search_utils.py

import ast
import base64
import json
import math
import operator
import re
import shlex

import sqlparse
from packaging.version import Version
from sqlparse.sql import (
    Comparison,
    Identifier,
    IdentifierList,
    Parenthesis,
    Statement,
    Token,
    TokenList,
)
from sqlparse.tokens import Token as TokenType

from mlflow.entities import RunInfo
from mlflow.entities.model_registry.model_version_stages import STAGE_DELETED_INTERNAL
from mlflow.exceptions import MlflowException
from mlflow.protos.databricks_pb2 import INVALID_PARAMETER_VALUE
from mlflow.store.db.db_types import MSSQL, MYSQL, POSTGRES, SQLITE
from mlflow.utils.mlflow_tags import (
    MLFLOW_DATASET_CONTEXT,
)


def _convert_like_pattern_to_regex(pattern, flags=0):
    if not pattern.startswith("%"):
        pattern = "^" + pattern
    if not pattern.endswith("%"):
        pattern = pattern + "$"
    return re.compile(pattern.replace("_", ".").replace("%", ".*"), flags)


def _like(string, pattern):
    return _convert_like_pattern_to_regex(pattern).match(string) is not None


def _ilike(string, pattern):
    return _convert_like_pattern_to_regex(pattern, flags=re.IGNORECASE).match(string) is not None


def _join_in_comparison_tokens(tokens):
    """
    Find a sequence of tokens that matches the pattern of an IN comparison or a NOT IN comparison,
    join the tokens into a single Comparison token. Otherwise, return the original list of tokens.
    """
    if Version(sqlparse.__version__) < Version("0.4.4"):
        # In sqlparse < 0.4.4, IN is treated as a comparison, we don't need to join tokens
        return tokens

    non_whitespace_tokens = [t for t in tokens if not t.is_whitespace]
    joined_tokens = []
    num_tokens = len(non_whitespace_tokens)
    iterator = enumerate(non_whitespace_tokens)
    while elem := next(iterator, None):
        index, first = elem
        # We need at least 3 tokens to form an IN comparison or a NOT IN comparison
        if num_tokens - index < 3:
            joined_tokens.extend(non_whitespace_tokens[index:])
            break

        # Wait until we encounter an identifier token
        if not isinstance(first, Identifier):
            joined_tokens.append(first)
            continue

        (_, second) = next(iterator)
        (_, third) = next(iterator)

        # IN
        if (
            isinstance(first, Identifier)
            and second.match(ttype=TokenType.Keyword, values=["IN"])
            and isinstance(third, Parenthesis)
        ):
            joined_tokens.append(Comparison(TokenList([first, second, third])))
            continue

        (_, fourth) = next(iterator, (None, None))
        if fourth is None:
            joined_tokens.extend([first, second, third])
            break

        # NOT IN
        if (
            isinstance(first, Identifier)
            and second.match(ttype=TokenType.Keyword, values=["NOT"])
            and third.match(ttype=TokenType.Keyword, values=["IN"])
            and isinstance(fourth, Parenthesis)
        ):
            joined_tokens.append(
                Comparison(TokenList([first, Token(TokenType.Keyword, "NOT IN"), fourth]))
            )
            continue

        joined_tokens.extend([first, second, third, fourth])

