relationalai 0.13.0.dev0__py3-none-any.whl → 0.13.1__py3-none-any.whl

relationalai/semantics/lqp/model2lqp.py

@@ -0,0 +1,877 @@
+from relationalai.semantics.metamodel import ir, builtins, helpers, types
+from relationalai.semantics.metamodel.visitor import collect_by_type
+from relationalai.semantics.metamodel.util import FrozenOrderedSet
+from relationalai.semantics.lqp import ir as lqp, utils, types as lqp_types, builtins as lqp_builtins
+from relationalai.semantics.lqp.primitives import lqp_avg_op, lqp_operator, build_primitive
+from relationalai.semantics.lqp.pragmas import pragma_to_lqp_name
+from relationalai.semantics.lqp.types import meta_type_to_lqp
+from relationalai.semantics.lqp.constructors import (
+    mk_abstraction, mk_and, mk_exists, mk_or, mk_pragma, mk_primitive,
+    mk_specialized_value, mk_type, mk_value, mk_attribute
+)
+from relationalai.semantics.lqp.utils import TranslationCtx, gen_unique_var
+from relationalai.semantics.lqp.validators import assert_valid_input
+from relationalai.semantics.lqp.rewrite.functional_dependencies import (
+    normalized_fd, contains_only_declarable_constraints
+)
+from decimal import Decimal as PyDecimal
+from datetime import datetime, date, timezone
+from typing import Tuple, cast, Union, Optional
+from warnings import warn
+import re
+import uuid
+
+# Main access point for translating metamodel to lqp. Converts the model IR to an LQP epoch.
+def to_lqp(model: ir.Model, fragment_name: bytes, ctx: TranslationCtx) -> tuple[Optional[tuple], lqp.Epoch]:
+    assert_valid_input(model)
+    decls: list[lqp.Declaration] = []
+    reads: list[lqp.Read] = []
+
+    # LQP only accepts logical tasks
+    # These are asserted at init time
+    root = cast(ir.Logical, model.root)
+    for subtask in root.body:
+        assert isinstance(subtask, ir.Logical)
+        new_decls = _translate_to_decls(ctx, subtask)
+        decls.extend(new_decls)
+
+    # Add pyrel error attributes to reads.
+    for err_id in _extract_pyrel_error_ids(ctx, model):
+        ctx.output_ids.append(err_id)
+
+    reads.extend(_get_output_reads(ctx.output_ids))
+
+    export_info = None
+    if len(ctx.export_ids) > 0:
+        export_filename, col_types, export_reads = _get_export_reads(ctx.export_ids)
+        reads.extend(export_reads)
+        export_info = (export_filename, col_types)
+
+    debug_info = lqp.DebugInfo(id_to_orig_name=ctx.rel_id_to_orig_name, meta=None)
+    fragment_id = lqp.FragmentId(id=fragment_name, meta=None)
+    fragment = lqp.Fragment(id=fragment_id, declarations=decls, meta=None, debug_info=debug_info)
+    define_op = lqp.Define(fragment=fragment, meta=None)
+
+    txn = lqp.Epoch(
+        reads=reads,
+        writes=[lqp.Write(write_type=define_op, meta=None)],
+        meta=None
+    )
+
+    return (export_info, txn)
+
+def _effect_bindings(effect: Union[ir.Output, ir.Update]) -> list[ir.Value]:
+    if isinstance(effect, ir.Output):
+        # Unions may not return anything. The generated IR contains a None value when this
+        # happens. We ignore it here.
+        return [v for v in helpers.output_values(effect.aliases) if v]
+    else:
+        return list(effect.args)
+
+def _get_output_reads(output_ids: list[tuple[lqp.RelationId, str]]) -> list[lqp.Read]:
+    reads = []
+    for (rel_id, name) in output_ids:
+        assert isinstance(rel_id, lqp.RelationId)
+        output = lqp.Output(name=name, relation_id=rel_id, meta=None)
+        reads.append(lqp.Read(read_type=output, meta=None))
+    return reads
+
+def _get_export_reads(export_ids: list[tuple[lqp.RelationId, int, lqp.Type]]) -> tuple[str, list, list[lqp.Read]]:
+    reads = []
+    csv_columns = []
+    col_info = []
+    for (col_id, col_num, col_type) in sorted(export_ids, key=lambda x: x[1]):
+        col_name = f"col{col_num:03}"
+        csv_columns.append(lqp.ExportCSVColumn(column_name=col_name, column_data=col_id, meta=None))
+        col_info.append((col_name, col_type))
+
+    # Generate a random name for the internal export path
+    export_filename = "export_" + str(uuid.uuid4()).replace("-", "_")
+
+    # Note that the engine will append the transaction id to the export path for access
+    # control reasons. So the actual final path will be
+    # `.../export/{export_filename}/data_{txn_id}`
+    export_path = f"snowflake://APP_STATE.RAI_INTERNAL_STAGE/export/{export_filename}/data"
+    export_csv_config = lqp.ExportCSVConfig(
+        path=export_path,
+        data_columns=csv_columns,
+        compression="gzip",
+        partition_size=200,
+        syntax_escapechar='"',  # To follow Snowflake's expected format
+        meta=None,
+    )
+    reads.append(lqp.Read(read_type=lqp.Export(config=export_csv_config, meta=None), meta=None))
+    return (export_filename, col_info, reads)
