hermes-admin/hermes-agent
1
0
Fork 0
Code Issues Pull Requests Actions 28 Packages Projects Releases Wiki Activity

refactor(state): compact hermes_state_common comments and docstrings, keeping every invariant and failure mode

Browse Source
This commit is contained in:
Teknium
2026-09-02 22:56:57 -07:00
parent 7f24f2f987
commit c4a6c8fbc6
1 changed files with 90 additions and 122 deletions
Download Patch File Download Diff File Expand all files Collapse all files

212
hermes_state_common.py
View File

@@ -1,8 +1,5 @@
"""Shared constants and helpers for the SessionDB family of modules.
Lives outside hermes_state so the mixin modules can import it without a cycle;
hermes_state re-exports every name for backward compatibility.
"""
"""Shared constants and helpers for the SessionDB family of modules. Lives outside hermes_state so
the mixin modules can import it without a cycle; hermes_state re-exports every name."""
import contextlib
import errno
@@ -18,19 +15,17 @@ from agent.context_compressor import (LEGACY_SUMMARY_PREFIX, SUMMARY_PREFIX, _ME
_MERGED_SUMMARY_DELIMITER, _SUMMARY_END_MARKER)
# Session preview = head of the first user message, shown wherever a session has no title. A /skill
# invocation embeds the whole skill body, so its plain head would preview the SKILL's prose; scaffolded rows
# carry a wider excerpt (whole message under budget, else head + tail where the typed instruction lands) so
# ``_shape_preview`` can recover ``/work — fix the title leak``.
# Session preview = head of the first user message (shown when a session has no title). A /skill invocation
# embeds the whole skill body, so scaffolded rows carry a wider excerpt (whole message under budget, else
# head + tail where the typed instruction lands) so ``_shape_preview`` can recover ``/work — fix ...``.
_PREVIEW_HEAD_CHARS = 63
_PREVIEW_SCAFFOLD_WINDOW = 400
_PREVIEW_MAX_CHARS = 60
def escape_like(text: str) -> str:
"""Escape LIKE wildcards (``%``, ``_``) so derived text matches literally; pair with
``ESCAPE '\\'``. ``_`` is common in branch names, titles and paths, and a documented
substring/prefix match must not silently widen."""
"""Escape LIKE wildcards (``%``, ``_``) so derived text matches literally; pair with ``ESCAPE '\\'``.
``_`` is common in branch names/titles/paths and a substring match must not silently widen."""
return text.replace("\\", "\\\\").replace("%", "\\%").replace("_", "\\_")
@@ -61,8 +56,8 @@ def _sql_after_marker(marker: str) -> str:
return f"SUBSTR(m.content, INSTR(m.content, {_sql_literal(marker)}) + {len(marker)})"
# Current and legacy long-form prefixes share this whole introduction; matching all of it keeps an ordinary
# message that merely starts with the bracketed label from counting as a compaction carrier.
# Current and legacy long-form prefixes share this introduction; matching all of it keeps an ordinary message
# that merely starts with the bracketed label from counting as a compaction carrier.
_PREVIEW_LONG_FORM_PREFIX = SUMMARY_PREFIX.split("Do NOT answer", 1)[0]
_PREVIEW_SUMMARY_PREFIXES = (_PREVIEW_LONG_FORM_PREFIX, LEGACY_SUMMARY_PREFIX)
_PREVIEW_STANDALONE_SUMMARY_SQL = _sql_starts_with("m.content", _PREVIEW_SUMMARY_PREFIXES)
@@ -78,16 +73,14 @@ _PREVIEW_MERGED_PRIOR_UNWRAPPED_SQL = (f"CASE WHEN SUBSTR({_PREVIEW_MERGED_PRIOR
f" ELSE {_PREVIEW_MERGED_PRIOR_SQL} END")
_PREVIEW_FORCE_USER_REMAINDER_SQL = _sql_after_marker(_SUMMARY_END_MARKER)
