relationalai 0.13.5__py3-none-any.whl → 1.0.0a2__py3-none-any.whl

v0/relationalai/semantics/lqp/executor.py

@@ -0,0 +1,474 @@
+from __future__ import annotations
+from collections import defaultdict
+from datetime import datetime, timezone
+import atexit
+import re
+
+from pandas import DataFrame
+from typing import Any, Optional, Literal, TYPE_CHECKING
+from snowflake.snowpark import Session
+import v0.relationalai as rai
+
+from v0.relationalai import debugging
+from v0.relationalai.errors import NonDefaultLQPSemanticsVersionWarning
+from v0.relationalai.semantics.lqp import result_helpers
+from v0.relationalai.semantics.metamodel import ir, factory as f, executor as e
+from v0.relationalai.semantics.lqp.compiler import Compiler
+from v0.relationalai.semantics.lqp.intrinsics import mk_intrinsic_datetime_now
+from v0.relationalai.semantics.lqp.constructors import mk_transaction
+from v0.relationalai.semantics.lqp.types import lqp_type_to_sql
+from lqp import print as lqp_print, ir as lqp_ir
+from lqp.parser import construct_configure
+from v0.relationalai.semantics.lqp.ir import convert_transaction, validate_lqp
+from v0.relationalai.clients.config import Config
+from v0.relationalai.clients.snowflake import APP_NAME
+from v0.relationalai.clients.types import TransactionAsyncResponse
+from v0.relationalai.clients.util import IdentityParser, escape_for_f_string
+from v0.relationalai.tools.constants import USE_DIRECT_ACCESS, QUERY_ATTRIBUTES_HEADER
+from v0.relationalai.tools.query_utils import prepare_metadata_for_headers
+
+if TYPE_CHECKING:
+    from v0.relationalai.semantics.snowflake import Table
+
+# Whenever the logic engine introduces a breaking change in behaviour, we bump this version
+# once the client is ready to handle it.
+DEFAULT_LQP_SEMANTICS_VERSION = "0"
+
+class LQPExecutor(e.Executor):
+    """Executes LQP using the RAI client."""
+
+    def __init__(
+        self,
+        database: str,
+        dry_run: bool = False,
+        keep_model: bool = True,
+        wide_outputs: bool = False,
+        connection: Session | None = None,
+        config: Config | None = None,
+        # In order to facilitate snapshot testing, we allow overriding intrinsic definitions
+        # like the current time, which would otherwise change between runs.
+        intrinsic_overrides: dict = {},
+    ) -> None:
+        super().__init__()
+        self.database = database
+        self.dry_run = dry_run
+        self.keep_model = keep_model
+        self.wide_outputs = wide_outputs
+        self.compiler = Compiler()
+        self.connection = connection
+        self.config = config or Config()
+        self.intrinsic_overrides = intrinsic_overrides
+        self._resources = None
+        self._last_model = None
+        self._last_sources_version = (-1, None)
+
+    @property
+    def resources(self):
+        if not self._resources:
+            with debugging.span("create_session"):
+                self.dry_run |= bool(self.config.get("compiler.dry_run", False))
+                resource_class = rai.clients.snowflake.Resources
+                if self.config.get("use_direct_access", USE_DIRECT_ACCESS):
+                    resource_class = rai.clients.snowflake.DirectAccessResources
+                if self.config.get("platform", "") == "local":
+                    resource_class = rai.clients.local.LocalResources
+                # NOTE: language="lqp" is not strictly required for LQP execution, but it
+                # will significantly improve performance.
+                self._resources = resource_class(
+                    dry_run=self.dry_run,
+                    config=self.config,
+                    generation=rai.Generation.QB,
+                    connection=self.connection,
+                    language="lqp",
+                )
+                if not self.dry_run:
+                    self.engine = self._resources.get_default_engine_name()
+                    if not self.keep_model:
+                        atexit.register(self._resources.delete_graph, self.database, True, "lqp")
+        return self._resources
+
+    # Checks the graph index and updates it if necessary
+    def prepare_data(self):
+        # Has to happen first, so self.dry_run is populated.
