[Bugfix][KV Offload] Bound unanswered P2P lookup waits (#56836)

Signed-off-by: jacklin78911-collab <jacklin78911@gmail.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: Itay Etelis <92247226+Etelis@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Or Ozeri <oro@il.ibm.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Liqian Lin
2026-10-15 08:23:34 +03:00
committed by GitHub
co-authored by Codex Robert Shaw Itay Etelis mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> Or Ozeri Kevin H. Luu
parent 81e1c3f183
commit 3f1bcfd82b
6 changed files with 386 additions and 24 deletions
@@ -32,6 +32,7 @@ logger = init_logger(__name__)
_LOAD_TIMEOUT_S = 30.0
_ABORT_ACK_TIMEOUT_S = 10.0
_LOOKUP_TIMEOUT_S = 30.0
@dataclass
@@ -59,10 +60,8 @@ class _ClientRequestState:
# -- Lookup phase (symmetric P2P only; untouched for PD) --
# Probe outcome per OffloadKey: None while in-flight (registered/sent
# but unresolved), True/False once a LookupRespMsg lands. There is no
# timeout — finish (via finish_request) is guaranteed after
# the request's lookup() calls and clears every probe, so an
# unanswered probe simply stays None until then.
# but unresolved), True/False once a LookupRespMsg lands. Unanswered
# probes expire the session so the manager can fall back to local compute.
probes: dict[OffloadKey, bool | None] = field(default_factory=dict)
# OffloadKeys registered but not yet flushed onto the wire. Drained and
# cleared by the next flush_pending_lookups.
@@ -136,6 +135,9 @@ class ClientRole:
# Kept in exact sync with ``st.loads``.
self._active_loads: set[str] = set()
self._completed_loads: list[LoadResult] = []
# Per-request progress deadlines. Only newly resolved probes extend
# them; additional keys and repeated scheduler polls do not.
self._lookup_deadlines: dict[str, float] = {}
# ------------------------------------------------------------------
# State helpers
@@ -172,6 +174,14 @@ class ClientRole:
"""True if any kv_request_id has a fetch in flight."""
return bool(self._active_loads)
def get_expired_lookup(self) -> tuple[str, float] | None:
"""Return an expired request and seconds without progress, if any."""
now = time.monotonic()
for req_id, deadline in self._lookup_deadlines.items():
if now >= deadline:
return req_id, now - deadline + _LOOKUP_TIMEOUT_S
return None
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
@@ -228,6 +238,7 @@ class ClientRole:
)
assert all(st.probes.get(key) is True for key in keys)
st.probes.clear()
self._lookup_deadlines.pop(kv_request_id, None)
def finish(self, kv_request_id: str) -> None:
"""Finish a request: abort in-flight loads and release lookup state.
@@ -272,6 +283,7 @@ class ClientRole:
)
st.probes.clear()
st.unsent.clear()
self._lookup_deadlines.pop(kv_request_id, None)
self._flush_pending.discard(kv_request_id)
self._maybe_prune(kv_request_id)
@@ -377,7 +389,7 @@ class ClientRole:
)
return None
def flush_pending_lookups(self) -> None:
def flush_pending_lookups(self, send_ready: bool) -> None:
"""Send a LookupMsg for each kv_request_id with unsent entries.
Called once per scheduler step from the manager's
@@ -387,13 +399,14 @@ class ClientRole:
keys, all tagged with the current round. register_lookup()
de-dups in-flight and already-resolved (req_id, key) pairs, so
each LookupMsg carries only the keys first probed in that step.
Send-gating is handled by the injected ``_send`` callback
(queues until ConnectAckMsg if needed).
Wait for ConnectAckMsg before sending or starting the response deadline.
Only requests that registered new keys since the last flush are
visited — the ``_flush_pending`` work-list avoids scanning every
live request each scheduler step.