# Pure compaction rows are ineligible for previews; force-user-leading and merged
# carriers are eligible only when authentic content survives.
# Pure compaction rows are ineligible; force-user-leading and merged carriers only when authentic content survives.
_PREVIEW_ELIGIBLE_SQL = (f"((NOT {_PREVIEW_STANDALONE_SUMMARY_SQL} AND NOT {_PREVIEW_MERGED_SUMMARY_SQL})"
f" OR ({_PREVIEW_STANDALONE_SUMMARY_SQL} AND INSTR(m.content, {_sql_literal(_SUMMARY_END_MARKER)}) > 0"
f" AND LENGTH({_sql_trim_whitespace(_PREVIEW_FORCE_USER_REMAINDER_SQL)}) > 0)"
f" OR ({_PREVIEW_MERGED_SUMMARY_SQL}"
f" AND LENGTH({_sql_trim_whitespace(_PREVIEW_MERGED_PRIOR_UNWRAPPED_SQL)}) > 0))")
# Shared ``_preview_raw`` SELECT expression for every listing query (scaffolded rows:
# head + tail spliced around SKILL_EXCERPT_JOINT when over budget).
# ``_preview_raw`` SELECT for every listing query (scaffolded rows: head + tail around SKILL_EXCERPT_JOINT).
_PREVIEW_RAW_SELECT = (
f"CASE WHEN {_PREVIEW_STANDALONE_SUMMARY_SQL} THEN {_PREVIEW_FORCE_USER_REMAINDER_SQL}"
f" WHEN {_PREVIEW_MERGED_SUMMARY_SQL} THEN {_PREVIEW_MERGED_PRIOR_UNWRAPPED_SQL}"
@@ -116,53 +109,48 @@ _PREVIEW_RAW_SUBQUERY_SQL = (f"COALESCE((SELECT {_PREVIEW_RAW_SELECT} FROM messa
# ── Session lineage predicates ({a} = sessions alias) ───────────────────────
# A /branch child (kept visible, never cascade-deleted): stable marker OR the legacy end_reason heuristic.
# /branch child (kept visible, never cascade-deleted): stable marker OR legacy end_reason heuristic.
_BRANCH_CHILD_SQL = ("json_extract(COALESCE({a}.model_config, '{{}}'), '$._branched_from') IS NOT NULL"
" OR EXISTS (SELECT 1 FROM sessions p WHERE p.id = {a}.parent_session_id"
" AND p.end_reason = 'branched' AND {a}.started_at >= p.ended_at)")
_COMPRESSION_CHILD_SQL = ("EXISTS (SELECT 1 FROM sessions p WHERE p.id = {a}.parent_session_id"
" AND p.end_reason = 'compression')")
# 'session_switch': switch_session() creates no child row, but pre-marker DBs hold legacy reset children
# whose parent later ended that way. Must stay identical to the recovery fence in
# find_latest_gateway_session_for_peer (interpolates the SQL form).
# 'session_switch' creates no child row today, but pre-marker DBs hold legacy reset children whose parent
# ended that way. Must stay identical to the recovery fence in find_latest_gateway_session_for_peer.
_RESET_END_REASONS = ("session_reset", "session_switch", "idle", "daily", "suspended", "resume_pending_expired")
_RESET_END_REASONS_SQL = ", ".join(f"'{reason}'" for reason in _RESET_END_REASONS)
# Accidental end reasons recovery treats as resumable (docs/session-lifecycle.md). Single source of truth:
# interpolated into recovery SQL AND exposed as SessionDB.RECOVERABLE_END_REASONS.
# superseded_by_resume: stale sentinel-parked runtime superseded by a fresh session.resume.
# startup_orphan_reap: startup sweep of rows orphaned by a dead gateway process (same accident class as
# ws_orphan_reap, kept distinct for forensics).
# Accidental end reasons recovery treats as resumable (docs/session-lifecycle.md); single source of truth for
# recovery SQL and SessionDB.RECOVERABLE_END_REASONS. superseded_by_resume = sentinel-parked runtime replaced
# by a fresh session.resume; startup_orphan_reap = dead-gateway sweep, same class as ws_orphan_reap but kept
# distinct for forensics.
_RECOVERABLE_END_REASONS = ("agent_close", "ws_orphan_reap", "superseded_by_resume", "startup_orphan_reap")
_RECOVERABLE_END_REASONS_SQL = ", ".join(f"'{reason}'" for reason in _RECOVERABLE_END_REASONS)