+        resources = self.resources
+
+        if self.dry_run:
+            return
+
+        from v0.relationalai.semantics.snowflake import Table
+        table_sources = Table._used_sources
+        if not table_sources.has_changed(self._last_sources_version):
+            return
+
+        model = self.database
+        app_name = resources.get_app_name()
+        engine_name = self.engine
+        engine_size = self.resources.config.get_default_engine_size()
+
+        program_span_id = debugging.get_program_span_id()
+        sources = [t._fqn for t in Table._used_sources]
+        self._last_sources_version = Table._used_sources.version()
+
+        assert self.engine is not None
+
+        with debugging.span("poll_use_index", sources=sources, model=model, engine=engine_name):
+            resources.maybe_poll_use_index(
+                app_name=app_name,
+                sources=sources,
+                model=model,
+                engine_name=self.engine,
+                engine_size=engine_size,
+                program_span_id=program_span_id,
+            )
+
+    def report_errors(self, problems: list[dict[str, Any]], abort_on_error=True):
+        from v0.relationalai import errors
+        all_errors = []
+        undefineds = []
+        pyrel_errors = defaultdict(list)
+        pyrel_warnings = defaultdict(list)
+
+        for problem in problems:
+            message = problem.get("message", "")
+            report = problem.get("report", "")
+            # TODO: we need to build source maps
+            # path = problem.get("path", "")
+            # source_task = self._install_batch.line_to_task(path, problem["start_line"]) or task
+            # source = debugging.get_source(source_task) or debugging.SourceInfo()
+            source = debugging.SourceInfo()
+            severity = problem.get("severity", "warning")
+            code = problem.get("code")
+
+            if severity in ["error", "exception"]:
+                if code == "UNDEFINED_IDENTIFIER":
+                    match = re.search(r'`(.+?)` is undefined', message)
+                    if match:
+                        undefineds.append((match.group(1), source))
+                    else:
+                        all_errors.append(errors.RelQueryError(problem, source))
+                elif "overflowed" in report:
+                    all_errors.append(errors.NumericOverflow(problem, source))
+                elif code == "PYREL_ERROR":
+                    pyrel_errors[problem["props"]["pyrel_id"]].append(problem)
+                elif abort_on_error:
+                    all_errors.append(errors.RelQueryError(problem, source))
+            else:
+                if code == "ARITY_MISMATCH":
+                    errors.ArityMismatch(problem, source)
+                elif code == "IC_VIOLATION":
+                    all_errors.append(errors.IntegrityConstraintViolation(problem, source))
+                elif code == "PYREL_ERROR":
+                    pyrel_warnings[problem["props"]["pyrel_id"]].append(problem)
+                else:
+                    errors.RelQueryWarning(problem, source)
+
+        if abort_on_error and len(undefineds):
+            all_errors.append(errors.UninitializedPropertyException(undefineds))
+
+        if abort_on_error:
+            for pyrel_id, pyrel_problems in pyrel_errors.items():
+                all_errors.append(errors.ModelError(pyrel_problems))
+
+        for pyrel_id, pyrel_problems in pyrel_warnings.items():
+            errors.ModelWarning(pyrel_problems)
+
+
+        if len(all_errors) == 1:
+            raise all_errors[0]
+        elif len(all_errors) > 1:
+            raise errors.RAIExceptionSet(all_errors)
+
+    def _export(self, txn_id: str, export_info: tuple, dest: Table, actual_cols: list[str], declared_cols: list[str], update: bool):
+        # At this point of the export, we assume that a CSV file has already been written
+        # to the Snowflake Native App stage area. Thus, the purpose of this method is to
+        # copy the data from the CSV file to the destination table.
