odps/df/backends/core.py

#!/usr/bin/env python # -*- coding: utf-8 -*- # Copyright 1999-2022 Alibaba Group Holding Ltd. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from __future__ import absolute_import import itertools import time import types import os import sys import threading from operator import itemgetter from ...compat import six, Enum, Iterable from ...models import Resource from ...config import options from ...dag import DAG from ...utils import init_progress_ui from ...ui.progress import create_instance_group from ...compat import futures from ...types import PartitionSpec from .. import utils from ..expr.expressions import Expr, CollectionExpr, Scalar from ..expr.core import ExprDictionary, ExprDAG from .context import context, ExecuteContext from .errors import DagDependencyError, CompileError from .formatter import ExprExecutionGraphFormatter class EngineTypes(Enum): ODPS = 'ODPS' PANDAS = 'PANDAS' SEAHAWKS = 'SEAHAWKS' SQLALCHEMY = 'SQLALCHEMY' ALGO = 'ALGO' class Backend(object): def visit_source_collection(self, expr): raise NotImplementedError def visit_project_collection(self, expr): raise NotImplementedError def visit_apply_collection(self, expr): raise NotImplementedError def visit_lateral_view(self, expr): raise NotImplementedError def visit_filter_collection(self, expr): raise NotImplementedError def visit_filter_partition_collection(self, expr): raise NotImplementedError def visit_algo(self, expr): raise NotImplementedError def visit_slice_collection(self, expr): raise NotImplementedError def visit_element_op(self, expr): raise NotImplementedError def visit_binary_op(self, expr): raise NotImplementedError def visit_unary_op(self, expr): raise NotImplementedError def visit_math(self, expr): raise NotImplementedError def visit_string_op(self, expr): raise NotImplementedError def visit_datetime_op(self, expr): raise NotImplementedError def visit_composite_op(self, expr): raise NotImplementedError def visit_groupby(self, expr): raise NotImplementedError def visit_mutate(self, expr): raise NotImplementedError def visit_reshuffle(self, expr): raise NotImplementedError def visit_value_counts(self, expr): raise NotImplementedError def visit_sort(self, expr): raise NotImplementedError def visit_sort_column(self, expr): raise NotImplementedError def visit_distinct(self, expr): raise NotImplementedError def visit_sample(self, expr): raise NotImplementedError def visit_pivot(self, expr): raise NotImplementedError def visit_reduction(self, expr): raise NotImplementedError def visit_user_defined_aggregator(self, expr): raise NotImplementedError def visit_column(self, expr): raise NotImplementedError def visit_function(self, expr): raise NotImplementedError def visit_builtin_function(self, expr): raise NotImplementedError def visit_sequence(self, expr): raise NotImplementedError def visit_cum_window(self, expr): raise NotImplementedError def visit_rank_window(self, expr): raise NotImplementedError def visit_shift_window(self, expr): raise NotImplementedError def visit_scalar(self, expr): raise NotImplementedError def visit_join(self, expr): raise NotImplementedError def visit_cast(self, expr): raise NotImplementedError def visit_union(self, expr): raise NotImplementedError def visit_concat(self, expr): raise NotImplementedError def visit_append_id(self, expr): raise NotImplementedError def visit_split(self, expr): raise NotImplementedError def visit_extract_kv(self, expr): raise NotImplementedError class ExecuteNode(object): def __init__(self, expr_dag, result_index=None, callback=None): self.expr_dag = expr_dag self.result_index = result_index self.callback = callback @property def expr(self): return self.expr_dag.root def run(self, **execute_kw): raise NotImplementedError def __call__(self, ui=None, progress_proportion=None): res = self.run(ui=ui, progress_proportion=progress_proportion) if self.callback: self.callback(res) return res def __repr__(self): raise NotImplementedError def _repr_html_(self): raise NotImplementedError class ExecuteDAG(DAG): def _run(self, ui, progress_proportion=1.0): curr_progress = ui.current_progress() or 0 try: calls = self.topological_sort() results = [None] * len(calls) result_idx = dict() for i, call in enumerate(calls): res = call(ui=ui, progress_proportion=progress_proportion / len(calls)) results[i] = res if call.result_index is not None: result_idx[call.result_index] = i return [results[result_idx[idx]] for idx in sorted(result_idx)] except Exception as e: if self._can_fallback() and self._need_fallback(e): ui.update(curr_progress) return self.fallback()._run(ui, progress_proportion) raise def _run_in_parallel(self, ui, n_parallel, wait=True, timeout=None, progress_proportion=1.0): submits_lock = threading.RLock() submits = dict() user_wait = dict() result_wait = dict() results = dict() curr_progress = ui.current_progress() or 0 def actual_run(dag=None, is_fallback=False): dag = dag or self calls = dag.topological_sort() result_calls = sorted([c for c in calls if c.result_index is not None], key=lambda x: x.result_index) fallback = threading.Event() if is_fallback: ui.update(curr_progress) def close_ui(*_): with submits_lock: if all(call in submits and call in results for call in result_calls): ui.close() executor = futures.ThreadPoolExecutor(max_workers=n_parallel) for call in calls: if call.result_index is not None and is_fallback: # if is fallback, we do not create new future # cause the future objects have been passed to user future = result_wait[call.result_index] else: future = futures.Future() user_wait[call] = future if call.result_index is not None: future.add_done_callback(close_ui) if not is_fallback: result_wait[call.result_index] = future for call in calls: def run(func): try: if fallback.is_set(): raise DagDependencyError('Node execution failed due to callback') if call.result_index is None or not is_fallback: user_wait[func].set_running_or_notify_cancel() prevs = dag.predecessors(func) if prevs: fs = [user_wait[p] for p in prevs] for f in fs: if f.exception(): raise DagDependencyError('Node execution failed due to failure of ' 'previous node, exception: %s' % f.exception()) res = func(ui=ui, progress_proportion=progress_proportion / len(calls)) results[func] = res user_wait[func].set_result(res) return res except: e, tb = sys.exc_info()[1:] if not is_fallback and self._can_fallback() and self._need_fallback(e): if not fallback.is_set(): fallback.set() new_dag = dag.fallback() actual_run(new_dag, True) if not fallback.is_set(): results[func] = (e, tb) if six.PY2: user_wait[func].set_exception_info(e, tb) else: user_wait[func].set_exception(e) raise finally: with submits_lock: for f in dag.successors(func): if f in submits: continue prevs = dag.predecessors(f) if all(p in submits and user_wait[p].done() for p in prevs): submits[f] = executor.submit(run, f) if not dag.predecessors(call): with submits_lock: submits[call] = executor.submit(run, call) if wait: dones, _ = futures.wait(user_wait.values()) for done in dones: done.result() return [results[c] for c in sorted([c for c in calls if c.result_index is not None], key=lambda x: x.result_index)] if timeout: futures.wait(user_wait.values(), timeout=timeout) actual_run() if wait: return [it[1].result() for it in sorted(result_wait.items(), key=itemgetter(0))] else: return [it[1] for it in sorted(result_wait.items(), key=itemgetter(0))] def execute(self, ui=None, async_=False, n_parallel=1, timeout=None, close_and_notify=True, progress_proportion=1.0, **kw): async_ = kw.get('async', async_) ui = ui or init_progress_ui(lock=async_) succeeded = False if not async_: try: if n_parallel <= 1: results = self._run(ui, progress_proportion) else: results = self._run_in_parallel(ui, n_parallel, progress_proportion=progress_proportion) succeeded = True return results finally: if close_and_notify or succeeded: ui.close() if succeeded: ui.notify('DataFrame execution succeeded') else: ui.notify('DataFrame execution failed') else: try: fs = self._run_in_parallel(ui, n_parallel, wait=not async_, timeout=timeout, progress_proportion=progress_proportion) succeeded = True return fs finally: if succeeded: ui.notify('DataFrame execution submitted') else: ui.notify('DataFrame execution failed to summit') def _can_fallback(self): return hasattr(self, 'fallback') and self.fallback is not None def _need_fallback(self, e): return hasattr(self, 'need_fallback') and self.need_fallback(e) def __repr__(self): return ExprExecutionGraphFormatter(self)._to_str() def _repr_html_(self): return ExprExecutionGraphFormatter(self)._to_html() class