Merge pull request 'fix: harden MCP server reliability, build reproducibility, and auth' (#1) from fix/reliability-pool-build-guard into main

The DB is at max_connections=100 and several stack services hold persistent
pools (several as the postgres superuser, idle for hours), so peaks hit
"too many connections". The MCP is a minor contributor but easy to bound:

- Dockerfile: uvicorn --workers 2 → 1. The MCP's connection budget is
  workers × MCP_POOL_MAX, so this caps it at 8 backends instead of 16. Scale
  via MCP_POOL_MAX, not workers, so the budget stays obvious. (Pairs with the
  minconn=0 lazy pool already on this branch: 0 connections held when idle.)
- analytics_ro_role.sql: add idle_session_timeout=5min so the DB reaps the
  MCP's idle POOLED connections (idle_in_transaction never reaps them — they're
  idle outside a txn) and returns the slots. Safe because the server now
  discards + transparently retries a reaped connection instead of erroring.

Note: the dominant fix is stack-wide (get the superuser app pools onto bounded,
timed roles; consider PgBouncer; or raise max_connections) — out of this repo's
scope but documented in the review.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

.dockerignore

@@ -1,3 +1,5 @@
+# NOTE: uv.lock is intentionally NOT ignored — the Dockerfile copies it for
+# reproducible `uv sync --frozen` builds.
 .git
 .venv
 __pycache__

.gitignore

@@ -7,4 +7,5 @@ __pycache__/
 .DS_Store
 .ruff_cache/
 .mypy_cache/
-uv.lock
+# uv.lock IS committed on purpose — the Docker build uses `uv sync --frozen` for
+# reproducible installs, and the Coolify build clones the repo (it must include it).

Dockerfile

@@ -11,15 +11,26 @@ COPY --from=ghcr.io/astral-sh/uv:latest /uv /uvx /bin/
 WORKDIR /app
 
 # Install ONLY dependencies (flat module — the project itself is not a package).
-COPY pyproject.toml ./
+# Copy the lockfile and build --frozen so rebuilds are reproducible: without it,
-RUN uv sync --no-dev --no-install-project
+# `uv sync` re-resolves the >= ranges in pyproject.toml and a redeploy could pull a
+# newer, behaviour-changed mcp/starlette and break the running server.
+COPY pyproject.toml uv.lock ./
+RUN uv sync --no-dev --no-install-project --frozen
 ENV PATH="/app/.venv/bin:$PATH"
 
 COPY analytics_mcp.py ./
 
+# Run as a non-root user (least privilege; nothing here needs root).
+RUN useradd -m -u 10001 app && chown -R app:app /app
+USER app
+
 EXPOSE 8892
 
 HEALTHCHECK --interval=30s --timeout=3s --start-period=10s --retries=3 \
   CMD python -c "import urllib.request,sys; sys.exit(0 if urllib.request.urlopen('http://localhost:8892/healthz').status==200 else 1)" || exit 1
 
-CMD ["uvicorn", "analytics_mcp:app", "--host", "0.0.0.0", "--port", "8892", "--workers", "2"]
+# Single worker: this is a low-traffic read-only proxy for a handful of analysts, and
+# the DB connection budget = workers × MCP_POOL_MAX. One worker (× default pool 8) caps
+# the MCP at 8 backends instead of 16, which matters on a shared 100-connection DB.
+# Scale out by raising MCP_POOL_MAX, not workers, so the budget stays obvious.
+CMD ["uvicorn", "analytics_mcp:app", "--host", "0.0.0.0", "--port", "8892", "--workers", "1"]

analytics_mcp.py

@@ -26,9 +26,11 @@ Env:
 """
 from __future__ import annotations
 
+import hmac
 import logging
 import os
 import re
+import threading
 import time
 from contextlib import contextmanager
 
@@ -57,6 +59,7 @@ def _get_logger(name: str) -> logging.Logger:
         )
         root.addHandler(handler)
         root.setLevel(logging.INFO)
+        root.propagate = False   # don't double-emit through uvicorn's root handler
     return root.getChild(name)
 