# End reasons written by AUTOMATIC cleanup (shutdown, orphan reapers, idle/LRU eviction), not by a
# deliberate conversation boundary: "some runtime went away", NOT "this conversation ended", so a writer
# that can prove liveness (e.g. a compression rotation holding the lease) may clear it. Superset of the
# recoverable set plus the TUI gateway's automatic reasons.
# End reasons written by AUTOMATIC cleanup (shutdown, orphan reapers, idle/LRU eviction), not a deliberate
# conversation boundary: "some runtime went away", so a writer that can prove liveness (e.g. a compression
# rotation holding the lease) may clear it. Recoverable set plus the TUI gateway's automatic reasons.
_AUTOMATIC_END_REASONS = frozenset(_RECOVERABLE_END_REASONS) | {
"tui_shutdown", "ws_disconnect", "idle_timeout", "lru_evict"}
def is_automatic_end_reason(reason) -> bool:
"""True when *reason* is an automatic-cleanup end stamp. Single owner of the
accidental-vs-deliberate predicate; compression-liveness sites must call this."""
"""True when *reason* is an automatic-cleanup end stamp; compression-liveness sites must call this."""
return isinstance(reason, str) and reason in _AUTOMATIC_END_REASONS
def _legacy_reset_child_sql(alias: str, reasons_sql: str) -> str:
"""Pre-marker reset-continuation heuristic: child rides its parent's exact non-empty routing key and the
parent ended at a reset boundary. Shared by ``_RESET_CHILD_SQL`` and ``reopen_session()``'s
marker-stamping UPDATE so the two cannot drift; ``reasons_sql`` is a literal or placeholder list."""
parent ended at a reset boundary. Shared by ``_RESET_CHILD_SQL`` and ``reopen_session()`` so the two
cannot drift; ``reasons_sql`` is a literal or placeholder list."""
return (f"EXISTS (SELECT 1 FROM sessions p WHERE p.id = {alias}.parent_session_id"
f" AND p.end_reason IN ({reasons_sql}) AND {alias}.session_key IS NOT NULL"
f" AND {alias}.session_key != '' AND {alias}.session_key = p.session_key)")
# A reset starts a separate user-visible conversation though rows keep parent_session_id
# for lineage. Stable marker, or the same-key fallback for pre-marker rows (the
# exact-key requirement keeps subagent children out).
# A reset starts a separate user-visible conversation though rows keep parent_session_id for lineage.
# Stable marker, or the same-key fallback for pre-marker rows (exact key keeps subagent children out).
_RESET_CHILD_SQL = ("json_extract(COALESCE({a}.model_config, '{{}}'), '$._reset_from') IS NOT NULL"
" OR " + _legacy_reset_child_sql("{a}", _RESET_END_REASONS_SQL))
@@ -178,9 +166,8 @@ def _ephemeral_child_sql(alias: str = "s") -> str:
def _sql_freshest_of(activity: str, session_id_expr: str, started: str) -> str:
"""Freshest of *activity* and the latest message timestamp for *session_id_expr*,
else *started*. Heartbeats are rate-limited (~60s) so ``last_activity_at`` can lag
a newer message; never prefer it alone."""
"""Freshest of *activity* and the latest message timestamp for *session_id_expr*, else *started*.
Heartbeats are rate-limited (~60s) so ``last_activity_at`` can lag a newer message; never use it alone."""
msg_max = f"(SELECT MAX(_act_m.timestamp) FROM messages _act_m WHERE _act_m.session_id = {session_id_expr})"
return (f"COALESCE((SELECT MAX(_act_v.v) FROM (SELECT {activity} AS v UNION ALL SELECT {msg_max}) _act_v), "
f"{started})")
@@ -200,14 +187,12 @@ def _sql_session_last_active_by_id(session_id_expr: str) -> str:
