Skip to content

The agent that grows with you

chore(sessions): meet code-health on the round-3 browse catch-up

Tekniumf9ba374d13eb
diff --git a/hermes_state_reconcile.py b/hermes_state_reconcile.py
index 4beac40299a..80f90b6e634 100644
--- a/hermes_state_reconcile.py
+++ b/hermes_state_reconcile.py
@@ -176,7 +176,7 @@ def _lockless_gateway_holder(path: Path) -> str | None:
try:
import psutil
argv = psutil.Process(pid).cmdline() or None
- except Exception:
+ except (ImportError, OSError, psutil.Error): # gone, denied, or no psutil: not provably ours
argv = None
if argv and _looks_like_hermes(argv) and 'gateway' in argv[1:]:
return describe_holder_pid(pid)
@@ -365,6 +365,32 @@ def _reserve_for_ddl(home: Path, path: Path, owned):
yield schema_drift_reason(path) is None
+def _verify_clean_before_lock(path: Path, too_slow):
+ """The pre-lock phase of the browse catch-up: a full read-only integrity scan OUTSIDE gateway.lock
+ (it may be slow on a large store and must never block a gateway boot). Returns the file signature
+ the scan vouches for. Raises :class:`CorruptStoreError` for a damaged store (latched, so polls do
+ not rescan it) and ``too_slow()`` when the scan cannot finish in its budget (remembered for the
+ same signature, so a polling sidebar does not rescan a multi-GB file every few seconds)."""
+ from hermes_state_health import STORAGE_CORRUPT, mark_storage_corrupt, storage_corrupt_reason, storage_state
+ if storage_state(path) == STORAGE_CORRUPT:
+ raise CorruptStoreError(storage_corrupt_reason(path) or 'latched corrupt')
+ key = str(path.resolve())
+ signature = _store_signature(path)
+ if _scan_too_slow.get(key) == signature:
+ raise too_slow()
+ try:
+ problem = _integrity_problem(path, _Budget(BROWSE_INTEGRITY_BUDGET_S))
+ except BrowseBudgetExceeded as exc:
+ _scan_too_slow[key] = signature
+ raise too_slow() from exc
+ if problem:
+ # A failing column probe hides damage elsewhere in the file: additive DDL is not
+ # corruption recovery, so never write into a store that is not verified clean.
+ mark_storage_corrupt(path, problem)
+ raise CorruptStoreError(problem)
+ return signature
+
+
def reconcile_drift_for_browse(path: Path) -> bool:
"""Catch a lagging store up for a read-only browse; True when DDL ran, False when already current.
@@ -373,7 +399,6 @@ def reconcile_drift_for_browse(path: Path) -> bool:
when the lagging store is also damaged. Never touches a store missing its history tables:
recreating them empty would hide lost sessions."""
from gateway.runtime_ownership import OwnershipConflict, live_runtime_owner
- from hermes_state_health import STORAGE_CORRUPT, mark_storage_corrupt, storage_corrupt_reason, storage_state
reason = schema_drift_reason(path)
if reason is None:
@@ -419,27 +444,10 @@ def reconcile_drift_for_browse(path: Path) -> bool:
raise owned(reason) from exc
if holder := _lockless_gateway_holder(path):
raise owned(reason, holder)
- # Already latched corrupt in this process: do not rescan a damaged file on every poll.
- if storage_state(path) == STORAGE_CORRUPT:
- raise CorruptStoreError(storage_corrupt_reason(path) or 'latched corrupt')
- key = str(path.resolve())
with _path_scan_lock(path): # one scan per store per process; concurrent polls wait for it
if schema_drift_reason(path) is None:
return False
- signature = _store_signature(path)
- if _scan_too_slow.get(key) == signature:
- raise too_slow()
- # Outside gateway.lock: read-only, may be slow on a large store, never blocks a gateway boot.
- try:
- problem = _integrity_problem(path, _Budget(BROWSE_INTEGRITY_BUDGET_S))
- except BrowseBudgetExceeded as exc:
- _scan_too_slow[key] = signature
- raise too_slow() from exc
- if problem:
- # A failing column probe hides damage elsewhere in the file: additive DDL is not
- # corruption recovery, so never write into a store that is not verified clean.
- mark_storage_corrupt(path, problem)
- raise CorruptStoreError(problem)
+ signature = _verify_clean_before_lock(path, too_slow)
try:
with _reserve_for_ddl(home, path, owned) as current:
if current:
diff --git a/tests/hermes_cli/test_browse_schema_drift.py b/tests/hermes_cli/test_browse_schema_drift.py
index 6a2c5244a67..36ab943f04a 100644
--- a/tests/hermes_cli/test_browse_schema_drift.py
+++ b/tests/hermes_cli/test_browse_schema_drift.py
@@ -740,7 +740,7 @@ def test_gateway_boot_never_waits_on_a_slow_integrity_scan(tmp_path):
browse = subprocess.Popen([sys.executable, "-c", _SLOW_SCAN_BROWSE, str(path), "4"], env=os.environ,
stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
assert browse.stdout.readline().strip() == "started"
- time.sleep(1.5) # inside the 4 s scan
+ time.sleep(1.5) # health: allow HX013 -- the child scan is a fixed 4 s stand-in; boot must land mid-scan
booting = ProfileOwnership()
try:
started = time.monotonic()
@@ -770,7 +770,7 @@ class _SlowProgress:
import time
def slow():
- time.sleep(0.01)
+ time.sleep(0.01) # health: allow HX013 -- the slowed VM IS the behaviour under test (cold disk)
return handler() if handler is not None else 0
return self._conn.set_progress_handler(slow, 5)