+
+def _translate_to_decls(ctx: TranslationCtx, rule: ir.Logical) -> list[lqp.Declaration]:
+    if contains_only_declarable_constraints(rule):
+        return _translate_to_constraint_decls(ctx, rule)
+    else:
+        return _translate_to_standard_decl(ctx, rule)
+
+def _translate_to_constraint_decls(ctx: TranslationCtx, rule: ir.Logical) -> list[lqp.Declaration]:
+    constraint_decls: list[lqp.Declaration] = []
+    for task in rule.body:
+        assert isinstance(task, ir.Require)
+        fd = normalized_fd(task)
+        assert fd is not None
+
+        # check for unresolved types
+        if any(types.is_any(var.type) for var in fd.keys + fd.values):
+            warn(f"Ignoring FD with unresolved type: {fd}")
+            continue
+
+        lqp_typed_keys = [_translate_term(ctx, key) for key in fd.keys]
+        lqp_typed_values = [_translate_term(ctx, value) for value in fd.values]
+        lqp_typed_vars:list[Tuple[lqp.Var, lqp.Type]] = lqp_typed_keys + lqp_typed_values # type: ignore
+        lqp_guard_atoms = [_translate_to_atom(ctx, atom) for atom in fd.guard]
+        lqp_guard = mk_abstraction(lqp_typed_vars, mk_and(lqp_guard_atoms))
+        lqp_keys:list[lqp.Var] = [var for (var, _) in lqp_typed_keys] # type: ignore
+        lqp_values:list[lqp.Var] = [var for (var, _) in lqp_typed_values] # type: ignore
+
+        fd_decl = lqp.FunctionalDependency(
+            guard=lqp_guard,
+            keys=lqp_keys,
+            values=lqp_values,
+            meta=None
+        )
+
+        constraint_decls.append(fd_decl)
+
+    return constraint_decls
+
+def _translate_to_standard_decl(ctx: TranslationCtx, rule: ir.Logical) -> list[lqp.Declaration]:
+    effects = collect_by_type((ir.Output, ir.Update), rule)
+    aggregates = collect_by_type(ir.Aggregate, rule)
+    ranks = collect_by_type(ir.Rank, rule)
+
+    # TODO: should this ever actually come in as input?
+    if len(effects) == 0:
+        return []
+
+    assert len(ranks) == 0 or len(aggregates) == 0, "rules cannot have both aggregates and ranks"
+
+    conjuncts = []
+    for task in rule.body:
+        if isinstance(task, (ir.Output, ir.Update)):
+            continue
+        conjuncts.append(_translate_to_formula(ctx, task))
+
+    # Aggregates reduce over the body
+    if aggregates or ranks:
+        aggr_body = mk_and(conjuncts)
+        conjuncts = []
+        for aggr in aggregates:
+            conjuncts.append(_translate_aggregate(ctx, aggr, aggr_body))
+        for rank in ranks:
+            conjuncts.append(_translate_rank(ctx, rank, aggr_body))
+
+    return [_translate_effect(ctx, effect, mk_and(conjuncts)) for effect in effects]
+
+def _translate_annotations(annotations: FrozenOrderedSet[ir.Annotation]) -> list[lqp.Attribute]:
+    attributes = []
+    for annotation in annotations:
+
+        if annotation.relation.name in lqp_builtins.annotations_to_emit:
+            if any(not isinstance(a, ir.Literal) for a in annotation.args):
+                warn("LQP currently ignores annotation parameters with non-literal values")
+                continue
+
+            # Convert literal arguments to LQP values
+            args = []
+            for a in annotation.args:
+                assert isinstance(a, ir.Literal)
+                args.append(mk_value(a.value))
+            attributes.append(mk_attribute(annotation.relation.name, args))
+    return attributes
+
+# Translates an effect (export, output, or update) into the corresponding def. Note that
+# this method only generates the def and not the LQP read operation, which is added later by
+# `_get_output_reads` and `_get_export_reads`.
+def _translate_effect(ctx: TranslationCtx, effect: Union[ir.Output, ir.Update], body: lqp.Formula) -> lqp.Declaration:
+    bindings = _effect_bindings(effect)
+
+    def _is_export(e):
+        return isinstance(e, ir.Output) and builtins.export_annotation in e.annotations
+
+    if isinstance(effect, ir.Output):
+        projection, eqs, suffix = _translate_output_bindings(ctx, bindings)
+        meta_id = effect.id
+
+        if _is_export(effect):
+            def_name = "export_relation" + suffix
+        else:
+            def_name = "output" + suffix
+
+        def_name = ctx.output_names.get_name_by_id(meta_id, def_name)
+        rel_id = get_output_id(ctx, def_name, meta_id)
+    else:
+        projection, eqs = _translate_bindings(ctx, bindings)
+        def_name = effect.relation.name
+        rel_id = get_relation_id(ctx, effect.relation, projection)
+
+    eqs.append(body)
+    new_body = mk_and(eqs)
+
+    # Context bookkeeping for exports and outputs
+    if _is_export(effect):
+        # The row id is the first n-1 elements, and the actual data is the last element. Its
+        # type is stored in the first element of the tuple.