SCHEMA_VERSION = 28
# Auto-maintenance VACUUMs only when at least this fraction of pages is on the freelist;
# below it a full rewrite costs more I/O than it returns.
# Auto-maintenance VACUUMs only above this freelist fraction; below it a rewrite costs more I/O than it returns.
AUTO_VACUUM_MIN_FREELIST_RATIO = 0.25
# FTS storage layout, tracked INDEPENDENTLY of SCHEMA_VERSION (state_meta ``fts_storage_version``): schema
# version advances freely on open, the FTS layout only changes when a DB is born fresh or explicitly
# optimized via ``hermes sessions optimize-storage``. Legacy DBs sit at 0 (marker absent) with a working
# inline index; 1 = v23 external-content layout.
# FTS layout, tracked INDEPENDENTLY of SCHEMA_VERSION (state_meta ``fts_storage_version``): it changes only
# when a DB is born fresh or via ``hermes sessions optimize-storage``. 0 (marker absent) = legacy inline
# index, still working; 1 = v23 external-content layout.
FTS_STORAGE_VERSION = 1
# Cap on user-controlled FTS5 query input before sanitizer processing.
@@ -215,8 +200,8 @@ MAX_FTS5_QUERY_CHARS = 2_048
def stat_db_file_identity(path) -> "tuple[int, int] | None":
"""``(st_dev, st_ino)`` for *path*, or None. st_ino=0 (Windows, some network FS)
would false-positive every replaced-file check, so it counts as unknown."""
"""``(st_dev, st_ino)`` for *path*, or None. st_ino=0 (Windows, some network FS) would false-positive
every replaced-file check, so it counts as unknown."""
try:
st = os.stat(path)
except OSError:
@@ -484,8 +469,7 @@ CREATE INDEX IF NOT EXISTS idx_async_delegations_delivery
ON async_delegations(delivery_state, completed_at);
"""
# Indexes on columns added in later schema versions must run AFTER
# _reconcile_columns() adds them, or executescript fails on legacy DBs.
# Indexes on later-added columns must run AFTER _reconcile_columns(), or executescript fails on legacy DBs.
DEFERRED_INDEX_SQL = """
CREATE INDEX IF NOT EXISTS idx_messages_session_active
ON messages(session_id, active, timestamp);
@@ -501,13 +485,11 @@ CREATE INDEX IF NOT EXISTS idx_sessions_system_prompt_hash
ON sessions(system_prompt_hash);
"""
# ── Deferred FTS rebuild bookkeeping ──
# While a background rebuild is pending, two state_meta keys define which rows are IN the FTS indexes: H =
# fts_rebuild_high_water (MAX(messages.id) when the old indexes were dropped), P = fts_rebuild_progress
# (highest backfilled id). A row is indexed iff id <= P OR id > H (AUTOINCREMENT ids: post-drop rows are
# indexed live by the insert triggers); rows in (P, H] are not. Every trigger gates on that predicate: an
# external-content 'delete' for a row NOT in the index corrupts it, and skipping one for an indexed row
# leaves a stale entry. With no rebuild pending both keys are absent and COALESCE makes it a tautology.
# Deferred FTS rebuild bookkeeping: while a background rebuild is pending, state_meta H = fts_rebuild_high_water
# (MAX(messages.id) when the old indexes were dropped) and P = fts_rebuild_progress (highest backfilled id)
# define the indexed rows: id <= P OR id > H (post-drop rows are indexed live by the insert triggers); (P, H]
# is not. Every trigger gates on that predicate: an external-content 'delete' for an unindexed row corrupts
# the index and skipping one for an indexed row leaves a stale entry. No rebuild pending => COALESCE tautology.
FTS_SQL = """
CREATE VIRTUAL TABLE IF NOT EXISTS messages_fts USING fts5(
content,
@@ -557,11 +539,10 @@ BEGIN
END;
"""
# Trigram FTS5 table for CJK substring search (unicode61 splits CJK into single tokens, breaking phrase
# matching). The trigram index is ~2.6x the text it covers and ``role='tool'`` rows are ~90% of message
# bytes of machine noise, so it reads through the ``messages_fts_trigram_src`` view, which excludes tool
# rows; those remain searchable via ``messages_fts``, and ``search_messages`` routes CJK queries filtered on
# role='tool' to LIKE.
# Trigram FTS5 table for CJK substring search (unicode61 splits CJK into single tokens). The trigram index is
# ~2.6x the text it covers and ``role='tool'`` rows are ~90% of message bytes, so it reads through the
# ``messages_fts_trigram_src`` view excluding tool rows; those stay searchable via ``messages_fts`` and
# ``search_messages`` routes CJK queries filtered on role='tool' to LIKE.
FTS_TRIGRAM_SQL = """
CREATE VIEW IF NOT EXISTS messages_fts_trigram_src AS
SELECT id, role, content, tool_name, tool_calls
@@ -621,23 +602,20 @@ END;
_FTS_CJK_TRIGGERS = ("messages_fts_cjk_insert", "messages_fts_cjk_delete", "messages_fts_cjk_update")