+        _exec = self.resources._exec
+        dest_database, dest_schema, dest_table, _ = IdentityParser(dest._fqn, require_all_parts=True).to_list()
+        filename = export_info[0]
+        result_table_name = filename + "_table"
+
+        with debugging.span("export", txn_id=txn_id, export_info=export_info, dest_table=dest_table):
+            with debugging.span("export_to_result_schema"):
+                # First, we need to persist from the CSV file to the results schema by calling the
+                # `persist_from_stage` stored procedure. This step also cleans up the CSV file in
+                # the stage area.
+                column_fields = []
+                for (col_name, col_type) in export_info[1]:
+                    column_fields.append([col_name, lqp_type_to_sql(col_type)])
+
+                # NOTE: the `str(column_fields)` depends on python formatting which surrounds
+                # strings with single quotes. If this changes, or if we ever get a single quote in
+                # the actual string, then we need to do something more sophisticated.
+                exec_str = f"call {APP_NAME}.api.persist_from_stage('{txn_id}', '{filename}', '{result_table_name}', {str(column_fields)})"
+                _exec(exec_str)
+
+            with debugging.span("write_table"):
+                # The result of the first step above is a table in the results schema,
+                # {app_name}.results.{result_table_name}.
+                # Second, we need to copy the data from the results schema to the
+                # destination table. This step also cleans up the result table.
+                out_sample = _exec(f"select * from {APP_NAME}.results.{result_table_name} limit 1;")
+                names = self._build_projection(declared_cols, actual_cols, column_fields, out_sample)
+                dest_fqn = dest._fqn
+                try:
+                    if not update:
+                        createTableLogic = f"""
+                                        CREATE TABLE {dest_fqn} AS
+                                        SELECT {names}
+                                        FROM {APP_NAME}.results.{result_table_name};
+                        """
+                        if dest._is_iceberg:
+                            assert dest._iceberg_config is not None
+                            external_volume_clause = ""
+                            if dest._iceberg_config.external_volume:
+                                external_volume_clause = f"EXTERNAL_VOLUME = '{dest._iceberg_config.external_volume}'"
+                            createTableLogic = f"""
+                                        CREATE ICEBERG TABLE {dest_fqn}
+                                        CATALOG = "SNOWFLAKE"
+                                        {external_volume_clause}
+                                        AS
+                                        SELECT {names}
+                                        FROM {APP_NAME}.results.{result_table_name};
+                            """
+
+                        _exec(f"""
+                            BEGIN
+                                -- Check if table exists
+                                IF (EXISTS (
+                                    SELECT 1
+                                    FROM {dest_database}.INFORMATION_SCHEMA.TABLES
+                                    WHERE table_schema = '{dest_schema}'
+                                    AND table_name = '{dest_table}'
+                                )) THEN
+                                    -- Insert into existing table
+                                    EXECUTE IMMEDIATE '
+                                        BEGIN
+                                            TRUNCATE TABLE {dest_fqn};
+                                            INSERT INTO {dest_fqn}
+                                            SELECT {names}
+                                            FROM {APP_NAME}.results.{result_table_name}
+                                            {'' if out_sample else 'WHERE 1=0'};
+                                        END;
+                                    ';
+                                ELSE
+                                    -- Create table based on the SELECT
+                                    EXECUTE IMMEDIATE '
+                                        {escape_for_f_string(createTableLogic)}
+                                    ';
+                                END IF;
+                            END;
+                        """)
+                    else:
+                        if out_sample:
+                            _exec(f"""
+                                BEGIN
+                                    INSERT INTO {dest_fqn}
+                                    SELECT {names}
+                                    FROM {APP_NAME}.results.{result_table_name};
+                                END;
+                            """)
+                finally:
+                    # Always try to drop the result table, even if the insert/create failed.