    return joined_tokens


class SearchUtils:
    LIKE_OPERATOR = "LIKE"
    ILIKE_OPERATOR = "ILIKE"
    ASC_OPERATOR = "asc"
    DESC_OPERATOR = "desc"
    VALID_ORDER_BY_TAGS = [ASC_OPERATOR, DESC_OPERATOR]
    VALID_METRIC_COMPARATORS = {">", ">=", "!=", "=", "<", "<="}
    VALID_PARAM_COMPARATORS = {"!=", "=", LIKE_OPERATOR, ILIKE_OPERATOR}
    VALID_TAG_COMPARATORS = {"!=", "=", LIKE_OPERATOR, ILIKE_OPERATOR}
    VALID_STRING_ATTRIBUTE_COMPARATORS = {"!=", "=", LIKE_OPERATOR, ILIKE_OPERATOR, "IN", "NOT IN"}
    VALID_NUMERIC_ATTRIBUTE_COMPARATORS = VALID_METRIC_COMPARATORS
    VALID_DATASET_COMPARATORS = {"!=", "=", LIKE_OPERATOR, ILIKE_OPERATOR, "IN", "NOT IN"}
    _BUILTIN_NUMERIC_ATTRIBUTES = {"start_time", "end_time"}
    _ALTERNATE_NUMERIC_ATTRIBUTES = {"created", "Created"}
    _ALTERNATE_STRING_ATTRIBUTES = {"run name", "Run name", "Run Name"}
    NUMERIC_ATTRIBUTES = set(
        list(_BUILTIN_NUMERIC_ATTRIBUTES) + list(_ALTERNATE_NUMERIC_ATTRIBUTES)
    )
    DATASET_ATTRIBUTES = {"name", "digest", "context"}
    VALID_SEARCH_ATTRIBUTE_KEYS = set(
        RunInfo.get_searchable_attributes()
        + list(_ALTERNATE_NUMERIC_ATTRIBUTES)
        + list(_ALTERNATE_STRING_ATTRIBUTES)
    )
    VALID_ORDER_BY_ATTRIBUTE_KEYS = set(
        RunInfo.get_orderable_attributes() + list(_ALTERNATE_NUMERIC_ATTRIBUTES)
    )
    _METRIC_IDENTIFIER = "metric"
    _ALTERNATE_METRIC_IDENTIFIERS = {"metrics"}
    _PARAM_IDENTIFIER = "parameter"
    _ALTERNATE_PARAM_IDENTIFIERS = {"parameters", "param", "params"}
    _TAG_IDENTIFIER = "tag"
    _ALTERNATE_TAG_IDENTIFIERS = {"tags"}
    _ATTRIBUTE_IDENTIFIER = "attribute"
    _ALTERNATE_ATTRIBUTE_IDENTIFIERS = {"attr", "attributes", "run"}
    _DATASET_IDENTIFIER = "dataset"
    _ALTERNATE_DATASET_IDENTIFIERS = {"datasets"}
    _IDENTIFIERS = [
        _METRIC_IDENTIFIER,
        _PARAM_IDENTIFIER,
        _TAG_IDENTIFIER,
        _ATTRIBUTE_IDENTIFIER,
        _DATASET_IDENTIFIER,
    ]
    _VALID_IDENTIFIERS = set(
        _IDENTIFIERS
        + list(_ALTERNATE_METRIC_IDENTIFIERS)
        + list(_ALTERNATE_PARAM_IDENTIFIERS)
        + list(_ALTERNATE_TAG_IDENTIFIERS)
        + list(_ALTERNATE_ATTRIBUTE_IDENTIFIERS)
        + list(_ALTERNATE_DATASET_IDENTIFIERS)
    )
    STRING_VALUE_TYPES = {TokenType.Literal.String.Single}
    DELIMITER_VALUE_TYPES = {TokenType.Punctuation}
    WHITESPACE_VALUE_TYPE = TokenType.Text.Whitespace
    NUMERIC_VALUE_TYPES = {TokenType.Literal.Number.Integer, TokenType.Literal.Number.Float}
    # Registered Models Constants
    ORDER_BY_KEY_TIMESTAMP = "timestamp"
    ORDER_BY_KEY_LAST_UPDATED_TIMESTAMP = "last_updated_timestamp"
    ORDER_BY_KEY_MODEL_NAME = "name"
    VALID_ORDER_BY_KEYS_REGISTERED_MODELS = {
        ORDER_BY_KEY_TIMESTAMP,
        ORDER_BY_KEY_LAST_UPDATED_TIMESTAMP,
        ORDER_BY_KEY_MODEL_NAME,
    }
    VALID_TIMESTAMP_ORDER_BY_KEYS = {ORDER_BY_KEY_TIMESTAMP, ORDER_BY_KEY_LAST_UPDATED_TIMESTAMP}
    # We encourage users to use timestamp for order-by
    RECOMMENDED_ORDER_BY_KEYS_REGISTERED_MODELS = {ORDER_BY_KEY_MODEL_NAME, ORDER_BY_KEY_TIMESTAMP}

    @staticmethod
    def get_comparison_func(comparator):
        return {
            ">": operator.gt,
            ">=": operator.ge,
            "=": operator.eq,
            "!=": operator.ne,
            "<=": operator.le,
            "<": operator.lt,
            "LIKE": _like,
            "ILIKE": _ilike,
            "IN": lambda x, y: x in y,
            "NOT IN": lambda x, y: x not in y,
        }[comparator]

