Keboola Explorer MCP Server

import abc import asyncio import json import logging import time from typing import Any, Literal, Mapping, Optional, Sequence from httpx import HTTPStatusError from pydantic import Field, TypeAdapter from pydantic.dataclasses import dataclass from keboola_mcp_server.clients.client import KeboolaClient LOG = logging.getLogger(__name__) @dataclass(frozen=True) class TableFqn: """The properly quoted parts of a fully qualified table name.""" # TODO: refactor this and probably use just a simple string db_name: str # project_id in a BigQuery schema_name: str # dataset in a BigQuery table_name: str quote_char: str = '' @property def identifier(self) -> str: """Returns the properly quoted database identifier.""" return '.'.join( f'{self.quote_char}{n}{self.quote_char}' for n in [self.db_name, self.schema_name, self.table_name] ) def __repr__(self) -> str: return self.identifier def __str__(self) -> str: return self.__repr__() QueryStatus = Literal['ok', 'error'] SqlSelectDataRow = Mapping[str, Any] @dataclass(frozen=True) class SqlSelectData: columns: Sequence[str] = Field(description='Names of the columns returned from SQL select.') rows: Sequence[SqlSelectDataRow] = Field( description='Selected rows, each row is a dictionary of column: value pairs.' ) @dataclass(frozen=True) class QueryResult: status: QueryStatus = Field(description='Status of running the SQL query.') data: SqlSelectData | None = Field(None, description='Data selected by the SQL SELECT query.') message: str | None = Field(None, description='Either an error message or the information from non-SELECT queries.') @property def is_ok(self) -> bool: return self.status == 'ok' @property def is_error(self) -> bool: return not self.is_ok class _Workspace(abc.ABC): def __init__(self, workspace_id: int) -> None: self._workspace_id = workspace_id @property def id(self) -> int: return self._workspace_id @abc.abstractmethod def get_sql_dialect(self) -> str: pass @abc.abstractmethod def get_quoted_name(self, name: str) -> str: pass @abc.abstractmethod async def get_table_fqn(self, table: Mapping[str, Any]) -> TableFqn | None: """Gets the fully qualified name of a Keboola table.""" # TODO: use a pydantic class for the 'table' param pass @abc.abstractmethod async def execute_query(self, sql_query: str) -> QueryResult: """Runs a SQL SELECT query.""" pass class _SnowflakeWorkspace(_Workspace): def __init__(self, workspace_id: int, schema: str, client: KeboolaClient): super().__init__(workspace_id) self._schema = schema # default schema created for the workspace self._client = client def get_sql_dialect(self) -> str: return 'Snowflake' def get_quoted_name(self, name: str) -> str: return f'"{name}"' # wrap name in double quotes async def get_table_fqn(self, table: Mapping[str, Any]) -> TableFqn | None: table_id = table['id'] db_name: str | None = None schema_name: str | None = None table_name: str | None = None if source_table := table.get('sourceTable'): # a table linked from some other project schema_name, table_name = source_table['id'].rsplit(sep='.', maxsplit=1) source_project_id = source_table['project']['id'] # sql = f"show databases like '%_{source_project_id}';" sql = ( f'select "DATABASE_NAME" from "INFORMATION_SCHEMA"."DATABASES" ' f'where "DATABASE_NAME" like \'%_{source_project_id}\';' ) result = await self.execute_query(sql) if result.is_ok and result.data and result.data.rows: db_name = result.data.rows[0]['DATABASE_NAME'] else: LOG.error(f'Failed to run SQL: {sql}, SAPI response: {result}') else: sql = 'select CURRENT_DATABASE() as "current_database";' result = await self.execute_query(sql) if result.is_ok and result.data and result.data.rows: row = result.data.rows[0] db_name = row['current_database'] if '.' in table_id: # a table local in a project for which the snowflake connection/workspace is open schema_name, table_name = table_id.rsplit(sep='.', maxsplit=1) else: # a table not in the project, but in the writable schema created for the workspace # TODO: we should never come here, because the tools for listing tables can only see # tables that are in the project schema_name = self._schema table_name = table['name'] else: LOG.error(f'Failed to run SQL: {sql}, SAPI response: {result}') if db_name and schema_name and table_name: fqn = TableFqn(db_name, schema_name, table_name, quote_char='"') return fqn else: return None async def execute_query(self, sql_query: str) -> QueryResult: resp = await self._client.storage_client.workspace_query(workspace_id=self.id, query=sql_query) return TypeAdapter(QueryResult).validate_python(resp) class _BigQueryWorkspace(_Workspace): _BQ_FIELDS = {'_timestamp'} def __init__(self, workspace_id: int, dataset_id: str, project_id: str, client: KeboolaClient): super().