[pull] develop from baserow:develop

backend/src/baserow/config/asgi.py

@@ -4,8 +4,8 @@
 
 from channels.routing import ProtocolTypeRouter, URLRouter
 
-from baserow.config.helpers import ConcurrencyLimiterASGI
-from baserow.core.mcp import baserow_mcp
+from baserow.config.helpers import ConcurrencyLimiterASGI, check_lazy_loaded_libraries
+from baserow.core.mcp import get_baserow_mcp_server
 from baserow.core.telemetry.telemetry import setup_logging, setup_telemetry
 from baserow.ws.routers import websocket_router
 
@@ -18,13 +18,16 @@
 # logging setup. Otherwise Django will try to destroy and log handlers we added prior.
 setup_logging()
 
+# Check that libraries meant to be lazy-loaded haven't been imported at startup.
+# This runs after Django is fully loaded, so it catches imports from all apps.
+check_lazy_loaded_libraries()
 
 application = ProtocolTypeRouter(
     {
         "http": ConcurrencyLimiterASGI(
             URLRouter(
                 [
-                    re_path(r"^mcp", baserow_mcp.sse_app()),
+                    re_path(r"^mcp", get_baserow_mcp_server().sse_app()),
                     re_path(r"", django_asgi_app),
                 ]
             ),

backend/src/baserow/config/celery.py

@@ -1,5 +1,8 @@
+from django.conf import settings
+
 from celery import Celery, signals
 
+from baserow.config.helpers import check_lazy_loaded_libraries
 from baserow.core.telemetry.tasks import BaserowTelemetryTask
 
 app = Celery("baserow")
@@ -26,3 +29,13 @@ def clear_local(*args, **kwargs):
 
 signals.task_prerun.connect(clear_local)
 signals.task_postrun.connect(clear_local)
+
+
+@signals.worker_process_init.connect
+def on_worker_init(**kwargs):
+    # This is only needed in asgi.py
+    settings.BASEROW_LAZY_LOADED_LIBRARIES.append("mcp")
+
+    # Check that libraries meant to be lazy-loaded haven't been imported at startup.
+    # This runs after Django is fully loaded, so it catches imports from all apps.
+    check_lazy_loaded_libraries()

backend/src/baserow/config/helpers.py

@@ -1,8 +1,54 @@
 import asyncio
+import sys
+
+from django.conf import settings
 
 from loguru import logger
 
 
+def check_lazy_loaded_libraries():
+    """
+    Check if any libraries that should be lazy-loaded have been imported at startup.
+
+    This function checks sys.modules against settings.BASEROW_LAZY_LOADED_LIBRARIES
+    and emits a warning if any of them have been loaded prematurely. This helps
+    catch accidental top-level imports that defeat the purpose of lazy loading
+    these heavy libraries to reduce memory footprint.
+
+    Only runs when DEBUG is True.
+    """
+
+    if not settings.DEBUG:
+        return
+
+    lazy_libs = getattr(settings, "BASEROW_LAZY_LOADED_LIBRARIES", [])
+    loaded_early = []
+
+    for lib in lazy_libs:
+        if lib in sys.modules:
+            loaded_early.append(lib)
+
+    if loaded_early:
+        libs_list = ", ".join(f'"{lib}"' for lib in loaded_early)
+        logger.warning(
+            f"The following libraries were loaded during startup but should be "
+            f"lazy-loaded to reduce memory footprint: {', '.join(loaded_early)}. "
+            f"Either import them inside functions/methods where they're used, or "
+            f"remove them from BASEROW_LAZY_LOADED_LIBRARIES if they're legitimately "
+            f"needed at startup. "
+            f"To debug, add the following code at the very top of your settings file "
+            f"(e.g., settings/dev.py, before any other imports):\n\n"
+            f"import sys, traceback\n"
+            f"class _T:\n"
+            f"    def find_module(self, n, p=None):\n"
+            f"        for lib in [{libs_list}]:\n"
+            f"            if n == lib or n.startswith(lib + '.'):\n"
+            f"                print(f'IMPORT: {{n}}'); traceback.print_stack(); sys.exit(1)\n"
+            f"        return None\n"
+            f"sys.meta_path.insert(0, _T())\n"
+        )
+
+
 class dummy_context:
     async def __aenter__(self):
         pass

backend/src/baserow/config/settings/base.py

@@ -11,10 +11,7 @@
 from django.core.exceptions import ImproperlyConfigured
 
 import dj_database_url
-import sentry_sdk
 from corsheaders.defaults import default_headers
-from sentry_sdk.integrations.django import DjangoIntegration
-from sentry_sdk.scrubber import DEFAULT_DENYLIST, EventScrubber
 
 from baserow.config.settings.utils import (
     Setting,
@@ -1303,18 +1300,42 @@ def __setitem__(self, key, value):
         print(e)
 