# Set when a tokenizer-less process dropped the cjk triggers to keep writes alive: the cjk index is missing
# rows and must not serve reads until `hermes sessions optimize-storage` rebuilds it on a capable host.
# Set when a tokenizer-less process dropped the cjk triggers to keep writes alive: the cjk index is missing rows
# and must not serve reads until `hermes sessions optimize-storage` rebuilds it on a capable host.
FTS_CJK_STALE_KEY = "fts_cjk_stale"
# Set when a base/trigram FTS index was detached after runtime corruption.
# While present, startup must rebuild the complete index before reinstalling
# sync triggers: rows written while they were absent leave an unknown gap.
# Set when a base/trigram FTS index was detached after runtime corruption. While present, startup must rebuild
# the complete index before reinstalling sync triggers: rows written while they were absent leave an unknown gap.
FTS_STALE_KEY = "fts_stale"
# Durable diagnostic for stale FTS recovery blocked across process restarts.
FTS_REBUILD_DEFERRAL_KEY = "fts_rebuild_deferral"
# ── Legacy (v22 / inline-content) FTS DDL ──────────────────────────────
# Used ONLY to keep a pre-v23 install's search working and its triggers repairable until
# `optimize_fts_storage()` migrates it: inline copies of content || tool_name || tool_calls, trigram over
# every row. Never created on a fresh install. Handing a legacy DB the v23 DDL would create the
# external-content trigram VIEW and leave it in a mixed, broken state.
# Legacy (v22 / inline-content) FTS DDL: ONLY keeps a pre-v23 install's search working and its triggers
# repairable until `optimize_fts_storage()` migrates it (inline content || tool_name || tool_calls, trigram over
# every row). Never created fresh; the v23 DDL on a legacy DB would leave a mixed, broken state.
LEGACY_FTS_SQL = """
CREATE VIRTUAL TABLE IF NOT EXISTS messages_fts USING fts5(
content
@@ -691,19 +669,16 @@ AFTER UPDATE OF content, tool_name, tool_calls ON messages BEGIN
END;
"""
# ── Cross-process full-FTS-rebuild admission (single authority) ──────────────
# Several Hermes processes share one state.db; a full structural FTS rebuild (FTS5 'rebuild' or the
# drop/recreate in `_recover_stale_fts`) must run in ONE of them at a time — concurrent rebuilds
# structurally corrupted state.db in production. Single authority for `rebuild_fts()`,
# `_rebuild_fts_indexes()` and `_recover_stale_fts()`; the chunked backfill (`fts_rebuild_step`) is
# deliberately NOT routed through it (it claims progress under SQLite transaction authority and is
# multi-process). Semantics mirror `hermes_state._cross_process_repair_lock`: portable (msvcrt on Windows,
# flock elsewhere), bounded wait, FAIL CLOSED. flock rides the open file description, so a forked child
# that inherited the fd holds it forever after the holder dies; the holder's pid + start time are recorded
# under the lock and a provably-dead holder's lock is broken by unlinking and retaking on a fresh inode.
# Indeterminate liveness still defers. `<db>.fts_rebuild.lock` is distinct from `<db>.repair.lock` (schema
# surgery on an EXCLUSIVE offline connection, minutes in VACUUM). Lives here because mixins cannot import
# hermes_state (cycle).
# Cross-process full-FTS-rebuild admission (single authority). Several processes share one state.db and a
# structural rebuild (FTS5 'rebuild' or `_recover_stale_fts`'s drop/recreate) must run in ONE at a time —
# concurrent rebuilds corrupted state.db in production. Gates `rebuild_fts()`, `_rebuild_fts_indexes()`,
# `_recover_stale_fts()`; the chunked backfill (`fts_rebuild_step`) is deliberately NOT routed through it (it
# claims progress under SQLite transaction authority). Mirrors `hermes_state._cross_process_repair_lock`:
# portable (msvcrt/flock), bounded wait, FAIL CLOSED. flock rides the open file description, so a forked
# child holds it forever after the holder dies; holder pid + start time are recorded under the lock and a
# provably-dead holder's lock is broken by unlinking and retaking on a fresh inode. Indeterminate liveness
# defers. `<db>.fts_rebuild.lock` is distinct from `<db>.repair.lock` (offline schema surgery, minutes in
# VACUUM). Lives here because mixins cannot import hermes_state (cycle).
logger = logging.getLogger("hermes_state")
@@ -711,25 +686,22 @@ _FTS_REBUILD_LOCK_TIMEOUT_SECONDS = 120.0
_FTS_REBUILD_LOCK_POLL_SECONDS = 0.1
_IS_WINDOWS = sys.platform == "win32"