+        col_type = projection[-1][1]
+        _col_num_match = re.search(r"export_relation_col([0-9]+)", def_name)
+        assert _col_num_match, f"Could not find column number in suffix: {def_name}"
+        col_num = int(_col_num_match.group(1))
+        ctx.export_ids.append((rel_id, col_num, col_type))
+    elif isinstance(effect, ir.Output):
+        ctx.output_ids.append((rel_id, def_name))
+
+    # First we collect annotations on the effect itself, e.g. from something like
+    # `select(...).annotate(...)`.
+    annotations = effect.annotations
+    if isinstance(effect, ir.Update):
+        # Then we translate annotations on the relation itself, e.g.
+        # ```
+        # Bar.foo = model.Relationship(...)
+        # Bar.foo.annotate(...)
+        # ```
+        annotations = annotations | effect.relation.annotations
+
+    return lqp.Def(
+        name = rel_id,
+        body = mk_abstraction(projection, new_body),
+        attrs = _translate_annotations(annotations),
+        meta = None,
+    )
+
+def _translate_output_bindings(ctx: TranslationCtx, bindings: list[ir.Value]) -> Tuple[list[Tuple[lqp.Var, lqp.Type]], list[lqp.Formula], str]:
+    symbol_literals = []
+    non_symbols = []
+    for binding in bindings:
+        if isinstance(binding, ir.Literal) and binding.type == types.Symbol:
+            symbol_literals.append(binding.value)
+        else:
+            non_symbols.append(binding)
+    projection, eqs = _translate_bindings(ctx, non_symbols)
+    if len(symbol_literals) > 0:
+        name_suffix = "_"
+        name_suffix += "_".join(symbol_literals)
+    else:
+        name_suffix = ""
+
+    return projection, eqs, name_suffix
+
+def _translate_rank(ctx: TranslationCtx, rank: ir.Rank, body: lqp.Formula) -> lqp.Formula:
+    # Ascending rank is constructed using rel_primitive_sort. If a limit is added to an
+    # ascending rank we can use rel_primitive_top for an efficient evaluation.
+    #
+    # Descending rank is constructed as an ascending sort, then the ascending rank is
+    # subtracted from a count of the elements (plus 1 so that we still start at 1).
+    # Adding a limit to a descending rank is done by adding a filter of rank <= limit.
+
+    # Limits are the sort plus a filter on rank <= limit.
+    if all(o for o in rank.arg_is_ascending):
+        ascending = True
+    elif all(not o for o in rank.arg_is_ascending):
+        ascending = False
+    else:
+        raise Exception("Mixed orderings in rank are not supported yet.")
+
+    # Filter out the group-by variables, since they are introduced outside the rank.
+    input_args, input_eqs = _translate_bindings(ctx, list(rank.args))
+    introduced_meta_projs = [arg for arg in rank.projection if arg not in rank.group and arg not in rank.args]
+    projected_args, projected_eqs = _translate_bindings(ctx, list(introduced_meta_projs))
+
+    # rank expects an Int128 result, but the primitive will return an Int64 result.
+    # we need to set up an intermediary variable to hold the Int64 result, and a cast
+    # to convert it to Int128.
+    result_var, _ = _translate_term(ctx, rank.result)
+    # The primitive will return an Int64 result, so we need a var to hold the intermediary.
+    result_64_var = gen_unique_var(ctx, "v_rank")
+    result_64_type = mk_type(lqp.TypeName.INT)
+
+    cast = lqp.Cast(input=result_64_var, result=result_var, meta=None)
+
+    body = mk_and([body] + input_eqs + projected_eqs)
+    abstr_args = input_args + projected_args
+
+    if ascending:
+        ranker = _translate_ascending_rank(ctx, rank.limit, result_64_var, body, abstr_args)
+    else:
+        ranker = _translate_descending_rank(ctx, rank.limit, result_64_var, body, abstr_args)
+
+    return mk_exists([(result_64_var, result_64_type)], mk_and([ranker, cast]))
+
+def _translate_descending_rank(ctx: TranslationCtx, limit: int, result: lqp.Var, body: lqp.Formula, abstr_args) -> lqp.Formula:
+    result_var = result
+    result_type = mk_type(lqp.TypeName.INT)
+
+    # Rename abstracted args in the body to new variable names
+    var_map = {var.name: gen_unique_var(ctx, 't_' + var.name) for (var, _) in abstr_args}
+    body = utils.rename_vars_formula(body, var_map)
+    new_abstr_args = [(var_map[var.name], typ) for (var, typ) in abstr_args]
+
+    # Construct a conjunction of the ranking, a counter for the body, a subtraction
+    # of the rank from the count and an addition of 1. Wrap this in an abstraction.
+    count_res = gen_unique_var(ctx, "count_res")
+
+    # Add one to the count to account for the rank starting at 1.
+    one, one_eq = constant_to_var(ctx, to_lqp_value(1, types.Int64), "one")
+    one_bigger = gen_unique_var(ctx, "one_bigger")
+    addition = mk_primitive("rel_primitive_add_monotype", [count_res, one, one_bigger])
+
+    # Subtract the rank from the count + 1
+    asc_rank = gen_unique_var(ctx, "asc_rank")
+    subtraction = mk_primitive("rel_primitive_subtract_monotype", [one_bigger, result_var, asc_rank])
+
+    # Construct the ranking
+    desc_ranking_terms = [asc_rank] + [v[0] for v in abstr_args]
+    ranking = lqp.FFI(
+        meta=None,
+        name="rel_primitive_sort",
+        args=[mk_abstraction(new_abstr_args, body)],
+        terms=desc_ranking_terms,
+    )
+
+    # Count the number of rows in the body
+    count_type = meta_type_to_lqp(types.Int64)
+    count_var, count_eq = constant_to_var(ctx, to_lqp_value(1, types.Int64), "counter")
+    desc_body = mk_and([body, count_eq])
+    aggr_abstr_args = new_abstr_args + [(count_var, count_type)]
+    count_aggr = lqp.Reduce(
+        op=lqp_operator(
+            ctx,
+            "count",
+            "count",
+            mk_type(lqp.TypeName.INT)
+        ),
+        body=mk_abstraction(aggr_abstr_args, desc_body),
+        terms=[count_res],
+        meta=None
+    )
+
+    # Bring it all together and do the maths.