 
+# Libraries that should be lazy-loaded (imported inside functions/methods) to reduce
+# memory footprint at startup. If any of these are found in sys.modules during startup,
+# a warning will be shown suggesting to either lazy-load them or remove them from this
+# list if they're legitimately needed at startup.
+BASEROW_LAZY_LOADED_LIBRARIES = [
+    "openai",
+    "anthropic",
+    "mistralai",
+    "ollama",
+    "langchain_core",
+    "jira2markdown",
+    "saml2",
+    "openpyxl",
+    "numpy",
+]
+
+
 SENTRY_BACKEND_DSN = os.getenv("SENTRY_BACKEND_DSN")
 SENTRY_DSN = SENTRY_BACKEND_DSN or os.getenv("SENTRY_DSN")
-SENTRY_DENYLIST = DEFAULT_DENYLIST + ["username", "email", "name"]
 
 if SENTRY_DSN:
+    import sentry_sdk
+    from sentry_sdk.integrations.django import DjangoIntegration
+    from sentry_sdk.scrubber import DEFAULT_DENYLIST, EventScrubber
+
+    SENTRY_DENYLIST = DEFAULT_DENYLIST + ["username", "email", "name"]
+
     sentry_sdk.init(
         dsn=SENTRY_DSN,
         integrations=[DjangoIntegration(signals_spans=False, middleware_spans=False)],
         send_default_pii=False,
         event_scrubber=EventScrubber(recursive=True, denylist=SENTRY_DENYLIST),
         environment=os.getenv("SENTRY_ENVIRONMENT", ""),
     )
+else:
+    BASEROW_LAZY_LOADED_LIBRARIES.append("sentry_sdk")
 
 BASEROW_OPENAI_API_KEY = os.getenv("BASEROW_OPENAI_API_KEY", None)
 BASEROW_OPENAI_ORGANIZATION = os.getenv("BASEROW_OPENAI_ORGANIZATION", "") or None

backend/src/baserow/config/settings/dev.py

@@ -20,6 +20,11 @@
 INSTALLED_APPS.insert(0, "daphne")  # noqa: F405
 INSTALLED_APPS += ["django_extensions"]  # noqa: F405
 
+# daphne imports numpy via autobahn -> flatbuffers, so we exclude it from the
+# lazy-load check in dev mode. In production, numpy should still be lazy-loaded.
+if "numpy" in BASEROW_LAZY_LOADED_LIBRARIES:  # noqa: F405
+    BASEROW_LAZY_LOADED_LIBRARIES.remove("numpy")  # noqa: F405
+
 BASEROW_ENABLE_SILK = str_to_bool(os.getenv("BASEROW_ENABLE_SILK", "on"))
 if BASEROW_ENABLE_SILK:
     INSTALLED_APPS += ["silk"]  # noqa: F405

backend/src/baserow/config/wsgi.py

@@ -7,8 +7,10 @@
 https://docs.djangoproject.com/en/2.2/howto/deployment/wsgi/
 """
 
+from django.conf import settings
 from django.core.wsgi import get_wsgi_application
 
+from baserow.config.helpers import check_lazy_loaded_libraries
 from baserow.core.telemetry.telemetry import setup_logging, setup_telemetry
 
 # The telemetry instrumentation library setup needs to run prior to django's setup.
@@ -19,3 +21,10 @@
 # It is critical to setup our own logging after django has been setup and done its own
 # logging setup. Otherwise Django will try to destroy and log handlers we added prior.
 setup_logging()
+
+# This is only needed in asgi.py
+settings.BASEROW_LAZY_LOADED_LIBRARIES.append("mcp")
+
+# Check that libraries meant to be lazy-loaded haven't been imported at startup.
+# This runs after Django is fully loaded, so it catches imports from all apps.
+check_lazy_loaded_libraries()

backend/src/baserow/contrib/database/mcp/rows/tools.py

@@ -1,6 +1,4 @@
 from asgiref.sync import sync_to_async
-from mcp import Tool
-from mcp.types import TextContent
 from rest_framework.response import Response
 from starlette.status import HTTP_204_NO_CONTENT
 