# Post-break re-acquire budget: the fresh inode is contended only by live
# processes, so a short wait suffices — never re-enter the full timeout.
# Post-break re-acquire budget: the fresh inode is contended only by live processes — never the full timeout.
_LOCK_BREAK_REACQUIRE_SECONDS = 5.0
# "Another process holds the lock": flock → EWOULDBLOCK/EAGAIN, msvcrt.locking → EACCES (EDEADLK when its
# retry gives up). Anything else (ESTALE, ENOTSUP, ENOLCK, EIO) is a persistent environment failure that
# polling cannot fix; treating it as contention burned the full timeout on every attempt.
# "Another process holds the lock": flock → EWOULDBLOCK/EAGAIN, msvcrt.locking → EACCES (EDEADLK when its retry
# gives up). Anything else (ESTALE, ENOTSUP, ENOLCK, EIO) is a persistent failure polling cannot fix.
_LOCK_CONTENTION_ERRNOS = {errno.EAGAIN, errno.EACCES, errno.EWOULDBLOCK, errno.EDEADLK}
def is_advisory_lock_contention(exc: BaseException) -> bool:
"""True when *exc* means another process holds the advisory lock. For any
other ``OSError`` callers must fail closed IMMEDIATELY: retrying cannot succeed."""
"""True when *exc* means another process holds the lock; on any other ``OSError`` fail closed at once."""
return isinstance(exc, BlockingIOError) or (isinstance(exc, OSError) and exc.errno in _LOCK_CONTENTION_ERRNOS)
def _proc_start_ticks(pid: int):
"""Kernel start time of *pid* (field 22 of ``/proc/<pid>/stat``), which with the PID uniquely identifies
a process; None off Linux or on any failure — callers must treat None as unknowable and FAIL CLOSED."""
"""Kernel start time of *pid* (field 22 of ``/proc/<pid>/stat``; with the PID it identifies a process
uniquely). None off Linux or on any failure — callers must treat None as unknowable and FAIL CLOSED."""
try:
with open(f"/proc/{pid}/stat", "rb") as fh:
stat = fh.read()
@@ -761,22 +733,20 @@ def _rewrite_lock_file(handle, payload: bytes) -> None:
def _write_lock_holder_record(handle) -> None:
"""Record this process as holder (best effort), written under the flock so timed-out
contenders can tell an orphaned-fd holder from a live wedged one."""
"""Record this process as holder (best effort) so timed-out contenders can tell an orphaned-fd holder
from a live wedged one."""
record = {"pid": os.getpid(), "start_ticks": _proc_start_ticks(os.getpid()), "acquired_at": time.time()}
_rewrite_lock_file(handle, json.dumps(record, sort_keys=True).encode("utf-8"))
def _clear_lock_holder_record(handle) -> None:
"""Erase holder metadata before a normal release, so a surviving record always
means an ABNORMAL exit — the only condition allowing a break."""
"""Erase holder metadata before a normal release: a surviving record means ABNORMAL exit (break allowed)."""
_rewrite_lock_file(handle, b"")
def _lock_holder_provably_dead(record) -> bool:
"""True ONLY when the recorded holder is provably dead or PID-recycled. Anything
indeterminate (no/malformed record, PID owned by another user, /proc unavailable)
is False — the caller must FAIL CLOSED and defer."""