+    ranking = mk_exists(
+        vars=[
+            (asc_rank, result_type),
+            (count_res, result_type),
+            (one, result_type),
+            (one_bigger, result_type)
+        ],
+        value=mk_and([ranking, count_aggr, one_eq, addition, subtraction])
+    )
+
+    # If there is a limit, we need to add a filter to the ranking.
+    # Wrap with a rank <= limit
+    if limit != 0:
+        limit_term, _ = _translate_term(ctx, ir.Literal(types.Int64, limit))
+        limiter = mk_primitive("rel_primitive_lt_eq_monotype", [result_var, limit_term])
+        ranking = mk_and([ranking, limiter])
+
+    return ranking
+
+def _translate_ascending_rank(ctx: TranslationCtx, limit: int, result_var: lqp.Var, body: lqp.Formula, abstr_args) -> lqp.Formula:
+    terms = [result_var] + [v[0] for v in abstr_args]
+
+    # Rename abstracted args in the body to new variable names
+    var_map = {var.name: gen_unique_var(ctx, 't_' + var.name) for (var, _) in abstr_args}
+    body = utils.rename_vars_formula(body, var_map)
+    new_abstr_args = [(var_map[var.name], typ) for (var, typ) in abstr_args]
+    sort_abstr = mk_abstraction(new_abstr_args, body)
+
+    if limit == 0:
+        return lqp.FFI(
+            meta=None,
+            name="rel_primitive_sort",
+            args=[sort_abstr],
+            terms=terms,
+        )
+    else:
+        limit_type = meta_type_to_lqp(types.Int64)
+        limit_var, limit_eq = constant_to_var(ctx, to_lqp_value(limit, types.Int64), "limit")
+        limit_abstr = mk_abstraction([(limit_var, limit_type)], limit_eq)
+        return lqp.FFI(
+            meta=None,
+            name="rel_primitive_top",
+            args=[sort_abstr, limit_abstr],
+            terms=terms,
+        )
+
+def _rename_shadowed_abstraction_vars(
+    ctx: TranslationCtx,
+    aggr: ir.Aggregate,
+    abstr_args: list[Tuple[lqp.Var, lqp.Type]],
+    body_conjs: list[lqp.Formula]
+) -> list[Tuple[lqp.Var, lqp.Type]]:
+    """
+    Rename abstraction variables that shadow group-by variables.
+
+    This can happen when the same variable appears in both aggr.group and as an input
+    to the aggregation, e.g., min(Person.age).per(Person.age). The group-by variables
+    are in the outer scope, while the abstraction parameters are in the inner scope,
+    so we need different names to avoid shadowing.
+    """
+    # Get the LQP names of group-by variables
+    group_var_names = set()
+    for group_var in aggr.group:
+        lqp_var = _translate_var(ctx, group_var)
+        group_var_names.add(lqp_var.name)
+
+    # Rename any abstraction parameters that conflict with group-by variables
+    renamed_abstr_args = []
+    for var, typ in abstr_args:
+        if var.name in group_var_names:
+            # This variable shadows a group-by variable, so rename it
+            fresh_var = gen_unique_var(ctx, var.name)
+            # Add an equality constraint: fresh_var == var
+            # var is a free variable referring to the outer scope group-by variable
+            body_conjs.append(mk_primitive("rel_primitive_eq", [fresh_var, var]))
+            renamed_abstr_args.append((fresh_var, typ))
+        else:
+            renamed_abstr_args.append((var, typ))
+
+    return renamed_abstr_args
+
+def _translate_aggregate(ctx: TranslationCtx, aggr: ir.Aggregate, body: lqp.Formula) -> Union[lqp.Reduce, lqp.Formula]:
+    # TODO: handle this properly
+    aggr_name = aggr.aggregation.name
+    supported_aggrs = ("sum", "count", "avg", "min", "max", "rel_primitive_solverlib_ho_appl")
+    assert aggr_name in supported_aggrs, f"only support {supported_aggrs} for now, not {aggr.aggregation.name}"
+
+    meta_output_terms = []
+    meta_input_terms = []
+
+    for (field, arg) in zip(aggr.aggregation.fields, aggr.args):
+        if field.input:
+            meta_input_terms.append(arg)
+        else:
+            meta_output_terms.append(arg)
+
+    output_terms = [_translate_term(ctx, term) for term in meta_output_terms]
+    output_vars = [term[0] for term in output_terms]
+
+    body_conjs = [body]
+    input_args, input_eqs = _translate_bindings(ctx, meta_input_terms)
+
+    # TODO: Can this safely be applied to all aggregates?
