trex/trex-core
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(slot,intent): slot strategy-state 저장소 (dcy-028) + sell retryable 플래그 (issue #27)

Browse Source 
BAT-042 (dcy-028 / issue #26 C-1/C-2/C-2b/C-3, REQ-079~082 + REQ-073/074 CHG):
- slot_strategy_state 별도 테이블 + StrategyStateStore(in-memory+PG) + occupancy_seq·version CAS + opaque UTF-8 data + 2단 size cap (REQ-079)
- trex.slots.strategy_state.set write RPC — reject reason 집합, 멱등 (REQ-080)
- Slot.occupancy_seq 필드 + inactive→pending_buy bump + →INACTIVE close 삭제 훅 → 점유 전환 ABA 구조적 차단 (REQ-081)
- slots.req 응답 행에 strategy_state 맵(batched 조립)·occupancy_seq additive (REQ-073 AC-7/REQ-074 AC-6)
- 마이그레이션 014(occupancy_seq)/015(slot_strategy_state), baseline 문서 정합 (REQ-082)

BAT-043 (issue #27 / REQ-007 CHG):
- rejected SellIntent 응답에 additive retryable: bool (core 소유 reason→retryable 매핑, 미인식 기본 true, bad_request* false) — 손절매 재시도 안전 경로 decoupling

검증: 487 passed, 47 skipped (신규 41: store CAS 6+ABA, occupancy bump/close 4경로, set RPC 9, slots read 5, retryable 8, PG 게이트), ruff clean.
REQ-079·080·081·082·073·074·007 → Implemented. issue #27 close. PG 경로 실 DB 검증은 release 위임.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
genichin
2026-06-15 19:27:57 +09:00
co-authored by Claude Opus 4.8
parent bfadd0d5c8
commit e0d4e71304
40 changed files with 1941 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
trex_core/adapters/messaging/rabbitmq.py
+6 -1
View File
@@ -41,6 +41,7 @@ from trex_core.adapters.messaging.topology import (
direct_routing_key,
intent_queue,
slots_queue,
strategy_state_set_queue,
)
from trex_core.application.messaging import (
EventHandler,
@@ -140,7 +141,11 @@ class RabbitMQTransport:
iq = intent_queue(instance_key)
sq = slots_queue(instance_key)
cq = control_queue(instance_key)
for q, channel_type in ((iq, "intent"), (sq, "slots"), (cq, "control")):
ssq = strategy_state_set_queue(instance_key) # dcy-028 / REQ-080
for q, channel_type in (
(iq, "intent"), (sq, "slots"), (cq, "control"),
(ssq, "strategy_state_set"),
):
ch.queue_declare(q, durable=True, arguments=dlx)
# Directed binding: trex.direct → {instance_key}.{channel_type}
ch.queue_bind(q, DIRECT_EXCHANGE,
trex_core/adapters/messaging/topology.py
+5
View File
@@ -14,6 +14,7 @@ from __future__ import annotations
# --- RPC queues (durable for intent; transient-ish for slots but kept
# durable for simplicity since the queue itself is shared) ---
SLOTS_REQ_QUEUE = "trex.slots.req" # REQ-073 / dcy-025: replaces snapshot.req
STRATEGY_STATE_SET_QUEUE = "trex.slots.strategy_state.set" # dcy-028 / REQ-080
INTENT_REQ_QUEUE = "trex.intent.req"
PAUSE_REQ_QUEUE = "trex.pause.req" # BAT-008 / REQ-032 (legacy alias of control)
CONTROL_REQ_QUEUE = "trex.control.req" # BAT-012 / REQ-043 operator control channel
@@ -44,6 +45,10 @@ def slots_queue(instance_key: str | None = DEFAULT_INSTANCE_KEY) -> str:
return _instance_queue(SLOTS_REQ_QUEUE, instance_key)
def strategy_state_set_queue(instance_key: str | None = DEFAULT_INSTANCE_KEY) -> str:
return _instance_queue(STRATEGY_STATE_SET_QUEUE, instance_key)
def control_queue(instance_key: str | None = DEFAULT_INSTANCE_KEY) -> str:
return _instance_queue(CONTROL_REQ_QUEUE, instance_key)
trex_core/adapters/persistence/postgresql.py
+121 -3
View File
@@ -24,6 +24,11 @@ from trex_core.application.ledger_outbox import (
)
from trex_core.application.ledger_port import LedgerTrade
from trex_core.application.ports import IntentRecord
from trex_core.application.strategy_state import (
SetResult,
StrategyStateEntry,
_cas,
)
from trex_core.domain.slot import Slot, SlotNotFoundError, SlotState, SlotTransitionRecord
_SLOT_COLUMNS = (
@@ -31,7 +36,7 @@ _SLOT_COLUMNS = (
"order_id, strategy_id, broker_order_id, portfolio_id, "
"sell_strategy_id, forced_close_at, stop_loss_price, take_profit_price, "
"created_at, updated_at, instance_key, retired_at, "
"timeout_sec, slippage_pct"
"timeout_sec, slippage_pct, occupancy_seq"
)
@@ -197,7 +202,7 @@ class PostgreSQLSlotRepository:
"broker_order_id=%s, portfolio_id=%s, sell_strategy_id=%s, "
"forced_close_at=%s, stop_loss_price=%s, take_profit_price=%s, "
"instance_key=%s, retired_at=%s, timeout_sec=%s, slippage_pct=%s, "
"updated_at=%s WHERE slot_id=%s",
"occupancy_seq=%s, updated_at=%s WHERE slot_id=%s",
(
slot.status.value, slot.symbol, slot.qty, slot.avg_price,
slot.limit_price, slot.order_id,
@@ -205,7 +210,7 @@ class PostgreSQLSlotRepository:
slot.sell_strategy_id, slot.forced_close_at,
slot.stop_loss_price, slot.take_profit_price,
slot.instance_key, slot.retired_at,
slot.timeout_sec, slot.slippage_pct,
slot.timeout_sec, slot.slippage_pct, slot.occupancy_seq,
slot.updated_at or datetime.now(tz=timezone.utc), slot.slot_id,
),
)
@@ -299,6 +304,7 @@ class PostgreSQLSlotRepository:
# forced_close_at(11) stop_loss_price(12) take_profit_price(13)
# created_at(14) updated_at(15) instance_key(16) retired_at(17)
# timeout_sec(18) slippage_pct(19) [dcy-022/REQ-068]
# occupancy_seq(20) [dcy-028/REQ-081]
return Slot(
slot_id=row[0] if isinstance(row[0], UUID) else UUID(str(row[0])),
status=SlotState(row[1]),
@@ -320,6 +326,7 @@ class PostgreSQLSlotRepository:
retired_at=row[17],
timeout_sec=int(row[18]) if row[18] is not None else None,
slippage_pct=f(row[19]),
occupancy_seq=int(row[20]) if len(row) > 20 and row[20] is not None else 0,
)
@@ -515,3 +522,114 @@ class PostgreSQLLedgerOutbox:
updated_at=row[15],
next_retry_at=row[16],
)
class PgStrategyStateStore:
"""slot_strategy_state-backed StrategyStateStore (dcy-028 / REQ-079).
Stores opaque per-strategy state keyed by (slot_id, strategy_id). The CAS
decision is the same `_cas` used by the in-memory double; occupancy_seq and
slot_active are supplied by the caller (fetched from the slot) so the store
only owns the slot_strategy_state table.