+                    _exec(f"call {APP_NAME}.api.drop_result_table('{result_table_name}');")
+
+    def _build_projection(self, declared_cols, actual_cols, column_fields, out_sample=None):
+        # map physical col -> type
+        col_type_map = {col.lower(): dtype for col, dtype in column_fields}
+        sample_keys = {k.lower() for k in out_sample[0].as_dict()} if out_sample else set()
+
+        fields = []
+        ix = 0
+
+        for name in declared_cols:
+            if name not in actual_cols:
+                # Declared but not present in results
+                fields.append(f"NULL as \"{name}\"")
+                continue
+
+            colname = f"col{ix:03}"
+            ix += 1
+
+            if colname in sample_keys:
+                # Actual column exists in sample
+                fields.append(f"{colname} as \"{name}\"")
+            else:
+                # No sample or missing key → fall back to type
+                dtype = col_type_map.get(colname.lower(), "VARCHAR")
+                fields.append(f"CAST(NULL AS {dtype}) as \"{name}\"")
+
+        return ", ".join(fields)
+
+    def _construct_configure(self):
+        config_dict = {}
+        # Only set the IVM flag if there is a value in `config`. Otherwise, let
+        # `construct_configure` set the default value.
+        ivm_flag = self.config.get('reasoner.rule.incremental_maintenance', None)
+        if ivm_flag:
+            config_dict['ivm.maintenance_level'] = lqp_ir.Value(value=ivm_flag, meta=None)
+
+        # Set semantics_version from config, defaulting to 0
+        semantics_version: str | Any = self.config.get('reasoner.rule.lqp.semantics_version', DEFAULT_LQP_SEMANTICS_VERSION)
+        config_dict['semantics_version'] = lqp_ir.Value(value=int(semantics_version), meta=None)
+
+        # Warn if a non-default semantics version is used. Most likely, this is due to a
+        # user manually reverting to an older version. We want them to not get stuck on that
+        # version for longer than necessary.
+        if semantics_version != DEFAULT_LQP_SEMANTICS_VERSION:
+            debugging.warn(NonDefaultLQPSemanticsVersionWarning(semantics_version, DEFAULT_LQP_SEMANTICS_VERSION))
+
+        return construct_configure(config_dict, None)
+
+    def _should_sync(self, model) :
+        if self._last_model != model:
+            return lqp_ir.Sync(fragments=[], meta=None)
+        else :
+            return None
+
+    def _compile_intrinsics(self) -> lqp_ir.Epoch:
+        """Construct an epoch that defines a number of built-in definitions used by the
+        emitter."""
+        with debugging.span("compile_intrinsics") as span:
+            span["compile_type"] = "intrinsics"
+
+            now = self.intrinsic_overrides.get('datetime_now', datetime.now(timezone.utc))
+
+            debug_info = lqp_ir.DebugInfo(id_to_orig_name={}, meta=None)
+            intrinsics_fragment = lqp_ir.Fragment(
+                id = lqp_ir.FragmentId(id=b"__pyrel_lqp_intrinsics", meta=None),
+                declarations = [mk_intrinsic_datetime_now(now)],
+                debug_info = debug_info,
+                meta = None,
+            )
+
+            return lqp_ir.Epoch(
+                writes=[
+                    lqp_ir.Write(write_type=lqp_ir.Define(fragment=intrinsics_fragment, meta=None), meta=None)
+                ],
+                meta=None,
+            )
+
+    # [RAI-40997] We eagerly undefine query fragments so they are not committed to storage
+    def _compile_undefine_query(self, query_epoch: lqp_ir.Epoch) -> lqp_ir.Epoch:
+        fragment_ids = []
+
+        for write in query_epoch.writes:
+            if isinstance(write.write_type, lqp_ir.Define):
+                fragment_ids.append(write.write_type.fragment.id)
+
+        # Construct new Epoch with Undefine operations for all collected fragment IDs
+        undefine_writes = [
+            lqp_ir.Write(
+                write_type=lqp_ir.Undefine(fragment_id=frag_id, meta=None),
+                meta=None
+            )
+            for frag_id in fragment_ids
+        ]
+
+        return lqp_ir.Epoch(
+            writes=undefine_writes,
+            meta=None,
+        )
+
+    def compile_lqp(self, model: ir.Model, task: ir.Task):
+        configure = self._construct_configure()
+        # Merge the epochs into a single transaction. Long term the query bits should all
+        # go into a WhatIf action and the intrinsics could be fused with either of them. But
+        # for now we just use separate epochs.