    @staticmethod
    def get_sql_comparison_func(comparator, dialect):
        import sqlalchemy as sa

        def comparison_func(column, value):
            if comparator == "LIKE":
                return column.like(value)
            elif comparator == "ILIKE":
                return column.ilike(value)
            elif comparator == "IN":
                return column.in_(value)
            elif comparator == "NOT IN":
                return ~column.in_(value)
            return SearchUtils.get_comparison_func(comparator)(column, value)

        def mssql_comparison_func(column, value):
            if not isinstance(column.type, sa.types.String):
                return comparison_func(column, value)

            collated = column.collate("Japanese_Bushu_Kakusu_100_CS_AS_KS_WS")
            return comparison_func(collated, value)

        def mysql_comparison_func(column, value):
            if not isinstance(column.type, sa.types.String):
                return comparison_func(column, value)

            # MySQL is case insensitive by default, so we need to use the binary operator to
            # perform case sensitive comparisons.
            templates = {
                # Use non-binary ahead of binary comparison for runtime performance
                "=": "({column} = :value AND BINARY {column} = :value)",
                "!=": "({column} != :value OR BINARY {column} != :value)",
                "LIKE": "({column} LIKE :value AND BINARY {column} LIKE :value)",
            }
            if comparator in templates:
                column = f"{column.class_.__tablename__}.{column.key}"
                return sa.text(templates[comparator].format(column=column)).bindparams(
                    sa.bindparam("value", value=value, unique=True)
                )

            return comparison_func(column, value)

        return {
            POSTGRES: comparison_func,
            SQLITE: comparison_func,
            MSSQL: mssql_comparison_func,
            MYSQL: mysql_comparison_func,
        }[dialect]

    @staticmethod
    def translate_key_alias(key):
        if key in ["created", "Created"]:
            return "start_time"
        if key in ["run name", "Run name", "Run Name"]:
            return "run_name"
        return key

    @classmethod
    def _trim_ends(cls, string_value):
        return string_value[1:-1]

    @classmethod
    def _is_quoted(cls, value, pattern):
        return len(value) >= 2 and value.startswith(pattern) and value.endswith(pattern)

    @classmethod
    def _trim_backticks(cls, entity_type):
        """Remove backticks from identifier like `param`, if they exist."""
        if cls._is_quoted(entity_type, "`"):
            return cls._trim_ends(entity_type)
        return entity_type

    @classmethod
    def _strip_quotes(cls, value, expect_quoted_value=False):
        """
        Remove quotes for input string.
        Values of type strings are expected to have quotes.
        Keys containing special characters are also expected to be enclose in quotes.
        """
        if cls._is_quoted(value, "'") or cls._is_quoted(value, '"'):
            return cls._trim_ends(value)
        elif expect_quoted_value:
            raise MlflowException(
                "Parameter value is either not quoted or unidentified quote "
                "types used for string value %s. Use either single or double "
                "quotes." % value,
                error_code=INVALID_PARAMETER_VALUE,
            )
        else:
            return value

    @classmethod
    def _valid_entity_type(cls, entity_type):
        entity_type = cls._trim_backticks(entity_type)
        if entity_type not in cls._VALID_IDENTIFIERS:
            raise MlflowException(
                f"Invalid entity type '{entity_type}'. Valid values are {cls._IDENTIFIERS}",
                error_code=INVALID_PARAMETER_VALUE,
            )

        if entity_type in cls._ALTERNATE_PARAM_IDENTIFIERS:
            return cls._PARAM_IDENTIFIER
        elif entity_type in cls._ALTERNATE_METRIC_IDENTIFIERS:
            return cls._METRIC_IDENTIFIER
        elif entity_type in cls._ALTERNATE_TAG_IDENTIFIERS:
            return cls._TAG_IDENTIFIER
        elif entity_type in cls._ALTERNATE_ATTRIBUTE_IDENTIFIERS:
            return cls._ATTRIBUTE_IDENTIFIER
        elif entity_type in cls._ALTERNATE_DATASET_IDENTIFIERS:
            return cls._DATASET_IDENTIFIER
        else:
            # one of ("metric", "parameter", "tag", or "attribute") since it a valid type
            return entity_type