+    if aggr_name in ("sum", "min", "max"):
+        assert len(output_terms) == 1, f"{aggr_name} expects a single output variable"
+        assert len(meta_input_terms) == 1, f"{aggr_name} expects a single input variable"
+        assert isinstance(meta_output_terms[0], ir.Var)
+        assert input_args[0][1] == output_terms[0][1], f"{aggr_name}({input_args[0][1].type_name}) had output type of {output_terms[0][1].type_name}"
+
+    # Filter out the group-by variables, since they are introduced outside the aggregation.
+    # Input terms are added later below.
+    introduced_meta_projs = [arg for arg in aggr.projection if arg not in aggr.group and arg not in meta_input_terms]
+    projected_args, projected_eqs = _translate_bindings(ctx, list(introduced_meta_projs))
+    body_conjs.extend(input_eqs)
+    body_conjs.extend(projected_eqs)
+    abstr_args: list[Tuple[lqp.Var, lqp.Type]] = projected_args + input_args
+
+    # Rename abstraction variables that shadow group-by variables
+    abstr_args = _rename_shadowed_abstraction_vars(ctx, aggr, abstr_args, body_conjs)
+
+    if aggr_name == "count":
+        assert len(output_terms) == 1, "Count and avg expect a single output variable"
+        assert isinstance(meta_output_terms[0], ir.Var)
+        # Count sums up "1" for each row. We use the expected output type for the type
+        # of the count variable.
+        typ = meta_type_to_lqp(meta_output_terms[0].type)
+        one_var, eq = constant_to_var(ctx, to_lqp_value(1, meta_output_terms[0].type), "one")
+        body_conjs.append(eq)
+        abstr_args.append((one_var, typ))
+
+    # Average needs to wrap the reduce in Exists(Conjunction(Reduce, div))
+    if aggr_name == "avg":
+        assert len(output_vars) == 1, "avg should only have one output variable"
+        output_var = output_vars[0]
+
+        # Count sums up "1" for each row. We make the reasonably safe assumption that there
+        # are less than 2^31 rows in the body.
+        count_type = meta_type_to_lqp(types.Int64)
+        one_var, eq = constant_to_var(ctx, to_lqp_value(1, types.Int64), "one")
+        body_conjs.append(eq)
+        abstr_args.append((one_var, count_type))
+
+        # The average will produce two output variables: sum and count.
+        sum_result = gen_unique_var(ctx, "sum")
+        count_result = gen_unique_var(ctx, "count")
+
+        # Second to last is the variable we're summing over.
+        (sum_var, sum_type) = abstr_args[-2]
+        body = mk_and(body_conjs)
+
+        result = lqp.Reduce(
+            op=lqp_avg_op(ctx, aggr.aggregation.name, sum_var.name, sum_type),
+            body=mk_abstraction(abstr_args, body),
+            terms=[sum_result, count_result],
+            meta=None,
+        )
+
+        if sum_type == count_type:
+            div = mk_primitive("rel_primitive_divide_monotype", [sum_result, count_result, output_var])
+            conjunction = mk_and([result, div])
+
+            # Finally, we need to wrap everything in an `exists` to project away the sum and
+            # count variables and only keep the result of the division.
+            result_terms = [(sum_result, sum_type), (count_result, count_type)]
+        else:
+            # If the sum type and count type don't match, we need to cast the count
+            count_casted = gen_unique_var(ctx, "count_casted")
+            count_cast = lqp.Cast(input=count_result, result=count_casted, meta=None)
+
+            div = mk_primitive("rel_primitive_divide_monotype", [sum_result, count_casted, output_var])
+            conjunction = mk_and([result, count_cast, div])
+
+            # Finally, we need to wrap everything in an `exists` to project away the sum and
+            # count variables and only keep the result of the division.
+            result_terms = [(sum_result, sum_type), (count_result, count_type), (count_casted, sum_type)]
+
+        return mk_exists(result_terms, conjunction)
+
+    # `input_args` hold the types of the input arguments, but they may have been modified
+    # if we're dealing with a count, so we use `abstr_args` to find the type.
+    (aggr_arg, aggr_arg_type) = abstr_args[-1]
+    body = mk_and(body_conjs)
+
+    # Group-bys do not need to be handled at all, since they are introduced outside already
+    reduce = lqp.Reduce(
+        op=lqp_operator(ctx, aggr.aggregation.name, aggr_arg.name, aggr_arg_type),
+        body=mk_abstraction(abstr_args, body),
+        terms=output_vars,
+        meta=None
+    )
+    return reduce
+
+def _translate_to_formula(ctx: TranslationCtx, task: ir.Task) -> lqp.Formula:
+    if isinstance(task, ir.Logical):
+        conjuncts = [_translate_to_formula(ctx, child) for child in task.body]
+        return mk_and(conjuncts)
+    elif isinstance(task, ir.Lookup):
+        return _translate_to_atom(ctx, task)
+    elif isinstance(task, ir.Not):
+        return lqp.Not(arg=_translate_to_formula(ctx, task.task), meta=None)
+    elif isinstance(task, ir.Exists):
+        lqp_vars, conjuncts = _translate_bindings(ctx, list(task.vars))
+        conjuncts.append(_translate_to_formula(ctx, task.task))
+        return mk_exists(lqp_vars, mk_and(conjuncts))
+    elif isinstance(task, ir.Construct):
+        assert len(task.values) >= 1, "Construct should have at least one value"
+        terms = [_translate_term(ctx, arg) for arg in task.values]
+        result_term = _translate_term(ctx, task.id_var)
+        terms.append(result_term)
+        assert result_term[1].type_name == lqp.TypeName.UINT128, \
+            f"Attempting to store a {task.id_var} in a type `{result_term[1].type_name}`"
+
+        return mk_primitive("rel_primitive_hash_tuple_uint128", [v for v, _ in terms])
+    elif isinstance(task, ir.Union):
+        # TODO: handle hoisted vars if needed
+        disjs = [_translate_to_formula(ctx, child) for child in task.tasks]
+        return mk_or(disjs)
+    elif isinstance(task, (ir.Aggregate, ir.Output, ir.Update)):
+        # Nothing to do here, handled in _translate_to_decls
+        return mk_and([])
+    elif isinstance(task, ir.Rank):
+        # Nothing to do here, handled in _translate_to_decls
+        return mk_and([])
+    else:
+        raise NotImplementedError(f"Unknown task type (formula): {type(task)}")
+
+# Only used for translating terms on relatoms, which can be specialized values.