+        epochs = []
+        epochs.append(self._compile_intrinsics())
+
+        sync = self._should_sync(model)
+
+        if sync is not None:
+            with debugging.span("compile", metamodel=model) as install_span:
+                install_span["compile_type"] = "model"
+                _, model_epoch = self.compiler.compile(model, {"fragment_id": b"model"})
+                epochs.append(model_epoch)
+                self._last_model = model
+
+        with debugging.span("compile", metamodel=task) as txn_span:
+            query = f.compute_model(f.logical([task]))
+            options = {
+                "wide_outputs": self.wide_outputs,
+                "fragment_id": b"query",
+            }
+            result, final_model = self.compiler.compile_inner(query, options)
+            export_info, query_epoch = result
+
+            epochs.append(query_epoch)
+            epochs.append(self._compile_undefine_query(query_epoch))
+
+            txn_span["compile_type"] = "query"
+            txn = mk_transaction(epochs=epochs, configure=configure, sync=sync)
+            txn_span["lqp"] = lqp_print.to_string(txn, {"print_names": True, "print_debug": False, "print_csv_filename": False})
+
+        validate_lqp(txn)
+
+        txn_proto = convert_transaction(txn)
+        return final_model, export_info, txn_proto
+
+    # TODO (azreika): This should probably be split up into exporting and other processing. There are quite a lot of arguments here...
+    def _process_results(self, task: ir.Task, final_model: ir.Model, raw_results: TransactionAsyncResponse, export_info: Optional[tuple], export_to: Optional[Table], update: bool) -> DataFrame:
+        cols, extra_cols = self._compute_cols(task, final_model)
+
+        df, errs = result_helpers.format_results(raw_results, cols)
+        self.report_errors(errs)
+
+        # Rename columns if wide outputs is enabled
+        if self.wide_outputs and len(cols) - len(extra_cols) == len(df.columns):
+            df.columns = cols[: len(df.columns)]
+
+        # Process exports
+        if export_to and not self.dry_run:
+            assert cols, "No columns found in the output"
+            assert isinstance(raw_results, TransactionAsyncResponse) and raw_results.transaction, "Invalid transaction result"
+
+            result_cols = export_to._col_names
+
+            if result_cols is not None:
+                assert all(col in result_cols or col in extra_cols for col in cols)
+            else:
+                result_cols = [col for col in cols if col not in extra_cols]
+            assert result_cols
+
+            assert export_info, "Export info should be populated if we are exporting results"
+            self._export(raw_results.transaction['id'], export_info, export_to, cols, result_cols, update)
+
+        return self._postprocess_df(self.config, df, extra_cols)
+
+    def execute(self, model: ir.Model, task: ir.Task, format: Literal["pandas", "snowpark"] = "pandas",
+                export_to: Optional[Table] = None,
+                update: bool = False, meta: dict[str, Any] | None = None) -> DataFrame:
+        self.prepare_data()
+        previous_model = self._last_model
+
+        final_model, export_info, txn_proto = self.compile_lqp(model, task)
+
+        if self.dry_run:
+            return DataFrame()
+
+        if format != "pandas":
+            raise ValueError(f"Unsupported format: {format}")
+
+        # Format meta as headers
+        json_meta = prepare_metadata_for_headers(meta)
+        headers = {QUERY_ATTRIBUTES_HEADER: json_meta} if json_meta else {}
+        raw_results = self.resources.exec_lqp(
+            self.database,
+            self.engine,
+            txn_proto.SerializeToString(),
+            # Current strategy is to run all queries as write transactions, in order to
+            # benefit from view caching. This will have to be revisited, because write
+            # transactions are serialized.
+            readonly=False,
+            nowait_durable=True,
+            headers=headers,
+        )
+        assert isinstance(raw_results, TransactionAsyncResponse)
+
+        try:
+            return self._process_results(task, final_model, raw_results, export_info, export_to, update)
+        except Exception as e:
+            # If processing the results failed, revert to the previous model.