__init__(workspace_id) self._dataset_id = dataset_id # default dataset created for the workspace self._project_id = project_id self._client = client def get_sql_dialect(self) -> str: return 'BigQuery' def get_quoted_name(self, name: str) -> str: return f'`{name}`' # wrap name in back tick async def get_table_fqn(self, table: Mapping[str, Any]) -> TableFqn | None: table_id = table['id'] schema_name: str | None = None table_name: str | None = None if '.' in table_id: # a table local in a project for which the workspace is open schema_name, table_name = table_id.rsplit(sep='.', maxsplit=1) schema_name = schema_name.replace('.', '_').replace('-', '_') else: # a table not in the project, but in the writable schema created for the workspace # TODO: we should never come here, because the tools for listing tables can only see # tables that are in the project schema_name = self._dataset_id table_name = table['name'] if schema_name and table_name: fqn = TableFqn(self._project_id, schema_name, table_name, quote_char='`') return fqn else: return None async def execute_query(self, sql_query: str) -> QueryResult: resp = await self._client.storage_client.workspace_query(workspace_id=self.id, query=sql_query) return TypeAdapter(QueryResult).validate_python(resp) @dataclass(frozen=True) class _WspInfo: id: int schema: str backend: str credentials: str | None # the backend credentials; it can contain serialized JSON data readonly: bool | None @staticmethod def from_sapi_info(sapi_wsp_info: Mapping[str, Any]) -> '_WspInfo': _id = sapi_wsp_info.get('id') backend = sapi_wsp_info.get('connection', {}).get('backend') _schema = sapi_wsp_info.get('connection', {}).get('schema') credentials = sapi_wsp_info.get('connection', {}).get('user') readonly = sapi_wsp_info.get('readOnlyStorageAccess') return _WspInfo(id=_id, schema=_schema, backend=backend, credentials=credentials, readonly=readonly) class WorkspaceManager: STATE_KEY = 'workspace_manager' MCP_META_KEY = 'KBC.McpServer.workspaceId' @classmethod def from_state(cls, state: Mapping[str, Any]) -> 'WorkspaceManager': instance = state[cls.STATE_KEY] assert isinstance(instance, WorkspaceManager), f'Expected WorkspaceManager, got: {instance}' return instance def __init__(self, client: KeboolaClient, workspace_schema: str | None = None): # We use the read-only workspace with access to all project data which lives in the production branch. # Hence we need KeboolaClient bound to the production/default branch. self._client = client.with_branch_id(None) self._workspace_schema = workspace_schema self._workspace: _Workspace | None = None self._table_fqn_cache: dict[str, TableFqn] = {} async def _find_ws_by_schema(self, schema: str) -> _WspInfo | None: """Finds the workspace info by its schema.""" for sapi_wsp_info in await self._client.storage_client.workspace_list(): assert isinstance(sapi_wsp_info, dict) wi = _WspInfo.from_sapi_info(sapi_wsp_info) if wi.id and wi.backend and wi.schema and wi.schema == schema: return wi return None async def _find_ws_by_id(self, workspace_id: int) -> _WspInfo | None: """Finds the workspace info by its ID.""" try: sapi_wsp_info = await self._client.storage_client.workspace_detail(workspace_id) assert isinstance(sapi_wsp_info, dict) wi = _WspInfo.from_sapi_info(sapi_wsp_info) if wi.id and wi.backend and wi.schema: return wi else: raise ValueError(f'Invalid workspace info: {sapi_wsp_info}') except HTTPStatusError as e: if e.response.status_code == 404: return None else: raise e async def _find_ws_in_branch(self) -> _WspInfo | None: """Finds the workspace info in the current branch.""" metadata = await self._client.storage_client.branch_metadata_get() for m in metadata: if m.get('key') == self.MCP_META_KEY: workspace_id = m.get('value') if workspace_id and (info := await self._find_ws_by_id(workspace_id)) and info.readonly: return info return None async def _create_ws(self, *, timeout_sec: float = 300.0) -> _WspInfo | None: """ Creates a new workspace in the current branch and returns its info. :param timeout_sec: The number of seconds to wait for the workspace creation job to finish. :return: The workspace info if the workspace was created successfully, None otherwise. """ # Verify token before creating workspace to ensure it has proper permissions token_info = await self._client.storage_client.verify_token() # Check for defaultBackend parameter in token info under owner object owner_info = token_info.get('owner', {}) default_backend = owner_info.get('defaultBackend') resp = None if default_backend == 'snowflake': resp = await self._client.storage_client.workspace_create( login_type='snowflake-person-sso', backend=default_backend, async_run=True, read_only_storage_access=True, ) elif default_backend == 'bigquery': resp = await self._client.storage_client.workspace_create( login_type='default', backend=default_backend, async_run=True, read_only_storage_access=True ) else: raise ValueError(f'Unexpected default backend: {default_backend}') assert 'id' in resp, f'Expected job ID in response: {resp}' assert isinstance(resp['id'], int) job_id = resp['id'] start_ts = time.perf_counter() LOG.info(f'Requested new workspace: job_id={job_id}, timeout={timeout_sec:.2f} seconds') while True: job_info = await self._client.storage_client.job_detail(job_id) job_status = job_info['status'] duration = time.perf_counter() - start_ts LOG.info( f'Job info: job_id={job_id}, status={job_status}, ' f'duration={duration:.2f} seconds, timeout={timeout_sec:.2f} seconds' ) if job_info['status'] == 'success': assert 'results' in job_info, f'Expected `results` in job info: {job_info}' assert isinstance(job_info['results'], dict) assert 'id' in job_info['results'], f'Expected `id` in `results` in job info: {job_info}' assert isinstance(job_info['results']['id'], int) workspace_id = job_info['results']['id'] LOG.info(f'Created workspace: {workspace_id}') return await self._find_ws_by_id(workspace_id) elif duration > timeout_sec: LOG.info(f'Workspace creation timed out after {duration:.2f} seconds.') return None else: remaining_time = max(0.0, timeout_sec - duration) await asyncio.sleep(min(5.0, remaining_time)) def _init_workspace(self, info: _WspInfo) -> _Workspace: """Creates a new `Workspace` instance based on the workspace info.""" if info.backend == 'snowflake': return _SnowflakeWorkspace(workspace_id=info.id, schema=info.schema, client=self._client) elif info.backend == 'bigquery': credentials = json.loads(info.credentials or '{}') if project_id := credentials.get('project_id'): return _BigQueryWorkspace( workspace_id=info.id, dataset_id=info.schema, project_id=project_id, client=self._client, ) else: raise ValueError(f'No credentials or no project ID in workspace: {info.schema}') else: raise ValueError(f'Unexpected backend type "{info.backend}" in workspace: {info.schema}') async def _get_workspace(self) -> _Workspace: if self._workspace: return self._workspace if self._workspace_schema: # use the workspace that was explicitly requested # this workspace must never be written to the default branch metadata LOG.info(f'Looking up workspace by schema: {self._workspace_schema}') if info := await self._find_ws_by_schema(self._workspace_schema): LOG.info(f'Found workspace: {info}') self._workspace = self._init_workspace(info) return self._workspace else: raise ValueError( f'No Keboola workspace found or the workspace has no read-only storage access: ' f'workspace_schema={self._workspace_schema}' ) LOG.info('Looking up workspace in the default branch.') if info := await self._find_ws_in_branch(): # use the workspace that has already been created by the MCP server and noted to the branch LOG.info(f'Found workspace: {info}') self._workspace = self._init_workspace(info) return self._workspace # create a new workspace and note its ID to the branch LOG.info('Creating workspace in the default branch.') if info := await self._create_ws(): # update the branch metadata with the workspace ID meta = await self._client.storage_client.branch_metadata_update({self.MCP_META_KEY: info.id}) LOG.info(f'Set metadata in the default branch: {meta}') # use the newly created workspace self._workspace = self._init_workspace(info) return self._workspace else: raise ValueError('Failed to initialize Keboola Workspace.') async def execute_query(self, sql_query: str) -> QueryResult: workspace = await self._get_workspace() return await workspace.execute_query(sql_query) async def get_table_fqn(self, table: Mapping[str, Any]) -> Optional[TableFqn]: table_id = table['id'] if table_id in self._table_fqn_cache: return self._table_fqn_cache[table_id] workspace = await self._get_workspace() fqn = await workspace.get_table_fqn(table) if fqn: self._table_fqn_cache[table_id] = fqn return fqn async def get_quoted_name(self, name: str) -> str: workspace = await self._get_workspace() return workspace.get_quoted_name(name) async def get_sql_dialect(self) -> str: workspace = await self._get_workspace() return workspace.get_sql_dialect() async def get_workspace_id(self) -> int: workspace = await self._get_workspace() return workspace.id