+def _translate_relterm(ctx: TranslationCtx, term: ir.Value) -> Tuple[lqp.RelTerm, lqp.Type]:
+    if isinstance(term, ir.Literal) and term.type == types.Symbol:
+        if isinstance(term.value, str):
+            value = mk_value(term.value)
+            return mk_specialized_value(value), meta_type_to_lqp(types.String)
+        elif isinstance(term.value, int):
+            value = mk_value(term.value)
+            return mk_specialized_value(value), meta_type_to_lqp(types.Int64)
+        else:
+            raise NotImplementedError(f"Cannot specialize literal of type {type(term.value)}")
+    return _translate_term(ctx, term)
+
+def _translate_term(ctx: TranslationCtx, term: ir.Value) -> Tuple[lqp.Term, lqp.Type]:
+    if isinstance(term, ir.ScalarType):
+        # TODO: ScalarType is not like other terms, should be handled separately.
+        return to_lqp_value(term.name, types.String), meta_type_to_lqp(types.String)
+    elif isinstance(term, ir.Var):
+        t = meta_type_to_lqp(term.type)
+        return _translate_var(ctx, term), t
+    else:
+        assert isinstance(term, ir.Literal), f"Cannot translate value {term!r} of type {type(term)} to LQP Term; neither Var nor Literal."
+        v = to_lqp_value(term.value, term.type)
+        return v, meta_type_to_lqp(term.type)
+
+def _translate_to_atom(ctx: TranslationCtx, task: ir.Lookup) -> lqp.Formula:
+    if task.relation == builtins.cast:
+        assert len(task.args) == 3, f"expected three terms for {task.relation.name}, got {len(task.args)}"
+
+        terms = []
+        for arg in task.args[1:]:
+            term, _ = _translate_relterm(ctx, arg)
+            terms.append(term)
+        return lqp.Cast(input=terms[0], result=terms[1], meta=None)
+
+    if task.relation == builtins.join:
+        return _translate_join(ctx, task)
+    elif task.relation == builtins.infomap:
+        return _translate_infomap(ctx, task)
+    elif task.relation == builtins.louvain:
+        return _translate_louvain(ctx, task)
+    elif task.relation == builtins.label_propagation:
+        return _translate_label_propagation(ctx, task)
+
+    terms = []
+    term_types = []
+    for arg in task.args:
+        # Handle varargs, which come wrapped in a tuple.
+        if isinstance(arg, tuple):
+            for vararg in arg:
+                term, ty = _translate_relterm(ctx, vararg)
+                terms.append(term)
+                term_types.append(ty)
+        else:
+            term, ty = _translate_relterm(ctx, arg)
+            terms.append(term)
+            term_types.append(ty)
+
+    if builtins.is_pragma(task.relation):
+        lqp_name = pragma_to_lqp_name(task.relation.name)
+        return mk_pragma(lqp_name, terms)
+
+    if builtins.is_builtin(task.relation):
+        return build_primitive(task.relation.name, terms, term_types)
+
+    if helpers.is_external(task.relation):
+        return lqp.RelAtom(name=task.relation.name, terms=terms, meta=None)
+
+    projection, _ = _translate_bindings(ctx, list(task.args))
+    rid = get_relation_id(ctx, task.relation, projection)
+    return lqp.Atom(name=rid, terms=terms, meta=None)
+
+
+def get_relation_id(ctx: TranslationCtx, relation: ir.Relation, projection: list[Tuple[lqp.Var, lqp.Type]] = []) -> lqp.RelationId:
+    types = "_".join([str(t.type_name) for (_, t) in projection])
+    if types:
+        types = "_" + types
+    if relation.id in ctx.def_names.id_to_name:
+        unique_name = ctx.def_names.id_to_name[relation.id]
+    else:
+        name = helpers.relation_name_prefix(relation) + relation.name
+        name = helpers.sanitize(name)
+        unique_name = ctx.def_names.get_name_by_id(relation.id, name)
+
+    return utils.gen_rel_id(ctx, unique_name, types)
+
+def get_output_id(ctx: TranslationCtx, orig_name: str, metamodel_id: int) -> lqp.RelationId:
+    unique_name = ctx.output_names.get_name_by_id(metamodel_id, orig_name)
+    return utils.gen_rel_id(ctx, unique_name)
+
+def _translate_bindings(ctx: TranslationCtx, bindings: list[ir.Value]) -> Tuple[list[Tuple[lqp.Var, lqp.Type]], list[lqp.Formula]]:
+    lqp_vars = []
+    conjuncts = []
+    for binding in bindings:
+        lqp_var, typ, eq = binding_to_lqp_var(ctx, binding)
+        lqp_vars.append((lqp_var, typ))
+        if eq is not None:
+            conjuncts.append(eq)
+
+    return lqp_vars, conjuncts
+
+def binding_to_lqp_var(ctx: TranslationCtx, binding: ir.Value) -> Tuple[lqp.Var, lqp.Type, Union[None, lqp.Formula]]:
+    if isinstance(binding, ir.Var):
+        var, typ = _translate_term(ctx, binding)
+        assert isinstance(var, lqp.Var)
+        return var, typ, None
+    elif isinstance(binding, ir.Literal):
+        lqp_value = to_lqp_value(binding.value, binding.type)
+        var, formula = constant_to_var(ctx, lqp_value)
+        return var, meta_type_to_lqp(binding.type), formula
+    else:
+        raise Exception(f"Unsupported binding type: {type(binding)}")
+
+def to_lqp_value(value: ir.PyValue, value_type: ir.Type) -> lqp.Value:
+    typ = meta_type_to_lqp(value_type)
+
+    # Ensure int values match the requested integer type.
