_traced("mutator.mutate_decky")
async def mutate_decky(
decky_name: str,
repo: BaseRepository,
bus: BaseBus | None = None,
trigger: MutationTrigger = "operator",
) -> bool:
"""
Perform an Intra-Archetype Shuffle for a specific decky.
Returns True if mutation succeeded, False otherwise.
"""
log.debug("mutate_decky: start decky=%s", decky_name)
state_dict = await repo.get_state("deployment")
if state_dict is None:
log.error("mutate_decky: no active deployment found in database")
console.print("[red]No active deployment found in database.[/]")
return False
@@ -54,6 +76,7 @@ async def mutate_decky(decky_name: str, repo: BaseRepository) -> bool:
console.print(f"[yellow]No services available for mutating '{decky_name}'.[/]")
return False
old_services = list(decky.services)
current_services = set(decky.services)
attempts = 0
@@ -73,58 +96,376 @@ async def mutate_decky(decky_name: str, repo: BaseRepository) -> bool:
# Still writes files for Docker to use
write_compose(config, compose_path)
log.info("mutation applied decky=%s services=%s", decky_name, ",".join(decky.services))
console.print(f"[cyan]Mutating '{decky_name}' to services: {', '.join(decky.services)}[/]")
try:
# Wrap blocking call in thread
await anyio.to_thread.run_sync(_compose_with_retry, "up", "-d", "--remove-orphans", compose_path)
except Exception as e:
log.error("mutation failed decky=%s error=%s", decky_name, e)
console.print(f"[red]Failed to mutate '{decky_name}': {e}[/]")
return False
await emit_decky_mutated(
bus,
decky=decky_name,
old_services=old_services,
new_services=list(decky.services),
trigger=trigger,
)
return True
async def mutate_all(repo: BaseRepository, force: bool = False) -> None:
"""
Check all deckies and mutate those that are due.
If force=True, mutates all deckies regardless of schedule.
@_traced("mutator.mutate_all")
async def mutate_all(
repo: BaseRepository,
force: bool = False,
bus: BaseBus | None = None,
only: set[str] | None = None,
) -> float | None:
"""Mutate all deckies that are due (or *only* the named ones).
Returns the number of seconds until the next scheduled mutation, or
``None`` if no deployment exists / no decky has an interval set. The
watch loop uses this to adaptively sleep instead of hard-polling at a
fixed cadence.
A missing ``deployment`` state row is *not* an error any more — the
host may simply not have run ``decnet deploy`` yet. The watch loop
edge-triggers the user-facing log for that state.
"""
log.debug("mutate_all: start force=%s only=%s", force, only)
state_dict = await repo.get_state("deployment")
if state_dict is None:
console.print("[red]No active deployment found.[/]")
return
log.debug("mutate_all: no active deployment found")
return None
config = DecnetConfig(**state_dict["config"])
now = time.time()
# Trigger derivation: explicit force / targeted only-list come from
# an operator action (CLI --all, API mutate-now, UI bus request).
# Scheduled-interval ticks carry trigger=scheduled.
trigger: MutationTrigger = "operator" if (force or only is not None) else "scheduled"
mutated_count = 0
next_due_in: float | None = None
for decky in config.deckies:
if only is not None and decky.name not in only:
continue
interval_mins = decky.mutate_interval or config.mutate_interval
if interval_mins is None and not force:
continue
if force:
if force or only is not None:
due = True
else:
elapsed_secs = now - decky.last_mutated
due = elapsed_secs >= (interval_mins * 60)
remaining = (interval_mins * 60) - elapsed_secs
if not due and (next_due_in is None or remaining < next_due_in):
next_due_in = remaining
if due:
success = await mutate_decky(decky.name, repo=repo)
success = await mutate_decky(
decky.name, repo=repo, bus=bus, trigger=trigger,
)
if success:
mutated_count += 1
if mutated_count == 0 and not force:
console.print("[dim]No deckies are due for mutation.[/]")
if mutated_count:
log.info("mutate_all: complete mutated_count=%d", mutated_count)
else:
log.debug("mutate_all: no deckies due for mutation")
return next_due_in
@_traced("mutator.reconcile_topologies")
async def reconcile_topologies(
repo: BaseRepository, bus: BaseBus | None = None,
) -> int:
"""Drain pending ``topology_mutations`` rows against live topologies.
For every topology in ``active|degraded`` with at least one pending
mutation, atomically claim the oldest via
:meth:`BaseRepository.claim_next_mutation`, dispatch to the matching
``apply_<op>`` in :mod:`decnet.mutator.ops`, and write the outcome
back (``applied`` or ``failed``).
On ``MutationError`` the topology is flipped to ``degraded`` — the
same state the future Healer will target — so operators can see that
a requested change was rejected without the repo drifting into an
inconsistent state.
Returns the number of mutations drained this tick.