 
@@ -78,26 +81,89 @@ READABLE_SCHEMAS = tuple(
 # Force read-only + a statement timeout at the connection level (belt + braces;
 # the analytics_ro role already sets these, but a self-contained server is safer
 # in case it is ever pointed at a less-restricted DSN).
-_pool = psycopg2.pool.ThreadedConnectionPool(
+POOL_MAX = int(os.getenv("MCP_POOL_MAX", "8"))
-    1,
+_POOL_OPTS = "-c default_transaction_read_only=on -c statement_timeout=30000 -c client_encoding=UTF8"
-    int(os.getenv("MCP_POOL_MAX", "8")),
+
-    DATABASE_URL,
+
-    options="-c default_transaction_read_only=on -c statement_timeout=30000 -c client_encoding=UTF8",
+def _init_pool(retries: int = 5, delay: float = 2.0) -> psycopg2.pool.ThreadedConnectionPool:
+    """Create the pool, retrying so a brief DB outage at deploy time doesn't crash
+    the worker into a boot loop. minconn=0 → no eager connect at import (connections
+    are opened lazily on first use)."""
+    last: Exception | None = None
+    for attempt in range(1, retries + 1):
+        try:
+            return psycopg2.pool.ThreadedConnectionPool(
+                0, POOL_MAX, DATABASE_URL, options=_POOL_OPTS
+            )
+        except psycopg2.OperationalError as exc:
+            last = exc
+            log.warning("pool init attempt %d/%d failed: %s", attempt, retries, exc)
+            if attempt < retries:
+                time.sleep(delay)
+    assert last is not None
+    raise last
+
+
+_pool = _init_pool()
+# FastMCP runs each sync tool in an anyio worker thread (default ~40). Gate checkouts
+# behind a bounded semaphore so that >POOL_MAX concurrent queries QUEUE (block in
+# their worker thread) instead of overflowing the pool and raising PoolError — a 500
+# to the analyst. The event loop is never blocked; only surplus worker threads wait.
+_pool_slots = threading.BoundedSemaphore(POOL_MAX)
+
+# SQLSTATE classes that mean the CONNECTION is gone (vs. a query that merely failed,
+# e.g. a statement_timeout, which is QueryCanceled / 57014 and leaves the conn usable).
+# class 08 = connection exception; 57P01/02/03 = admin/crash shutdown.
+_DISCONNECT_SQLSTATES = frozenset(
+    {"08000", "08003", "08006", "08001", "08004", "08007", "57P01", "57P02", "57P03"}
 )
 
 
+def _is_disconnect(exc: Exception) -> bool:
+    """True only for a genuinely lost connection — so we discard/retry on a dropped
+    socket but NOT on a statement_timeout or other in-session query error."""
+    if isinstance(exc, psycopg2.InterfaceError):
+        return True
+    if isinstance(exc, psycopg2.OperationalError):
+        code = getattr(exc, "pgcode", None)
+        return code is None or code in _DISCONNECT_SQLSTATES   # None = socket-level failure
+    return False
+
+
 @contextmanager
 def _ro_conn():
-    """Read-only connection; the transaction is ALWAYS rolled back (never commits)."""
+    """Read-only connection; the transaction is ALWAYS rolled back (never commits).
+
+    Dead connections (DB restart, network blip, crash) are DISCARDED rather than
+    recycled — otherwise a single broken socket poisons the pool and every later
+    query handed that connection fails until the container is recreated. A query that
+    merely errors (e.g. statement_timeout) leaves the connection healthy, so it is
+    rolled back and returned to the pool as normal."""
+    _pool_slots.acquire()
+    try:
         conn = _pool.getconn()
+    except Exception:
+        _pool_slots.release()
+        raise
+    broken = False
     try:
         conn.set_session(readonly=True, autocommit=False)
         yield conn
+    except Exception as exc:
+        broken = _is_disconnect(exc)
+        raise
     finally:
+        try:
+            if broken or conn.closed:
+                _pool.putconn(conn, close=True)
+            else:
                 try:
                     conn.rollback()
-        finally:
                     _pool.putconn(conn)
+                except (psycopg2.OperationalError, psycopg2.InterfaceError):
+                    _pool.putconn(conn, close=True)
+        finally:
+            _pool_slots.release()
 
 
 def _rows(cur) -> list[dict]:
@@ -135,6 +201,13 @@ def _strip_comments(sql: str) -> str:
     return sql.strip()
 