+    if typ.type_name == lqp.TypeName.INT and isinstance(value, int):
+        assert lqp_types.INT_MIN <= value <= lqp_types.INT_MAX, f"{value} out of range for a 64-bit INT value"
+        val = value
+    elif typ.type_name == lqp.TypeName.INT128 and isinstance(value, int):
+        assert lqp_types.INT128_MIN <= value <= lqp_types.INT128_MAX, f"{value} out of range for an INT128 value"
+        val = lqp.Int128Value(value=value, meta=None)
+    elif typ.type_name == lqp.TypeName.UINT128 and isinstance(value, int):
+        assert lqp_types.UINT128_MIN <= value <= lqp_types.UINT128_MAX, f"{value} out of range for an UINT128 value"
+        val = lqp.UInt128Value(value=value, meta=None)
+    elif typ.type_name == lqp.TypeName.FLOAT and isinstance(value, float):
+        val = value
+    elif typ.type_name == lqp.TypeName.STRING and isinstance(value, str):
+        val = value
+    elif typ.type_name == lqp.TypeName.DECIMAL and isinstance(value, (int, float, PyDecimal)):
+        precision = typ.parameters[0].value
+        scale = typ.parameters[1].value
+        assert isinstance(precision, int) and isinstance(scale, int)
+        val = lqp.DecimalValue(precision=precision, scale=scale, value=PyDecimal(value), meta=None)
+    elif typ.type_name == lqp.TypeName.DATE and isinstance(value, date):
+        val = lqp.DateValue(value=value, meta=None)
+    elif typ.type_name == lqp.TypeName.DATETIME and isinstance(value, datetime):
+        utc_value = value.astimezone(timezone.utc) if value.tzinfo is not None else value # Convert to UTC cf. Iceberg
+        val = lqp.DateTimeValue(value=utc_value, meta=None)
+    elif typ.type_name == lqp.TypeName.BOOLEAN and isinstance(value, bool):
+        val = lqp.BooleanValue(value=value, meta=None)
+    else:
+        raise Exception(f"Unsupported type for constant: {typ.type_name} with value {value} of type {type(value)}")
+
+    return mk_value(val)
+
+def constant_to_var(ctx: TranslationCtx, value: lqp.Value, name_hint: str = "cvar") -> Tuple[lqp.Var, lqp.Formula]:
+    var = gen_unique_var(ctx, name_hint)
+    eq = mk_primitive("rel_primitive_eq", [var, value])
+    return var, eq
+
+def _extract_pyrel_error_ids(ctx: TranslationCtx, model: ir.Model) -> list[Tuple[lqp.RelationId, str]]:
+    effects = collect_by_type(ir.Update, model)
+    pyrel_error_attrs = []
+    # We use a separate counter here to avoid modifying the current model.
+    i = 1
+    for effect in effects:
+        if "pyrel_error_attrs" not in effect.relation.name:
+            continue
+        projection, _ = _translate_bindings(ctx, _effect_bindings(effect))
+        rel_id = get_relation_id(ctx, effect.relation, projection)
+        name = f"{effect.relation.name}_{i}"
+        pyrel_error_attrs.append((rel_id, name))
+        i += 1
+
+    return pyrel_error_attrs
+
+# Translate a relation reference into an abstraction over its fields.
+def _translate_relation_ref(ctx: TranslationCtx, relation: ir.Relation) -> lqp.Abstraction:
+    projection = []
+    for field in relation.fields:
+        var = gen_unique_var(ctx, field.name)
+        typ = meta_type_to_lqp(field.type)
+        projection.append((var, typ))
+
+    rid = get_relation_id(ctx, relation, projection)
+    atom = lqp.Atom(name=rid, terms=[var for (var, _) in projection], meta=None)
+    return mk_abstraction(projection, atom)
+
+# Common translation logic for graph algorithms.