"""
if not send_ready:
return
for req_id in self._flush_pending:
st = self._requests.get(req_id)
if st is None or not st.unsent:
@@ -401,6 +414,9 @@ class ClientRole:
# The peer now holds lookup state for this id; finish owes a
# terminal empty FetchMsg until a fetch closes it.
st.peer_lookup_open = True
self._lookup_deadlines.setdefault(
req_id, time.monotonic() + _LOOKUP_TIMEOUT_S
)
logger.debug(
"P2P LOOKUP client %s: SEND LookupMsg kv_request_id=%s keys=%d",
self._peer_id,
@@ -443,10 +459,16 @@ class ClientRole:
st = self._requests.get(kv_request_id)
if st is None:
return
progressed = False
for h, hit in zip(keys, hits):
key = OffloadKey(h)
if key in st.probes:
progressed |= st.probes[key] is None
st.probes[key] = hit
if all(hit is not None for hit in st.probes.values()):
self._lookup_deadlines.pop(kv_request_id, None)
elif progressed:
self._lookup_deadlines[kv_request_id] = time.monotonic() + _LOOKUP_TIMEOUT_S
def collect_results(self) -> list[LoadResult]:
"""Walk load timeouts and drain completed loads.
@@ -454,9 +476,6 @@ class ClientRole:
Loads past ``_LOAD_TIMEOUT_S`` get an AbortFetchMsg sent and
enter the aborting phase. Aborting loads past
``_ABORT_ACK_TIMEOUT_S`` are surfaced as failed.
Lookups have no timeout: an unanswered probe stays None (RETRY)
until finish_request clears it — see ``_ClientRequestState.probes``.
"""
now = time.monotonic()
to_remove: list[tuple[str, int]] = []
@@ -525,4 +544,5 @@ class ClientRole:
self._flush_pending.clear()
self._active_loads.clear()
self._completed_loads.clear()
self._lookup_deadlines.clear()
return ClientCloseResult(failed_jobs=failed_jobs, failed_req_ids=failed_req_ids)
@@ -755,10 +755,6 @@ class ServerRole:
"""
results: list[StoreResult] = self._timeout_pending_store_jobs()
if self._pending_store_results:
results.extend(self._pending_store_results)
self._pending_store_results.clear()
# Scope the poll to this peer: the transport is shared across all peer
# sessions of the engine, and poll() drains completed handles. An
# unscoped poll here would consume sibling sessions' completions and
@@ -833,6 +829,11 @@ class ServerRole:
self._finalize_cancelled(xfer.kv_request_id, xfer.round_key)
self._maybe_prune(xfer.kv_request_id)
# Finalizing transfers can queue results for supplied but unfetched blocks.
if self._pending_store_results:
results.extend(self._pending_store_results)
self._pending_store_results.clear()
return results
def collect_idle_timeouts(self) -> list[StoreResult]:
@@ -264,10 +264,9 @@ class P2PSession:
"""Flush any aggregated symmetric-P2P lookups for this peer.
Called once per scheduler step from the manager's
``on_schedule_end()``. Send-gating is handled inside the
client's ``_send`` callback (queues until ConnectAckMsg).
``on_schedule_end()``. Probes wait for ConnectAckMsg before sending.
"""
self._client.flush_pending_lookups()
self._client.flush_pending_lookups(send_ready=self._send_ready)
def serve_external_requests(self, parent: ParentManager) -> None:
"""Resolve inbound peer lookups against the tiering manager.
@@ -329,6 +328,21 @@ class P2PSession:
stores = self._server.collect_results()
self._server.drain_pending_aborts()
# Drain buffered replies and completed transfers before expiring a peer.
expired_lookup = self._client.get_expired_lookup()
if expired_lookup is not None:
req_id, elapsed = expired_lookup
logger.warning(
"P2PSession %s: lookup response timed out for kv_request_id=%s "
"after %.1fs without progress; disconnecting peer session",
self.peer_id,
req_id,
elapsed,
)
with contextlib.suppress(Exception):
self._conn.send({TYPE_KEY: DisconnectMsg.TYPE})
self._conn.mark_dead()
new_fetch_ids = self._new_fetch_ids
self._new_fetch_ids = []
return SessionPollResult(