    @classmethod
    def _get_identifier(cls, identifier, valid_attributes):
        try:
            tokens = identifier.split(".", 1)
            if len(tokens) == 1:
                key = tokens[0]
                entity_type = cls._ATTRIBUTE_IDENTIFIER
            else:
                entity_type, key = tokens
        except ValueError:
            raise MlflowException(
                "Invalid identifier '%s'. Columns should be specified as "
                "'attribute.<key>', 'metric.<key>', 'tag.<key>', 'dataset.<key>', or "
                "'param.'." % identifier,
                error_code=INVALID_PARAMETER_VALUE,
            )
        identifier = cls._valid_entity_type(entity_type)
        key = cls._trim_backticks(cls._strip_quotes(key))
        if identifier == cls._ATTRIBUTE_IDENTIFIER and key not in valid_attributes:
            raise MlflowException.invalid_parameter_value(
                f"Invalid attribute key '{key}' specified. Valid keys are '{valid_attributes}'"
            )
        elif identifier == cls._DATASET_IDENTIFIER and key not in cls.DATASET_ATTRIBUTES:
            raise MlflowException.invalid_parameter_value(
                f"Invalid dataset key '{key}' specified. Valid keys are '{cls.DATASET_ATTRIBUTES}'"
            )
        return {"type": identifier, "key": key}

    @classmethod
    def _get_value(cls, identifier_type, key, token):
        if identifier_type == cls._METRIC_IDENTIFIER:
            if token.ttype not in cls.NUMERIC_VALUE_TYPES:
                raise MlflowException(
                    f"Expected numeric value type for metric. Found {token.value}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
            return token.value
        elif identifier_type == cls._PARAM_IDENTIFIER or identifier_type == cls._TAG_IDENTIFIER:
            if token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            raise MlflowException(
                "Expected a quoted string value for "
                f"{identifier_type} (e.g. 'my-value'). Got value "
                f"{token.value}",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif identifier_type == cls._ATTRIBUTE_IDENTIFIER:
            if key in cls.NUMERIC_ATTRIBUTES:
                if token.ttype not in cls.NUMERIC_VALUE_TYPES:
                    raise MlflowException(
                        f"Expected numeric value type for numeric attribute: {key}. "
                        f"Found {token.value}",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                return token.value
            elif token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            elif isinstance(token, Parenthesis):
                if key != "run_id":
                    raise MlflowException(
                        "Only the 'run_id' attribute supports comparison with a list of quoted "
                        "string values.",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                return cls._parse_run_ids(token)
            else:
                raise MlflowException(
                    f"Expected a quoted string value for attributes. Got value {token.value}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
        elif identifier_type == cls._DATASET_IDENTIFIER:
            if key in cls.DATASET_ATTRIBUTES and (
                token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier)
            ):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            elif isinstance(token, Parenthesis):
                if key not in ("name", "digest", "context"):
                    raise MlflowException(
                        "Only the dataset 'name' and 'digest' supports comparison with a list of "
                        "quoted string values.",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                return cls._parse_run_ids(token)
            else:
                raise MlflowException(
                    "Expected a quoted string value for dataset attributes. "
                    f"Got value {token.value}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
        else:
            # Expected to be either "param" or "metric".
            raise MlflowException(
                "Invalid identifier type. Expected one of "
                f"{[cls._METRIC_IDENTIFIER, cls._PARAM_IDENTIFIER]}."
            )

    @classmethod
    def _validate_comparison(cls, tokens):
        base_error_string = "Invalid comparison clause"
        if len(tokens) != 3:
            raise MlflowException(
                f"{base_error_string}. Expected 3 tokens found {len(tokens)}",
                error_code=INVALID_PARAMETER_VALUE,
            )
        if not isinstance(tokens[0], Identifier):
            raise MlflowException(
                f"{base_error_string}. Expected 'Identifier' found '{tokens[0]}'",
                error_code=INVALID_PARAMETER_VALUE,
            )
        if not isinstance(tokens[1], Token) and tokens[1].ttype != TokenType.Operator.Comparison:
            raise MlflowException(
                f"{base_error_string}. Expected comparison found '{tokens[1]}'",
                error_code=INVALID_PARAMETER_VALUE,
            )
        if not isinstance(tokens[2], Token) and (
            tokens[2].ttype not in cls.STRING_VALUE_TYPES.union(cls.NUMERIC_VALUE_TYPES)
            or isinstance(tokens[2], Identifier)
        ):
            raise MlflowException(
                f"{base_error_string}. Expected value token found '{tokens[2]}'",
                error_code=INVALID_PARAMETER_VALUE,
            )

    @classmethod
    def _get_comparison(cls, comparison):
        stripped_comparison = [token for token in comparison.tokens if not token.is_whitespace]
        cls._validate_comparison(stripped_comparison)
        comp = cls._get_identifier(stripped_comparison[0].value, cls.VALID_SEARCH_ATTRIBUTE_KEYS)
        comp["comparator"] = stripped_comparison[1].value
        comp["value"] = cls._get_value(comp.get("type"), comp.get("key"), stripped_comparison[2])
        return comp