+# task.args[0] : normalized weight list (int64, int64, float)
+# task.args[1] : normalized node count (relation or constant int64)
+# task.args[2] : normalized edge count (relation or constant int64)
+# task.args[3:-3] : algorithm parameters (var or constant)
+# task.args[-3] : diagnostic info
+# task.args[-2] : node index
+# task.args[-1] : community ident
+def _translate_graph_common(name: str, ctx: TranslationCtx, task: ir.Lookup):
+    abstractions = []
+
+    assert isinstance(task.args[0], ir.Relation), \
+        f"Expected relation as first arg to {name}, got {task.args[0]}:{type(task.args[0])}"
+    abstractions.append(_translate_relation_ref(ctx, task.args[0]))
+
+    # Allow constant args for node and edge count
+    for arg in task.args[1:3]:
+        if isinstance(arg, ir.Relation):
+            abst = _translate_relation_ref(ctx, arg)
+            typ = abst.vars[0][1]
+            assert typ.type_name == lqp.TypeName.INT, \
+                f"Expected Int64 types for node and edge counts, got type {typ.type_name}"
+            abstractions.append(abst)
+        else:
+            var, typ, eq = binding_to_lqp_var(ctx, arg)
+            assert eq is not None, \
+                f"Expected equality formula for {name} arg {arg}:{type(arg)}"
+            abstractions.append(mk_abstraction([(var, typ)], mk_and([eq])))
+
+    for arg in task.args[3:-3]:
+        var, typ, eq = binding_to_lqp_var(ctx, arg)
+        if eq:
+            abstractions.append(mk_abstraction([(var, typ)], mk_and([eq])))
+        else:
+            print(f"Primitive graph algorithm arg without eq:\n var:{var}, typ:{typ}\n arg:{arg}:{type(arg)}")
+            abstractions.append(mk_abstraction([(var, typ)], mk_and([])))
+
+    terms = []
+    for arg in task.args[-3:]:
+        term, _ = _translate_relterm(ctx, arg)
+        terms.append(term)
+
+    return (abstractions, terms)
+
+def _translate_infomap(ctx: TranslationCtx, task: ir.Lookup) -> lqp.Formula:
+    abstractions, terms = _translate_graph_common("infomap", ctx, task)
+
+    return lqp.FFI(
+        meta=None,
+        name="rel_primitive_infomap",
+        args=abstractions,
+        terms=terms,
+    )
+
+def _translate_louvain(ctx: TranslationCtx, task: ir.Lookup) -> lqp.Formula:
+    abstractions, terms = _translate_graph_common("louvain", ctx, task)
+
+    return lqp.FFI(
+        meta=None,
+        name="rel_primitive_louvain",
+        args=abstractions,
+        terms=terms,
+    )
+
+def _translate_label_propagation(ctx: TranslationCtx, task: ir.Lookup) -> lqp.Formula:
+    abstractions, terms = _translate_graph_common("label_propagation", ctx, task)
+
+    return lqp.FFI(
+        meta=None,
+        name="rel_primitive_async_label_propagation",
+        args=abstractions,
+        terms=terms,
+    )
+
+# Hard-coded implementation of Rel's string_join
+def _translate_join(ctx: TranslationCtx, task: ir.Lookup) -> lqp.Formula:
+    assert len(task.args) == 3
+    (strs, separator, target) = task.args
+    assert isinstance(separator, ir.Node)
+    assert isinstance(target, ir.Var)
+    assert isinstance(strs, tuple)
+
+    # Enumerate the strings we're going to be reducing over
+    enumerated_conjunctions = []
+    index_var, index_type = gen_unique_var(ctx, "idx"), meta_type_to_lqp(types.Int64)
+    string_var, string_type = gen_unique_var(ctx, "s"), meta_type_to_lqp(types.String)
+    for i, s in enumerate(strs):
+        index = to_lqp_value(i, types.Int64)
+        string = _translate_term(ctx, s)[0]
+        eq_idx = mk_primitive("rel_primitive_eq", [index_var, index])
+        eq_string = mk_primitive("rel_primitive_eq", [string_var, string])
+        enumerated_conjunctions.append(mk_and([eq_idx, eq_string]))
+    body = mk_abstraction([(index_var, index_type), (string_var, string_type)], mk_or(enumerated_conjunctions))
+
+    # Make the function we're reducing
+    a_var, a_type = gen_unique_var(ctx, "a"), meta_type_to_lqp(types.String)
+    b_var, b_type = gen_unique_var(ctx, "b"), meta_type_to_lqp(types.String)
+    curr_var, curr_type = gen_unique_var(ctx, "curr"), meta_type_to_lqp(types.String)
+    res_var, res_type = gen_unique_var(ctx, "res"), meta_type_to_lqp(types.String)
+
+    sep, _ = _translate_term(ctx, separator)
+
+    op = mk_abstraction(
+        [(a_var, a_type), (b_var, b_type), (curr_var, curr_type)],
+        mk_exists([(res_var, res_type)],
+            mk_and([
+                mk_primitive("rel_primitive_concat", [res_var, b_var, curr_var]),
+                mk_primitive("rel_primitive_concat", [a_var, sep, res_var])
+            ])
+        )
+    )
+
+    output_term = _translate_term(ctx, target)[0]
+
+    return lqp.Reduce(meta=None, op=op, body=body, terms=[output_term])
+
+def _translate_var(ctx: TranslationCtx, term: ir.Var) -> lqp.Var:
+    name = ctx.var_names.get_name_by_id(term.id, term.name)
+    return lqp.Var(name=name, meta=None)