Engine(object): def stop(self): pass @classmethod def _convert_table(cls, expr): if isinstance(expr, Expr): return utils.to_collection(expr) expr_dag = expr root = utils.to_collection(expr_dag.root) if root is not expr_dag.root: new_expr_dag = ExprDAG(root, dag=expr_dag) new_expr_dag.ensure_all_nodes_in_dag() return new_expr_dag return expr_dag def _cache(self, expr_dag, dag, expr, **kwargs): # should return the data raise NotImplementedError def _dispatch(self, expr_dag, expr, ctx): if expr._need_cache: if not ctx.is_cached(expr): def h(): def inner(*args, **kwargs): ret = self._cache(*args, **kwargs) if ret: data, node = ret ctx.cache(expr, data) return node return inner return h() else: cached = ctx.get_cached(expr) if isinstance(expr, CollectionExpr): cached = cached.copy() expr_dag.substitute(expr, cached) elif expr._deps: return self._handle_dep def _new_analyzer(self, expr_dag, on_sub=None): raise NotImplementedError def _new_rewriter(self, expr_dag): return def _analyze(self, expr_dag, dag, **kwargs): from .optimize import Optimizer def sub_has_dep(_, sub): if sub._deps is not None: kw = dict(kwargs) kw['finish'] = False kw.pop('head', None) kw.pop('tail', None) self._handle_dep(ExprDAG(sub, dag=expr_dag), dag, sub, **kw) # analyze first self._new_analyzer(expr_dag, on_sub=sub_has_dep).analyze() # optimize return Optimizer(expr_dag).optimize() def _rewrite(self, expr_dag): # rewrite if exist rewriter = self._new_rewriter(expr_dag) if rewriter: return rewriter.rewrite() return expr_dag.root def _new_execute_node(self, expr_dag): return ExecuteNode(expr_dag) def _handle_dep(self, expr_dag, dag, expr, **kwargs): root = expr_dag.root execute_nodes = [] for dep in root._deps: if isinstance(dep, tuple): if len(dep) == 3: node, action, callback = dep else: node, callback = dep action = '_execute' else: node, action, callback = dep, '_execute', None if callback: def dep_callback(res): callback(res, expr) else: dep_callback = None execute_node = getattr(self, action)(ExprDAG(node, dag=expr_dag), dag, node, analyze=False, **kwargs) execute_node.callback = dep_callback execute_nodes.append(execute_node) return execute_nodes def _handle_expr_args_kwargs(self, expr_args_kwargs): if len(expr_args_kwargs) == 1 and not isinstance(expr_args_kwargs[0], Expr) and \ all(isinstance(it, Expr) for it in expr_args_kwargs[0]): expr_args_kwargs = expr_args_kwargs[0] if all(isinstance(it, Expr) for it in expr_args_kwargs): expr_args_kwargs = [('_execute', it, (), {}) for it in expr_args_kwargs] return expr_args_kwargs def _process(self, *expr_args_kwargs): expr_args_kwargs = self._handle_expr_args_kwargs(expr_args_kwargs) def h(e): if isinstance(e, Scalar) and e.name is None: return e.rename('__rand_%s' % int(time.time())) if isinstance(e, CollectionExpr) and hasattr(e, '_proxy') and \ e._proxy is not None: return e._proxy return e src_exprs = [h(it[1]) for it in expr_args_kwargs] exprs_dags = self._build_expr_dag([self._convert_table(e) for e in src_exprs]) return exprs_dags, expr_args_kwargs def _compile_dag(self, expr_args_kwargs, exprs_dags): ctx = ExecuteContext() # expr -> new_expr dag = ExecuteDAG() for idx, it, expr_dag in zip(itertools.count(0), expr_args_kwargs, exprs_dags): action, src_expr, args, kwargs = it for node in expr_dag.traverse(): if hasattr(self, '_selecter') and not self._selecter.force_odps and hasattr(node, '_algo'): raise NotImplementedError h = self._dispatch(expr_dag, node, ctx) if h: kw = dict(kwargs) kw['finish'] = False if node is expr_dag.root: node_dag = expr_dag else: node_dag = ExprDAG(node, dag=expr_dag) h(node_dag, dag, node, **kw) args = args + (expr_dag, dag, src_expr) n = getattr(self, action)(*args, **kwargs) n.result_index = idx return dag def compile(self, *expr_args_kwargs): exprs_dags, expr_args_kwargs = self._process(*expr_args_kwargs) return self._compile_dag(expr_args_kwargs, exprs_dags) def _action(self, *exprs_args_kwargs, **kwargs): ui = kwargs.pop('ui', None) progress_proportion = kwargs.pop('progress_proportion', 1.0) async_ = kwargs.pop('async_', kwargs.pop('async', False)) n_parallel = kwargs.pop('n_parallel', 1) timeout = kwargs.pop('timeout', None) batch = kwargs.pop('batch', False) action = kwargs.pop('action', None) def transform(*exprs_args_kw): for expr_args_kwargs