Files
mcp-nextcloud/nextcloud_mcp_server/usage/store.py
T
Chris CoutinhoandClaude Opus 4.8 973f80e7b9 feat(usage): rename metrics → tokens_embedded/pages_embedded + export token cost to Prometheus
Billing product model finalized (Deck #281): bill pages externally, record
tokens internally. Rename the data-plane metric literals to match the now-
canonical contract (Deck #284) — the control plane's METRIC_EVENT_NAMES is
already renamed, so the old names would be unmapped and never sync to Stripe.

Rename (values unchanged):
- embeddings_queries → tokens_embedded (value = real token count, already
  emitted by this PR; the unit upstream providers bill on).
- pages_chunks → pages_embedded (value kept as len(chunk_texts) interim;
  TODO(#282): real normalized "pages indexed" count — real pages for paginated
  types, chars/tokens-per-page constant otherwise — is deferred to the
  instrumentation card, this only lands the name/contract).
- All literals, log strings, docstrings, comments, the migration comment, and
  tests renamed; grep confirms zero old strings remain.

Observability (new): export embedding token cost to Prometheus as
astrolabe_embedding_tokens_total{provider,operation} (operation = index|query)
so the billed cost unit is visible in Grafana, not just the per-tenant billing
DB. Dedicated counter (doesn't inflate the existing chunk/request metrics) and
always-on (independent of USAGE_METERING_ENABLED, so OSS/self-host gets it).
Wired on both the indexing batch embed and the search query embed (query inside
the per-request cache-miss branch, so reused embeddings aren't double-counted).

Note: the rename orphans any pre-existing embeddings_queries/pages_chunks rows
in tenant app DBs (CP no longer maps them) — acceptable; pipeline is inert with
throwaway dev/sandbox data.

Deck #284 (folded into PR #875).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-08 13:17:53 +02:00