 
+def _strip_literals(sql: str) -> str:
+    """Blank out the contents of single-quoted string literals so the keyword guard
+    does not fire on DATA (e.g. WHERE summary ILIKE '%please delete%'). '' escapes
+    are handled so we don't mis-terminate a literal mid-string."""
+    return re.sub(r"'(?:[^']|'')*'", "''", sql)
+
+
 def _guard(sql: str) -> str:
     """Validate a single read-only statement; return the cleaned statement."""
     stripped = _strip_comments(sql)
@@ -146,7 +219,11 @@ def _guard(sql: str) -> str:
     stmt = parts[0].strip()
     if not re.match(r"^(select|with)\b", stmt, re.IGNORECASE):
         raise ValueError("Only SELECT / WITH queries are allowed.")
-    if _FORBIDDEN.search(stmt):
+    # Scan with string literals blanked: the blocklist's real job is to reject
+    # data-modifying CTEs (WITH x AS (DELETE ... RETURNING ...)), not to trip over a
+    # keyword that merely appears inside a filter value. (Writes are impossible anyway
+    # via the analytics_ro GRANTs + the read-only, rolled-back transaction.)
+    if _FORBIDDEN.search(_strip_literals(stmt)):
         raise ValueError("Query contains a forbidden (write/DDL) keyword.")
     return stmt
 
@@ -179,18 +256,31 @@ mcp = FastMCP("fireside-analytics", stateless_http=True, transport_security=_tra
 def query(sql: str, max_rows: int = 1000) -> dict:
     """Run a read-only SELECT/WITH query against the fleet database.
 
-    Only the reporting.* and tracksolid.* schemas are readable. Single statement
+    Readable schemas are the analytics_ro grant surface (reporting, tracksolid,
-    only; write/DDL is rejected. Returns up to `max_rows` rows (default 1000, hard
+    tickets, fuel by default). Single statement only; write/DDL is rejected. Returns
-    cap 10000). A LIMIT is auto-applied when absent. Result: {row_count, truncated, rows}.
+    up to `max_rows` rows (default 1000, hard cap 10000). A LIMIT is auto-applied when
+    absent. Result: {row_count, truncated, rows}.
     """
     stmt = _guard(sql)
     cap = max(1, min(int(max_rows), MAX_ROWS_CEIL))
     if not re.search(r"\blimit\b", stmt, re.IGNORECASE):
         stmt = f"{stmt}\nLIMIT {cap + 1}"   # +1 row to detect truncation
     t0 = time.monotonic()
+    # Retry once on a dead connection (only): _ro_conn() discards broken sockets, so a
+    # second attempt gets a fresh one — making a recycled-but-stale pool connection
+    # invisible to the analyst. A real query error (statement_timeout, bad SQL) is NOT
+    # retried; it surfaces immediately.
+    rows = []
+    for attempt in (1, 2):
+        try:
             with _ro_conn() as conn, conn.cursor() as cur:
                 cur.execute(stmt)
                 rows = _rows(cur)
+            break
+        except Exception as exc:
+            if attempt == 2 or not _is_disconnect(exc):
+                raise
+            log.warning("stale DB connection on attempt %d — retrying once", attempt)
     truncated = len(rows) > cap
     rows = rows[:cap]
     dur_ms = int((time.monotonic() - t0) * 1000)
@@ -200,7 +290,7 @@ def query(sql: str, max_rows: int = 1000) -> dict:
 
 @mcp.tool()
 def list_schemas() -> list[dict]:
-    """List the readable schemas (reporting, tracksolid) with their object counts."""
+    """List the readable schemas (reporting, tracksolid, tickets, fuel) with object counts."""
     with _ro_conn() as conn, conn.cursor() as cur:
         cur.execute(
             "SELECT table_schema AS schema, count(*) AS objects "
@@ -213,7 +303,7 @@ def list_schemas() -> list[dict]:
 