"""True ONLY when the recorded holder is provably dead or PID-recycled. Anything indeterminate
(no/malformed record, PID owned by another user, /proc unavailable) is False: FAIL CLOSED and defer."""
if not isinstance(record, dict):
return False
try:
@@ -800,15 +770,14 @@ def _lock_holder_provably_dead(record) -> bool:
def _acquire_db_flock(lock_path, handle, timeout_seconds, poll_seconds, description):
"""Bounded POSIX flock acquire with orphaned-holder staleness break. Returns ``(acquired, handle)``;
*handle* may have been re-opened and the caller closes whichever comes back. *acquired* is True, False
(a holder kept the lock past the deadline), or None (non-contention ``OSError``, already logged; callers
treat it as not acquired without the held-by-another-process warning). ``flock`` belongs to the open
file DESCRIPTION, which ``fork()`` duplicates: a holder that forks then dies leaves the lock held forever
by a child that never releases. When the process that ACQUIRED is provably dead yet the flock is held,
the file is unlinked and retaken on a fresh inode; the orphan's flock stays on the old inode blocking
nobody. Every successful acquire verifies its inode still names *lock_path*, so a racer that locked a
dead inode retries instead of running alongside the breaker. Indeterminate liveness defers."""
"""Bounded POSIX flock acquire with orphaned-holder break. Returns ``(acquired, handle)``; *handle* may
have been re-opened and the caller closes whichever comes back. *acquired*: True, False (a holder kept
the lock past the deadline) or None (non-contention ``OSError``, already logged; callers treat it as not
acquired without the held-by-another-process warning). ``flock`` belongs to the open file DESCRIPTION,
which ``fork()`` duplicates, so a holder that forks then dies leaves the lock held forever; when the
acquirer is provably dead the file is unlinked and retaken on a fresh inode (the orphan's flock stays on
the old inode blocking nobody). Every successful acquire verifies its inode still names *lock_path*,
so a racer that locked a dead inode retries instead of running alongside the breaker."""
import fcntl
deadline = time.monotonic() + timeout_seconds
broke_lock = False
@@ -843,8 +812,7 @@ def _acquire_db_flock(lock_path, handle, timeout_seconds, poll_seconds, descript
broke_lock = True
deadline = time.monotonic() + _LOCK_BREAK_REACQUIRE_SECONDS
continue
# Verify the path still names our inode: a breaker may have replaced
# the file while we waited, and a lock on a dead inode excludes nobody.
# A breaker may have replaced the file while we waited; a lock on a dead inode excludes nobody.
try:
fd_stat, path_stat = os.fstat(handle.fileno()), os.stat(lock_path)
same_file = fd_stat.st_dev == path_stat.st_dev and fd_stat.st_ino == path_stat.st_ino
@@ -894,10 +862,11 @@ def _acquire_msvcrt_lock(lock_path, handle, timeout):
@contextlib.contextmanager
def fts_rebuild_admission(db_path, *, timeout_seconds=None):
"""Serialize full structural FTS rebuilds on *db_path* across processes. Yields True when this process holds the authority, False when the bounded acquire timed out or the lock
file could not be opened. On False the caller must NOT rebuild (fail closed); the stale breadcrumb
guarantees a retry. ``db_path`` None (in-memory DB) yields True. Opportunistic in-process retries pass
``timeout_seconds=0`` so a live holder never stalls a long-lived writer; the orphan break still applies."""
"""Serialize full structural FTS rebuilds on *db_path* across processes. Yields True when this process
holds the authority, False when the bounded acquire timed out or the lock file could not be opened; on
False the caller must NOT rebuild (fail closed; the stale breadcrumb guarantees a retry). ``db_path``
None (in-memory DB) yields True. Opportunistic in-process retries pass ``timeout_seconds=0`` so a live
holder never stalls a long-lived writer; the orphan break still applies."""
if db_path is None:
yield True
return
@@ -906,11 +875,10 @@ def fts_rebuild_admission(db_path, *, timeout_seconds=None):
try:
handle = open(lock_path, "a+b")
except OSError as exc:
# Fail closed like a timed-out acquire: an unopenable lock file means the FS is
# out of space/inodes/descriptors, and a sibling that opened its handle earlier
# may still be rebuilding — yielding True gave every process on a full disk a
# concurrent rebuild of the same DB. Deferring costs nothing (the breadcrumb
# retries, and the rebuild's own writes could not have committed either).
# Fail closed like a timed-out acquire: an unopenable lock file means the FS is out of
# space/inodes/descriptors and a sibling that opened earlier may still be rebuilding — yielding True
# gave every process on a full disk a concurrent rebuild. Deferring costs nothing (the breadcrumb
# retries; the rebuild's own writes could not have committed either).
logger.warning("Could not open FTS rebuild lock %s (%s) — deferring this rebuild "
"rather than running it without cross-process authority.", lock_path, exc)
yield False