+            self._last_model = previous_model
+            raise e

v0/relationalai/semantics/lqp/intrinsics.py

@@ -0,0 +1,24 @@
+from datetime import datetime
+
+from v0.relationalai.semantics.lqp import ir as lqp
+from v0.relationalai.semantics.lqp.constructors import mk_abstraction, mk_value, mk_type, mk_primitive
+from v0.relationalai.semantics.lqp.utils import lqp_hash
+
+# Constructs a definition of the current datetime.
+def mk_intrinsic_datetime_now(dt: datetime) -> lqp.Def:
+    """Constructs a definition of the current datetime."""
+    id = lqp_hash("__pyrel_lqp_intrinsic_datetime_now")
+    out = lqp.Var(name="out", meta=None)
+    out_type = mk_type(lqp.TypeName.DATETIME)
+    now = mk_value(lqp.DateTimeValue(value=dt, meta=None))
+    datetime_now = mk_abstraction(
+        [(out, out_type)],
+        mk_primitive("rel_primitive_eq", [out, now]),
+    )
+
+    return lqp.Def(
+        name = lqp.RelationId(id=id, meta=None),
+        body = datetime_now,
+        attrs = [],
+        meta = None,
+    )

v0/relationalai/semantics/lqp/ir.py

@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+__all__ = [
+    "SourceInfo",
+    "LqpNode",
+    "Declaration",
+    "FunctionalDependency",
+    "Def",
+    "Loop",
+    "Abstraction",
+    "Formula",
+    "Exists",
+    "Reduce",
+    "Conjunction",
+    "Disjunction",
+    "Not",
+    "FFI",
+    "Atom",
+    "Pragma",
+    "Primitive",
+    "RelAtom",
+    "Cast",
+    "Var",
+    "Int128Value",
+    "UInt128Value",
+    "DecimalValue",
+    "DateValue",
+    "DateTimeValue",
+    "BooleanValue",
+    "Term",
+    "Value",
+    "SpecializedValue",
+    "RelTerm",
+    "Attribute",
+    "RelationId",
+    "Type",
+    "TypeName",
+    "FragmentId",
+    "Fragment",
+    "Define",
+    "Write",
+    "Output",
+    "Export",
+    "ExportCSVConfig",
+    "ExportCSVColumn",
+    "Read",
+    "Epoch",
+    "Transaction",
+    "Sync",
+    "DebugInfo",
+    "Configure",
+    "IVMConfig",
+    "MaintenanceLevel",
+    "ir_to_proto",
+    "convert_transaction",
+    "validate_lqp",
+    "construct_configure",
+]
+
+from lqp.ir import (
+    SourceInfo,
+    LqpNode,
+    Declaration,
+    FunctionalDependency,
+    Def,
+    Loop,
+    Abstraction,
+    Formula,
+    Exists,
+    Reduce,
+    Conjunction,
+    Disjunction,
+    Not,
+    FFI,
+    Atom,
+    Pragma,
+    Primitive,
+    RelAtom,
+    Cast,
+    Var,
+    Int128Value,
+    UInt128Value,
+    DecimalValue,
+    DateValue,
+    DateTimeValue,
+    BooleanValue,
+    Term,
+    Value,
+    SpecializedValue,
+    RelTerm,
+    Attribute,
+    RelationId,
+    Type,
+    TypeName,
+    FragmentId,
+    Fragment,
+    Define,
+    Write,
+    Output,
+    Export,
+    ExportCSVConfig,
+    ExportCSVColumn,
+    Read,
+    Epoch,
+    Transaction,
+    Sync,
+    DebugInfo,
+    Configure,
+    IVMConfig,
+    MaintenanceLevel,
+)
+
+from lqp.emit import (
+    ir_to_proto,
+    convert_transaction,
+)
+
+from lqp.validator import (
+    validate_lqp
+)
+
+from lqp.parser import (
+    construct_configure
+)