@@ -23,6 +21,8 @@ class ListRowsMcpTool(MCPTool):
     name = "list_table_rows"
 
     async def list(self, endpoint):
+        from mcp import Tool
+
         return [
             Tool(
                 name=self.name,
@@ -64,6 +64,8 @@ async def call(
         name_parameters,
         call_arguments,
     ):
+        from mcp.types import TextContent
+
         table_id = call_arguments["table_id"]
         if not await sync_to_async(table_in_workspace_of_endpoint)(endpoint, table_id):
             return [TextContent(type="text", text="Table not in endpoint workspace.")]
@@ -92,6 +94,8 @@ class CreateRowMcpTool(MCPTool):
     name = "create_row_table_{id}"
 
     async def list(self, endpoint):
+        from mcp import Tool
+
         tables = await sync_to_async(get_all_tables)(endpoint)
         tables = await sync_to_async(remove_table_no_permission)(
             endpoint, tables, CreateRowDatabaseTableOperationType
@@ -127,6 +131,8 @@ async def call(
         name_parameters,
         call_arguments,
     ):
+        from mcp.types import TextContent
+
         table_id = name_parameters["id"]
         if not await sync_to_async(table_in_workspace_of_endpoint)(endpoint, table_id):
             return [TextContent(type="text", text="Table not in endpoint workspace.")]
@@ -148,6 +154,8 @@ class UpdateRowMcpTool(MCPTool):
     name = "update_row_table_{id}"
 
     async def list(self, endpoint):
+        from mcp import Tool
+
         tables = await sync_to_async(get_all_tables)(endpoint)
         tables = await sync_to_async(remove_table_no_permission)(
             endpoint, tables, UpdateDatabaseRowOperationType
@@ -187,6 +195,8 @@ async def call(
         name_parameters,
         call_arguments,
     ):
+        from mcp.types import TextContent
+
         table_id = name_parameters["id"]
         if not await sync_to_async(table_in_workspace_of_endpoint)(endpoint, table_id):
             return [TextContent(type="text", text="Table not in endpoint workspace.")]
@@ -211,6 +221,8 @@ class DeleteRowMcpTool(MCPTool):
     name = "delete_table_row"
 
     async def list(self, endpoint):
+        from mcp import Tool
+
         return [
             Tool(
                 name=self.name,
@@ -241,6 +253,8 @@ async def call(
         name_parameters,
         call_arguments,
     ):
+        from mcp.types import TextContent
+
         table_id = call_arguments["table_id"]
         if not await sync_to_async(table_in_workspace_of_endpoint)(endpoint, table_id):
             return [TextContent(type="text", text="Table not in endpoint workspace.")]

backend/src/baserow/contrib/database/mcp/table/tools.py

@@ -1,8 +1,6 @@
 import json
 
 from asgiref.sync import sync_to_async
-from mcp import Tool
-from mcp.types import TextContent
 
 from baserow.contrib.database.api.tables.serializers import (
     TableWithoutDataSyncSerializer,
@@ -16,6 +14,8 @@ class ListTablesMcpTool(MCPTool):
     name = "list_tables"
 
     async def list(self, endpoint):
+        from mcp import Tool
+
         return [
             Tool(
                 name=self.name,
@@ -34,6 +34,8 @@ async def call(
         name_parameters,
         call_arguments,
     ):
+        from mcp.types import TextContent
+
         tables = await sync_to_async(get_all_tables)(endpoint)
         serializer = TableWithoutDataSyncSerializer(tables, many=True)
         table_json = json.dumps(serializer.data)

backend/src/baserow/contrib/integrations/ai/service_types.py

@@ -3,8 +3,6 @@
 
 from django.contrib.auth.models import AbstractUser
 
-from langchain_core.exceptions import OutputParserException
-from langchain_core.prompts import PromptTemplate
 from rest_framework import serializers
 from rest_framework.exceptions import ValidationError as DRFValidationError
 
@@ -18,7 +16,7 @@
 )
 from baserow.core.generative_ai.registries import generative_ai_model_type_registry
 from baserow.core.integrations.handler import IntegrationHandler
-from baserow.core.output_parsers import StrictEnumOutputParser
+from baserow.core.output_parsers import get_strict_enum_output_parser
 from baserow.core.services.dispatch_context import DispatchContext
 from baserow.core.services.exceptions import (
     ServiceImproperlyConfiguredDispatchException,
@@ -170,6 +168,9 @@ def dispatch_data(
         resolved_values: Dict[str, Any],
         dispatch_context: DispatchContext,
     ) -> Dict[str, Any]:
+        from langchain_core.exceptions import OutputParserException
+        from langchain_core.prompts import PromptTemplate
+
         if not service.ai_generative_ai_type:
             raise ServiceImproperlyConfiguredDispatchException(
                 "The AI provider type is missing."
@@ -228,7 +229,7 @@ def dispatch_data(
             choices_enum = enum.Enum(
                 "Choices", {f"OPTION_{i}": choice for i, choice in enumerate(choices)}
             )
-            output_parser = StrictEnumOutputParser(enum=choices_enum)
+            output_parser = get_strict_enum_output_parser(enum=choices_enum)
             format_instructions = output_parser.get_format_instructions()
             prompt_template = PromptTemplate(
                 template=prompt + "\n\nGiven this user query:\n\n{format_instructions}",