 @mcp.tool()
 def list_tables(schema: str) -> list[dict]:
-    """List tables + views in a schema (must be reporting or tracksolid)."""
+    """List tables + views in a schema (must be one of the readable schemas)."""
     if schema not in READABLE_SCHEMAS:
         raise ValueError(f"schema must be one of {READABLE_SCHEMAS}")
     with _ro_conn() as conn, conn.cursor() as cur:
@@ -280,13 +370,26 @@ _TOKENS = {
 }
 
 
+def _lookup_token(token: str) -> str | None:
+    """Constant-time token match: compare against every known token so the response
+    time does not reveal how far a guessed prefix matched. Cheap for a handful of
+    per-analyst tokens."""
+    if not token:
+        return None
+    match = None
+    for known, name in _TOKENS.items():
+        if hmac.compare_digest(token, known):
+            match = name
+    return match
+
+
 class BearerAuth(BaseHTTPMiddleware):
     async def dispatch(self, request, call_next):
         if request.url.path == "/healthz":
             return await call_next(request)
         auth = request.headers.get("authorization", "")
         token = auth[7:] if auth.lower().startswith("bearer ") else ""
-        caller = _TOKENS.get(token)
+        caller = _lookup_token(token)
         if caller is None:
             return JSONResponse({"error": "unauthorized"}, status_code=401)
         request.state.caller = caller
@@ -294,7 +397,7 @@ class BearerAuth(BaseHTTPMiddleware):
 
 
 async def healthz(_request):
-    return JSONResponse({"ok": True, "tokens": len(_TOKENS)})
+    return JSONResponse({"ok": True})
 
 
 app = mcp.streamable_http_app()

deploy.sh

@@ -65,14 +65,18 @@ docker rm -f "$NAME" 2>/dev/null || true
 docker run -d --name "$NAME" --restart unless-stopped \
   --network "$APPNET" \
   --env-file "$ENV_FILE" \
+  --log-opt max-size=10m --log-opt max-file=5 \
   --label 'traefik.enable=true' \
   --label 'traefik.docker.network=coolify' \
   --label 'traefik.http.middlewares.redirect-to-https.redirectscheme.scheme=https' \
+  --label 'traefik.http.middlewares.fleetmcp-ratelimit.ratelimit.average=30' \
+  --label 'traefik.http.middlewares.fleetmcp-ratelimit.ratelimit.burst=60' \
   --label "traefik.http.routers.http-0-fleetmcp.entryPoints=http" \
   --label "traefik.http.routers.http-0-fleetmcp.middlewares=redirect-to-https" \
   --label "traefik.http.routers.http-0-fleetmcp.rule=Host(\`${HOST_DOMAIN}\`)" \
   --label "traefik.http.routers.https-0-fleetmcp.entryPoints=https" \
   --label "traefik.http.routers.https-0-fleetmcp.rule=Host(\`${HOST_DOMAIN}\`)" \
+  --label "traefik.http.routers.https-0-fleetmcp.middlewares=fleetmcp-ratelimit" \
   --label "traefik.http.routers.https-0-fleetmcp.tls=true" \
   --label "traefik.http.routers.https-0-fleetmcp.tls.certresolver=letsencrypt" \
   --label "traefik.http.services.fleetmcp.loadbalancer.server.port=${PORT}" \

scripts/analytics_ro_role.sql

@@ -62,3 +62,9 @@ ALTER DEFAULT PRIVILEGES FOR ROLE postgres         IN SCHEMA fuel       GRANT EX
 ALTER ROLE analytics_ro SET default_transaction_read_only = on;
 ALTER ROLE analytics_ro SET statement_timeout = '30s';
 ALTER ROLE analytics_ro SET idle_in_transaction_session_timeout = '60s';
+-- Cap idle POOLED connections too (these sit idle but NOT in a transaction, so the
+-- idle_in_transaction timeout never reaps them). On a shared 100-connection DB this
+-- returns slots the MCP isn't using. Safe with the server's dead-connection handling:
+-- a reaped pooled connection is discarded + transparently retried, not surfaced as an
+-- error. (Requires PostgreSQL 14+.)
+ALTER ROLE analytics_ro SET idle_session_timeout = '5min';