    @classmethod
    def _invalid_statement_token_search_runs(cls, token):
        if (
            isinstance(token, Comparison)
            or token.is_whitespace
            or token.match(ttype=TokenType.Keyword, values=["AND"])
        ):
            return False
        return True

    @classmethod
    def _process_statement(cls, statement):
        # check validity
        tokens = _join_in_comparison_tokens(statement.tokens)
        invalids = list(filter(cls._invalid_statement_token_search_runs, tokens))
        if len(invalids) > 0:
            invalid_clauses = ", ".join(f"'{token}'" for token in invalids)
            raise MlflowException(
                f"Invalid clause(s) in filter string: {invalid_clauses}",
                error_code=INVALID_PARAMETER_VALUE,
            )
        return [cls._get_comparison(si) for si in tokens if isinstance(si, Comparison)]

    @classmethod
    def parse_search_filter(cls, filter_string):
        if not filter_string:
            return []
        try:
            parsed = sqlparse.parse(filter_string)
        except Exception:
            raise MlflowException(
                f"Error on parsing filter '{filter_string}'", error_code=INVALID_PARAMETER_VALUE
            )
        if len(parsed) == 0 or not isinstance(parsed[0], Statement):
            raise MlflowException(
                f"Invalid filter '{filter_string}'. Could not be parsed.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif len(parsed) > 1:
            raise MlflowException(
                "Search filter contained multiple expression '%s'. "
                "Provide AND-ed expression list." % filter_string,
                error_code=INVALID_PARAMETER_VALUE,
            )
        return cls._process_statement(parsed[0])

    @classmethod
    def is_metric(cls, key_type, comparator):
        if key_type == cls._METRIC_IDENTIFIER:
            if comparator not in cls.VALID_METRIC_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' "
                    f"not one of '{cls.VALID_METRIC_COMPARATORS}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
            return True
        return False

    @classmethod
    def is_param(cls, key_type, comparator):
        if key_type == cls._PARAM_IDENTIFIER:
            if comparator not in cls.VALID_PARAM_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' "
                    f"not one of '{cls.VALID_PARAM_COMPARATORS}'",
                    error_code=INVALID_PARAMETER_VALUE,
                )
            return True
        return False

    @classmethod
    def is_tag(cls, key_type, comparator):
        if key_type == cls._TAG_IDENTIFIER:
            if comparator not in cls.VALID_TAG_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' not one of '{cls.VALID_TAG_COMPARATORS}"
                )
            return True
        return False

    @classmethod
    def is_string_attribute(cls, key_type, key_name, comparator):
        if key_type == cls._ATTRIBUTE_IDENTIFIER and key_name not in cls.NUMERIC_ATTRIBUTES:
            if comparator not in cls.VALID_STRING_ATTRIBUTE_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' not one of "
                    f"'{cls.VALID_STRING_ATTRIBUTE_COMPARATORS}'"
                )
            return True
        return False

    @classmethod
    def is_numeric_attribute(cls, key_type, key_name, comparator):
        if key_type == cls._ATTRIBUTE_IDENTIFIER and key_name in cls.NUMERIC_ATTRIBUTES:
            if comparator not in cls.VALID_NUMERIC_ATTRIBUTE_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' not one of "
                    f"'{cls.VALID_STRING_ATTRIBUTE_COMPARATORS}"
                )
            return True
        return False

    @classmethod
    def is_dataset(cls, key_type, comparator):
        if key_type == cls._DATASET_IDENTIFIER:
            if comparator not in cls.VALID_DATASET_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' "
                    f"not one of '{cls.VALID_DATASET_COMPARATORS}"
                )
            return True
        return False

    @classmethod
    def _does_run_match_clause(cls, run, sed):
        key_type = sed.get("type")
        key = sed.get("key")
        value = sed.get("value")
        comparator = sed.get("comparator").upper()

        key = SearchUtils.translate_key_alias(key)