"""
# Local imports keep the flat-fleet hot path free of MazeNET cost.
from decnet.mutator.ops import MutationError, dispatch as _op_dispatch
from decnet.topology.persistence import transition_status
from decnet.topology.status import TopologyStatus, TopologyStatusError
drained = 0
for tid in await repo.list_live_topology_ids():
while True:
mut = await repo.claim_next_mutation(tid)
if mut is None:
break # no more work for this topology this tick.
await _publish_safely(
bus,
_topics.topology_mutation(tid, _topics.MUTATION_APPLYING),
{"mutation_id": mut["id"], "op": mut["op"], "payload": mut["payload"]},
event_type=_topics.MUTATION_APPLYING,
)
try:
await _op_dispatch(repo, tid, mut["op"], mut["payload"])
await repo.mark_mutation_applied(mut["id"])
drained += 1
log.info(
"topology %s mutation %s applied op=%s",
tid, mut["id"], mut["op"],
)
await _publish_safely(
bus,
_topics.topology_mutation(tid, _topics.MUTATION_APPLIED),
{"mutation_id": mut["id"], "op": mut["op"]},
event_type=_topics.MUTATION_APPLIED,
)
except (MutationError, Exception) as exc: # noqa: BLE001
reason = f"{type(exc).__name__}: {exc}"
await repo.mark_mutation_failed(mut["id"], reason)
log.warning(
"topology %s mutation %s failed: %s",
tid, mut["id"], reason,
)
await _publish_safely(
bus,
_topics.topology_mutation(tid, _topics.MUTATION_FAILED),
{"mutation_id": mut["id"], "op": mut["op"], "reason": reason},
event_type=_topics.MUTATION_FAILED,
)
try:
await transition_status(
repo, tid, TopologyStatus.DEGRADED, reason=reason,
)
await _publish_safely(
bus,
_topics.topology_status(tid),
{"state": TopologyStatus.DEGRADED, "reason": reason},
event_type=_topics.TOPOLOGY_STATUS,
)
except TopologyStatusError:
# Already degraded / in a state that can't degrade
# further — leave as is.
pass
# Stop draining this topology on first failure so the
# operator can inspect before a cascade.
break
return drained
@_traced("mutator.reconcile_agent_resyncs")
async def reconcile_agent_resyncs(repo: BaseRepository) -> int:
"""Re-push agent-targeted topologies flagged by the heartbeat handler.
The heartbeat sets ``needs_resync=True`` when an agent's reported
applied_version_hash diverges from master's expectation. Here we
re-run the agent branch of ``deploy_topology`` which pushes the
current hydrated blob back down over mTLS and clears the flag on
success. Any push failure leaves the flag set so the next tick
retries — it also logs loudly so ops can see that a specific agent
is stuck.
"""
from decnet.engine import deployer as _deployer
try:
pending = await repo.list_topologies_needing_resync()
except NotImplementedError:
return 0
drained = 0
for topo in pending:
tid = topo["id"]
try:
await _deployer.resync_agent_topology(repo, tid)
await repo.set_topology_resync(tid, False)
drained += 1
log.info("topology %s resynced to agent %s",
tid, topo.get("target_host_uuid"))
except Exception as exc: # noqa: BLE001
log.warning(
"topology %s resync failed (will retry): %s", tid, exc,
)
return drained
@_traced("mutator.watch_loop")
async def run_watch_loop(repo: BaseRepository, poll_interval_secs: int = 10) -> None:
"""Run an infinite loop checking for deckies that need mutation."""
"""Run an infinite loop checking for deckies that need mutation.
Two independent responsibilities, in strict order per tick:
1. Flat-fleet service rotation (``mutate_all``) — runs every tick
regardless of MazeNET state, preserving phase-1 timing.
2. MazeNET live-mutation reconciliation — runs only when the cheap
guard ``has_pending_topology_mutation`` (indexed composite
lookup) returns True. Zero-topology and idle-topology hosts pay
exactly one indexed query per tick.
"""
log.info("mutator watch loop started poll_interval_secs=%d", poll_interval_secs)
console.print(f"[green]DECNET Mutator Watcher started (polling every {poll_interval_secs}s).[/]")
# Connect to the bus for publish + wake-on-enqueue. Failure here is
# non-fatal: a mutator without a bus still works, it just runs at
# poll-interval latency and doesn't push notifications to UI clients.
bus: BaseBus | None = None
wake = asyncio.Event()
mutate_requests: set[str] = set()
wake_tasks: list[asyncio.Task] = []
heartbeat_task: asyncio.Task | None = None
try:
candidate = get_bus(client_name="mutator")
await candidate.connect()
bus = candidate
wake_tasks.append(asyncio.create_task(_wake_on_enqueue(bus, wake)))
wake_tasks.append(asyncio.create_task(
_wake_on_mutate_request(bus, wake, mutate_requests),
))
heartbeat_task = asyncio.create_task(
_run_health_heartbeat(bus, "mutator"),
)