"""
def __init__(self, dsn: str, *, psycopg_module: Any = None) -> None:
self._psycopg = _require_psycopg(psycopg_module)
self._dsn = dsn
def set(
self,
slot_id: UUID,
strategy_id: str,
*,
expected_occupancy_seq: int,
expected_version: int,
data: str,
current_occupancy_seq: int,
slot_active: bool,
) -> SetResult:
with self._psycopg.connect(self._dsn, autocommit=False) as conn:
try:
with conn.cursor() as cur:
cur.execute(
"SELECT occupancy_seq, version, data, updated_at "
"FROM slot_strategy_state "
"WHERE slot_id = %s AND strategy_id = %s FOR UPDATE",
(slot_id, strategy_id),
)
row = cur.fetchone()
existing = (
StrategyStateEntry(
strategy_id=strategy_id,
occupancy_seq=int(row[0]),
version=int(row[1]),
data=row[2],
updated_at=row[3],
)
if row is not None
else None
)
result = _cas(
existing,
expected_occupancy_seq=expected_occupancy_seq,
expected_version=expected_version,
data=data,
current_occupancy_seq=current_occupancy_seq,
slot_active=slot_active,
)
if not result.ok:
conn.rollback()
return result
cur.execute(
"INSERT INTO slot_strategy_state "
"(slot_id, strategy_id, occupancy_seq, version, data, updated_at) "
"VALUES (%s, %s, %s, %s, %s, %s) "
"ON CONFLICT (slot_id, strategy_id) DO UPDATE SET "
"occupancy_seq = EXCLUDED.occupancy_seq, "
"version = EXCLUDED.version, data = EXCLUDED.data, "
"updated_at = EXCLUDED.updated_at",
(
slot_id, strategy_id, current_occupancy_seq,
result.version, data, datetime.now(tz=timezone.utc),
),
)
conn.commit()
return result
except BaseException:
conn.rollback()
raise
def get_for_slots(
self, slot_ids: list[UUID], current_occupancy_seq: dict[UUID, int]
) -> dict[UUID, dict[str, StrategyStateEntry]]:
if not slot_ids:
return {}
out: dict[UUID, dict[str, StrategyStateEntry]] = {}
with self._psycopg.connect(self._dsn) as conn:
with conn.cursor() as cur:
cur.execute(
"SELECT slot_id, strategy_id, occupancy_seq, version, data, updated_at "
"FROM slot_strategy_state WHERE slot_id = ANY(%s)",
(list(slot_ids),),
)
for row in cur.fetchall():
sid = row[0] if isinstance(row[0], UUID) else UUID(str(row[0]))
# stale-occupancy entries are invisible (dcy-028 FR-3 defense)
if int(row[2]) != current_occupancy_seq.get(sid):
continue
out.setdefault(sid, {})[row[1]] = StrategyStateEntry(
strategy_id=row[1],
occupancy_seq=int(row[2]),
version=int(row[3]),
data=row[4],
updated_at=row[5],
)
return out
def delete_all(self, slot_id: UUID) -> None:
with self._psycopg.connect(self._dsn, autocommit=False) as conn:
with conn.cursor() as cur:
cur.execute(
"DELETE FROM slot_strategy_state WHERE slot_id = %s", (slot_id,)
)
conn.commit()
trex_core/adapters/persistence/sql/014_add_slot_occupancy_seq.sql
+9
View File
@@ -0,0 +1,9 @@
-- BAT-042 / dcy-028 / REQ-081: per-occupancy guard counter on the slot.
--
-- The permanent asset slot is reused across occupancies; both a closed and a
-- freshly re-entered occupancy of the same slot start strategy_state at version
-- 0, so a version-only CAS cannot distinguish them at the first write. A
-- monotonic occupancy_seq (bumped on inactive→pending_buy) lets the
-- strategy-state write CAS reject a write computed against a prior occupancy
-- (occupancy_changed). Additive, idempotent. BIGINT so it never wraps.
ALTER TABLE daily_slot ADD COLUMN IF NOT EXISTS occupancy_seq BIGINT NOT NULL DEFAULT 0;
trex_core/adapters/persistence/sql/015_create_slot_strategy_state.sql
+18
View File
@@ -0,0 +1,18 @@
-- BAT-042 / dcy-028 / REQ-079: opaque per-strategy state store on slots.
--
-- A strategy daemon accumulates opaque state across a slot's occupancy (e.g. a
-- TSL peak). trex-core stores it server-side keyed by (slot_id, strategy_id),
-- never interpreting `data` (opaque UTF-8, byte-capped at the app layer). The
-- state's lifecycle matches the occupancy: trex-core deletes it on close
-- (→INACTIVE) so a reused slot starts empty (dcy-028 C-3). occupancy_seq mirrors
-- the slot's value at write time for the CAS occupancy guard (dcy-028 R-1).
-- version is the compare-and-set counter (idempotent / retry-safe writes).
CREATE TABLE IF NOT EXISTS slot_strategy_state (
slot_id UUID NOT NULL,
strategy_id TEXT NOT NULL,
occupancy_seq BIGINT NOT NULL,
version BIGINT NOT NULL,
data TEXT NOT NULL,
updated_at TIMESTAMPTZ NOT NULL DEFAULT (now() AT TIME ZONE 'UTC'),
PRIMARY KEY (slot_id, strategy_id)
);