159 lines
6.9 KiB
Python

"""Best-effort usage-event recording for per-tenant metering (Deck #67).
A tenant Pod records billable operations (embedding queries, pages/chunks
embedded) into the app-DB ``usage_events`` table; the control plane later pulls
that table read-only into the billing ledger and syncs to Stripe Meter Events
(see control-plane ``usage-metering.md``). This module owns only the data-plane
recording side.
Design contract:
- **Flag-gated.** Writes are a no-op unless ``USAGE_METERING_ENABLED`` is true,
so OSS self-hosters and unmetered deployments do zero DB work.
- **Best-effort.** A metering-write failure is logged and dropped, never raised
into the user-facing operation. ``ON CONFLICT (event_id) DO NOTHING`` makes a
retried write a no-op.
- **Engine reuse.** Rather than opening its own engine, this store borrows the
process-wide :class:`RefreshTokenStorage` singleton (``get_shared_storage()``)
— same app DB, NullPool, dialect handling, and ``_DBConn`` shim. The shared
storage guarantees Alembic migrations (incl. ``usage_events``) already ran.
"""
import json
import logging
import time
import uuid
from datetime import datetime, timezone
from typing import Any
import anyio
from nextcloud_mcp_server.auth.storage import RefreshTokenStorage, get_shared_storage
from nextcloud_mcp_server.config import get_settings
from nextcloud_mcp_server.observability.metrics import record_db_operation
logger = logging.getLogger(__name__)
# Parameters bind untyped through the ``sa.text(...)`` shim; asyncpg infers
# each placeholder's type from its target column. For ``occurred_at``
# (TIMESTAMPTZ) it wants a real ``datetime`` (a string is rejected even with a
# CAST), so we bind the aware datetime object on Postgres; SQLite's sqlite3
# driver can't bind a ``datetime`` on Python 3.12+, so we bind an ISO string
# there. ``metadata`` (JSONB) takes a JSON string on both — asyncpg's jsonb
# codec accepts ``str`` directly, so no cast is needed. Same SQL both ways;
# only the ``occurred_at`` bind value differs by dialect.
_INSERT_SQL = (
"INSERT INTO usage_events (event_id, occurred_at, metric, value, metadata) "
"VALUES (?, ?, ?, ?, ?) "
"ON CONFLICT (event_id) DO NOTHING"
)
class UsageEventStore:
"""Append-only writer for the app-DB ``usage_events`` table."""
# Process-wide cached instance returned by ``shared()`` so the hot search
# path doesn't allocate a fresh wrapper per metered query. The store is
# stateless beyond its storage handle, so one instance is reusable.
# ``anyio.Lock()`` doesn't bind to an event loop at construction, so a
# class-level instance is safe to define here (mirrors
# ``get_shared_storage``'s ``_shared_lock``).
_shared_instance: "UsageEventStore | None" = None
_shared_lock: anyio.Lock = anyio.Lock()
def __init__(self, storage: RefreshTokenStorage) -> None:
self._storage = storage
@classmethod
async def shared(cls) -> "UsageEventStore":
"""Return the process-wide store backed by the storage singleton.
Cached after first build: ``get_shared_storage()`` already returns the
cached :class:`RefreshTokenStorage` (running ``initialize()`` / Alembic
on first access, so ``usage_events`` exists), and the wrapper itself is
stateless, so reusing one instance avoids a per-call allocation on the
``nc_semantic_search`` hot path. The lock mirrors ``get_shared_storage``
so two concurrent cold-start callers don't both build (and one silently
overwrite) the instance.
Tests should construct ``UsageEventStore(storage)`` directly rather than
via ``shared()``: the cache is a process global with no teardown hook,
so a test that called ``shared()`` would leak its storage into the next.
"""
async with cls._shared_lock:
if cls._shared_instance is None:
cls._shared_instance = cls(await get_shared_storage())
return cls._shared_instance
async def record_usage_event(
self,
*,
metric: str,
value: int,
occurred_at: datetime | None = None,
metadata: dict[str, Any] | None = None,
event_id: str | None = None,
enabled: bool | None = None,
) -> None:
"""Record one billable usage event (best-effort, flag-gated).
Does nothing unless ``USAGE_METERING_ENABLED`` is true. Any failure is
logged and swallowed — this must never break the caller's operation.
Args:
metric: Catalog metric, e.g. ``"tokens_embedded"`` or
``"pages_embedded"``.
value: Count/quantity for this event.
occurred_at: Operation completion time; defaults to now (UTC).
metadata: Optional rawest-unit context (provider, model, tokens,
doc_type, ...). Stored as JSONB (Postgres) / JSON text (SQLite).
event_id: Optional idempotency key; defaults to a fresh UUID4.
enabled: The resolved ``USAGE_METERING_ENABLED`` value. ``None``
(default) re-reads it via ``get_settings()`` so the store stays
self-gating for standalone/test use. Hot-path callers that
already hold the flag should pass it to avoid a second uncached
``Settings`` build (``get_settings()`` is non-cached per
ADR-024 and ``nc_semantic_search`` is on the query path).
"""
if enabled is None:
enabled = get_settings().usage_metering_enabled
if not enabled:
return
start = time.time()
try:
event_id = event_id or str(uuid.uuid4())
when = occurred_at or datetime.now(timezone.utc)
# asyncpg takes the datetime object directly; sqlite3 needs a string.
when_bind = (
when if self._storage.dialect == "postgresql" else when.isoformat()
)
# json.dumps lives inside the best-effort try: a non-serializable
# metadata dict must be swallowed like any other write failure, not
# raised into the caller's operation (see the contract above).
params = (
event_id,
when_bind,
metric,
value,
json.dumps(metadata, sort_keys=True) if metadata is not None else None,
)
async with self._storage.acquire() as db:
await db.execute(_INSERT_SQL, params)
await db.commit()
record_db_operation(
self._storage.dialect, "insert", time.time() - start, "success"
)
except Exception:
# Best-effort: never surface a metering failure to the user op.
record_db_operation(
self._storage.dialect, "insert", time.time() - start, "error"
)
logger.warning(
"usage metering write dropped (metric=%s, value=%s)",
metric,
value,
exc_info=True,
)