# Control listener: SIGTERM-based so the existing shutdown path
# (cancel wake_tasks + heartbeat_task) runs unchanged.
wake_tasks.append(asyncio.create_task(
_run_control_listener_signal(bus, "mutator"),
))
except Exception as exc: # noqa: BLE001
log.warning("mutator: bus unavailable, running in poll-only mode: %s", exc)
# Edge-triggered "no deployment" state so we don't spam the console
# every 10 seconds on a host that hasn't deployed yet. Start as None
# so the first observation fires exactly one line.
deployment_present: bool | None = None
try:
while True:
await mutate_all(force=False, repo=repo)
await asyncio.sleep(poll_interval_secs)
requested = mutate_requests.copy()
mutate_requests.clear()
next_due = await mutate_all(
repo=repo,
force=False,
bus=bus,
only=requested or None,
)
has_deployment = (
next_due is not None or await repo.get_state("deployment") is not None
)
if has_deployment and deployment_present is not True:
log.info("mutator: active deployment observed — entering normal cadence")
console.print("[green]Active deployment observed.[/]")
deployment_present = True
elif not has_deployment and deployment_present is not False:
log.info("mutator: no active deployment — idling until one lands")
console.print("[dim]No active deployment; mutator idling.[/]")
deployment_present = False
# Gate reconciler on the O(log n) guard query — avoids
# entering the dispatch body when there's nothing to do.
try:
if await repo.has_pending_topology_mutation():
await reconcile_topologies(repo, bus=bus)
except NotImplementedError:
# Backend without MazeNET support — nothing to reconcile.
pass
try:
await reconcile_agent_resyncs(repo)
except NotImplementedError:
pass
except Exception:
log.exception("reconcile_agent_resyncs tick raised")
# Adaptive sleep: wake at the earlier of (next decky due) or
# (poll_interval_secs), bounded below by 1s so a thrashing
# schedule can't spin the loop. A bus wake (enqueue or
# mutate_request) short-circuits the wait.
if next_due is None or next_due > poll_interval_secs:
timeout = float(poll_interval_secs)
else:
timeout = max(1.0, next_due)
try:
await asyncio.wait_for(wake.wait(), timeout=timeout)
except asyncio.TimeoutError:
pass
wake.clear()
except KeyboardInterrupt:
log.info("mutator watch loop stopped")
console.print("\n[dim]Mutator watcher stopped.[/]")
finally:
for t in wake_tasks:
t.cancel()
if heartbeat_task is not None:
heartbeat_task.cancel()
for t in (*wake_tasks, heartbeat_task):
if t is None:
continue
with contextlib.suppress(asyncio.CancelledError, Exception):
await t
if bus is not None:
with contextlib.suppress(Exception):
await bus.close()
async def _wake_on_enqueue(bus: BaseBus, wake: asyncio.Event) -> None:
"""Flip *wake* every time a ``mutation.enqueued`` event lands.
Subscribes to the wildcard ``topology.*.mutation.enqueued`` — a single
subscription covers every topology on the host. Runs until cancelled
or the bus closes (NullBus yields nothing and returns immediately,
which is fine: the poll-interval fallback still ticks).
"""
pattern = f"{_topics.TOPOLOGY}.*.mutation.{_topics.MUTATION_ENQUEUED}"
try:
sub = bus.subscribe(pattern)
async with sub:
async for _event in sub:
wake.set()
except asyncio.CancelledError:
raise
except Exception as exc: # noqa: BLE001
log.warning("mutator: wake subscriber died (%s); falling back to poll", exc)
async def _wake_on_mutate_request(
bus: BaseBus,
wake: asyncio.Event,
pending: set[str],
) -> None:
"""Collect on-demand ``decky.<name>.mutate_request`` events.
API/CLI/UI callers publish to ``decky.{name}.mutate_request`` to force
an immediate mutation without waiting for the scheduled interval. We
stash the target decky name in *pending* so the next tick can feed it
to ``mutate_all(only=...)``, then flip *wake* to short-circuit the
sleep. Payload is optional — the topic's second token is the name.
"""
pattern = f"{_topics.DECKY}.*.{_topics.DECKY_MUTATE_REQUEST}"
try:
sub = bus.subscribe(pattern)
async with sub:
async for event in sub:
topic = getattr(event, "topic", "") or ""
parts = topic.split(".")
name = parts[1] if len(parts) >= 3 else ""
payload = getattr(event, "payload", None) or {}
if not name and isinstance(payload, dict):
name = payload.get("name", "") or ""
if name:
pending.add(name)
wake.set()
except asyncio.CancelledError:
raise
except Exception as exc: # noqa: BLE001
log.warning(
"mutator: mutate_request subscriber died (%s); falling back to poll",
exc,
)