in exprs_args_kw: if len(expr_args_kwargs) == 3: expr, args, kw = expr_args_kwargs act = action else: act, expr, args, kw = expr_args_kwargs kw = kw.copy() kw.update(kwargs) yield act, expr, args, kw dag = self.compile(*transform(*exprs_args_kwargs)) try: res = self._execute_dag(dag, ui=ui, async_=async_, n_parallel=n_parallel, timeout=timeout, progress_proportion=progress_proportion) except KeyboardInterrupt: self.stop() sys.exit(1) if not batch: return res[0] return res def _do_execute(self, expr_dag, expr, **kwargs): raise NotImplementedError def _execute(self, expr_dag, dag, expr, **kwargs): # analyze first analyze = kwargs.pop('analyze', True) if analyze: kw = dict(kwargs) kw.pop('execute_kw', None) self._analyze(expr_dag, dag, **kw) engine = self execute_node = self._new_execute_node(expr_dag) group_key = kwargs.get('group') or self._create_progress_group(expr) def run(s, **execute_kw): kw = dict(kwargs) kw.update(kw.pop('execute_kw', dict())) kw.update(execute_kw) kw['group'] = group_key if 'ui' in kw: kw['ui'].add_keys(group_key) result = engine._do_execute(expr_dag, expr, **kw) if 'ui' in kw: kw['ui'].remove_keys(group_key) return result execute_node.run = types.MethodType(run, execute_node) self._add_node(execute_node, dag) return execute_node @classmethod def _reorder(cls, expr, table, cast=False, with_partitions=False): from .odpssql.engine import types as odps_engine_types from .. import NullScalar df_schema = odps_engine_types.odps_schema_to_df_schema(table.table_schema) expr_schema = expr.schema.to_ignorecase_schema() expr_table_schema = odps_engine_types.df_schema_to_odps_schema(expr_schema) case_dict = dict((n.lower(), n) for n in expr.schema.names) for col in expr_table_schema.columns: if col.name.lower() not in table.table_schema: raise CompileError('Column(%s) does not exist in target table %s, ' 'writing cannot be performed.' % (col.name, table.name)) t_col = table.table_schema[col.name.lower()] if not cast and not t_col.type.can_implicit_cast(col.type): raise CompileError('Cannot implicitly cast column %s from %s to %s.' % ( col.name, col.type, t_col.type)) if ( table.table_schema.names == expr_schema.names and df_schema.types[:len(table.table_schema.names)] == expr_schema.types ): return expr def field(name): expr_name = case_dict[name] if expr[expr_name].dtype == df_schema[name].type: return expr[expr_name] elif df_schema[name].type.can_implicit_cast(expr[expr_name].dtype) or cast: return expr[expr_name].astype(df_schema[name].type) else: raise CompileError('Column %s\'s type does not match, expect %s, got %s' % ( expr_name, expr[expr_name].dtype, df_schema[name].type)) names = [c.name for c in table.table_schema.columns] if with_partitions else table.table_schema.names return expr[[field(name) if name in expr_schema else NullScalar(df_schema[name].type).rename(name) for name in names]] @classmethod def _get_partition(cls, partition, table=None): if isinstance(partition, dict): p_spec = PartitionSpec() for name, val in six.iteritems(partition): p_spec[name] = val elif isinstance(partition, PartitionSpec): p_spec = partition else: if not isinstance(partition, six.string_types): raise TypeError('`partition` should be a str or dict, ' 'got {0} instead'.format(type(partition))) p_spec = PartitionSpec(partition) if table is not None: part_names = [p.name for p in table.table_schema.partitions] for name in part_names: if name not in p_spec: raise ValueError('Table has partition column {0} ' 'which not specified by `partition`'.format(name)) for name in p_spec.keys: if name not in table.table_schema._partition_schema: raise ValueError('Table does not have partition({0}) ' 'which specified in `partition`'.format(name)) if p_spec.keys != part_names: old_p_spec = p_spec p_spec = PartitionSpec() for n in part_names: p_spec[n] = old_p_spec[n] return p_spec def _do_persist(self, expr_dag, expr, name, **kwargs): raise NotImplementedError def _persist(self, name, expr_dag, dag, expr, **kwargs): # analyze first analyze = kwargs.pop('analyze', True) if analyze: self._analyze(expr_dag, dag, **kwargs) engine = self execute_node = self._new_execute_node(expr_dag) group_key = self._create_progress_group(expr) def run(s, **execute_kw): kw = dict(kwargs) kw.update(execute_kw) kw['group'] = group_key if 'ui' in kw: kw['ui'].add_keys(group_key) result = engine._do_persist(expr_dag, expr, name, **kw) if 'ui' in kw: kw['ui'].remove_keys(group_key) return result execute_node.run = types.MethodType(run, execute_node) self._add_node(execute_node, dag) return execute_node @classmethod def _handle_params(cls, *expr_args_kwargs, **kwargs): if isinstance(expr_args_kwargs[0], Expr): return [(expr_args_kwargs[0], expr_args_kwargs[1:], {})], kwargs elif isinstance(expr_args_kwargs[0], Iterable) and \ all(isinstance(e, Expr) for e in expr_args_kwargs[0]): args = expr_args_kwargs[1:] kwargs['batch'] = True return [(e, args, {}) for e in expr_args_kwargs[0]], kwargs else: kwargs['batch'] = True return expr_args_kwargs, kwargs @staticmethod def _create_ui(**kwargs): existing_ui = kwargs.get('ui') if existing_ui: return existing_ui async_ = kwargs.get('async_', kwargs.get('async', False)) ui = init_progress_ui(lock=async_, use_console=not async_) ui.status('Preparing') return ui @staticmethod def _create_progress_group(expr): node_name = getattr(expr, 'node_name', expr.__class__.__name__) return create_instance_group(node_name) def execute(self, *exprs_args_kwargs, **kwargs): exprs_args_kwargs, kwargs = self._handle_params(*exprs_args_kwargs, **kwargs) kwargs['ui'] = self._create_ui(**kwargs) kwargs['action'] = '_execute' return self._action(*exprs_args_kwargs, **kwargs) def persist(self, *exprs_args_kwargs, **kwargs): exprs_args_kwargs, kwargs = self._handle_params(*exprs_args_kwargs, **kwargs) kwargs['ui'] = self._create_ui(**kwargs) kwargs['action'] = '_persist' return self._action(*exprs_args_kwargs, **kwargs) def batch(self, *action_exprs_args_kwargs, **kwargs): args = [] for action_expr_args_kwargs in action_exprs_args_kwargs: action, others = action_expr_args_kwargs[0], action_expr_args_kwargs[1:] action = '_%s' % action if not action.startswith('_') else action args.append((action, ) + tuple(others)) kwargs = kwargs.copy() kwargs['batch'] = True return self._action(*args, **kwargs) def _get_cached_sub_expr(self, cached_expr, ctx=None): ctx = ctx or context data = ctx.get_cached(cached_expr) return cached_expr.get_cached(data) def _build_expr_dag(self, exprs, on_copy=None): cached_exprs = ExprDictionary() def find_cached(_, n): if context.is_cached(n) and hasattr(n, 'get_cached'): cached_exprs[n] = True if on_copy is not None: if not isinstance(on_copy, Iterable): on_copy = (on_copy, ) else: on_copy = tuple(on_copy) on_copy = on_copy + (find_cached, ) else: on_copy = find_cached res = tuple(expr.to_dag(copy=True, on_copy=on_copy, validate=False) for expr in exprs) for cached in cached_exprs: sub = self._get_cached_sub_expr(cached) if sub is not None: for dag in res: if dag.contains_node(cached): dag.substitute(cached, sub) return res def _add_node(self, dag_node, dag): nodes = dag.nodes() dag.add_node(dag_node) for node in nodes: node_expr = node.expr if dag_node.expr.is_ancestor(node_expr): dag.add_edge(node, dag_node) elif node_expr.is_ancestor(dag_node.expr): dag.add_edge(dag_node, node) @classmethod def _execute_dag(cls, dag, ui=None, async_=False, n_parallel=1, timeout=None, close_and_notify=True, progress_proportion=1.0, **kw): async_ = kw.pop('async', async_) return dag.execute(ui=ui, async_=async_, n_parallel=n_parallel, timeout=timeout, close_and_notify=close_and_notify, progress_proportion=progress_proportion) def _get_libraries(self, libraries): def conv(libs): if libs is None: return None if isinstance(libs, (six.binary_type, six.text_type, Resource)): return conv([libs, ]) new_libs = [] for lib in libs: if not isinstance(lib, (Resource, six.string_types)): raise ValueError('Resource %s not acceptable: illegal input type %s.' % (repr(lib), type(lib).__name__)) if isinstance(lib, Resource): new_libs.append(lib) elif '/' not in lib and self._odps.exist_resource(lib): new_libs.append(self._odps.get_resource(lib)) elif os.path.isfile(lib) and lib.endswith('.py'): new_libs.append(lib) elif os.path.isdir(lib): new_libs.append(lib) else: raise ValueError('Resource %s not found.' % repr(lib)) return new_libs libraries = conv(libraries) or [] if options.df.libraries is not None: libraries.extend(conv(options.df.libraries)) if len(libraries) == 0: return return list(set(libraries))

odps/df/backends/core.py (646 lines of code) (raw):