        if cls.is_metric(key_type, comparator):
            lhs = run.data.metrics.get(key, None)
            value = float(value)
        elif cls.is_param(key_type, comparator):
            lhs = run.data.params.get(key, None)
        elif cls.is_tag(key_type, comparator):
            lhs = run.data.tags.get(key, None)
        elif cls.is_string_attribute(key_type, key, comparator):
            lhs = getattr(run.info, key)
        elif cls.is_numeric_attribute(key_type, key, comparator):
            lhs = getattr(run.info, key)
            value = int(value)
        elif cls.is_dataset(key_type, comparator):
            if key == "context":
                return any(
                    SearchUtils.get_comparison_func(comparator)(tag.value if tag else None, value)
                    for dataset_input in run.inputs.dataset_inputs
                    for tag in dataset_input.tags
                    if tag.key == MLFLOW_DATASET_CONTEXT
                )
            else:
                return any(
                    SearchUtils.get_comparison_func(comparator)(
                        getattr(dataset_input.dataset, key), value
                    )
                    for dataset_input in run.inputs.dataset_inputs
                )
        else:
            raise MlflowException(
                f"Invalid search expression type '{key_type}'", error_code=INVALID_PARAMETER_VALUE
            )
        if lhs is None:
            return False

        return SearchUtils.get_comparison_func(comparator)(lhs, value)

    @classmethod
    def filter(cls, runs, filter_string):
        """Filters a set of runs based on a search filter string."""
        if not filter_string:
            return runs
        parsed = cls.parse_search_filter(filter_string)

        def run_matches(run):
            return all(cls._does_run_match_clause(run, s) for s in parsed)

        return [run for run in runs if run_matches(run)]

    @classmethod
    def _validate_order_by_and_generate_token(cls, order_by):
        try:
            parsed = sqlparse.parse(order_by)
        except Exception:
            raise MlflowException(
                f"Error on parsing order_by clause '{order_by}'",
                error_code=INVALID_PARAMETER_VALUE,
            )
        if len(parsed) != 1 or not isinstance(parsed[0], Statement):
            raise MlflowException(
                f"Invalid order_by clause '{order_by}'. Could not be parsed.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        statement = parsed[0]
        ttype_for_timestamp = (
            TokenType.Name.Builtin
            if Version(sqlparse.__version__) >= Version("0.4.3")
            else TokenType.Keyword
        )

        if len(statement.tokens) == 1 and isinstance(statement[0], Identifier):
            token_value = statement.tokens[0].value
        elif len(statement.tokens) == 1 and statement.tokens[0].match(
            ttype=ttype_for_timestamp, values=[cls.ORDER_BY_KEY_TIMESTAMP]
        ):
            token_value = cls.ORDER_BY_KEY_TIMESTAMP
        elif (
            statement.tokens[0].match(
                ttype=ttype_for_timestamp, values=[cls.ORDER_BY_KEY_TIMESTAMP]
            )
            and all(token.is_whitespace for token in statement.tokens[1:-1])
            and statement.tokens[-1].ttype == TokenType.Keyword.Order
        ):
            token_value = cls.ORDER_BY_KEY_TIMESTAMP + " " + statement.tokens[-1].value
        else:
            raise MlflowException(
                f"Invalid order_by clause '{order_by}'. Could not be parsed.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        return token_value

    @classmethod
    def _parse_order_by_string(cls, order_by):
        token_value = cls._validate_order_by_and_generate_token(order_by)
        is_ascending = True
        tokens = shlex.split(token_value.replace("`", '"'))
        if len(tokens) > 2:
            raise MlflowException(
                f"Invalid order_by clause '{order_by}'. Could not be parsed.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif len(tokens) == 2:
            order_token = tokens[1].lower()
            if order_token not in cls.VALID_ORDER_BY_TAGS:
                raise MlflowException(
                    f"Invalid ordering key in order_by clause '{order_by}'.",
                    error_code=INVALID_PARAMETER_VALUE,
                )
            is_ascending = order_token == cls.ASC_OPERATOR
            token_value = tokens[0]
        return token_value, is_ascending

    @classmethod
    def parse_order_by_for_search_runs(cls, order_by):
        token_value, is_ascending = cls._parse_order_by_string(order_by)
        identifier = cls._get_identifier(token_value.strip(), cls.VALID_ORDER_BY_ATTRIBUTE_KEYS)
        return identifier["type"], identifier["key"], is_ascending

    @classmethod
    def parse_order_by_for_search_registered_models(cls, order_by):
        token_value, is_ascending = cls._parse_order_by_string(order_by)
        token_value = token_value.strip()
        if token_value not in cls.VALID_ORDER_BY_KEYS_REGISTERED_MODELS:
            raise MlflowException(
                f"Invalid order by key '{token_value}' specified. Valid keys "
                f"are '{cls.RECOMMENDED_ORDER_BY_KEYS_REGISTERED_MODELS}'",
                error_code=INVALID_PARAMETER_VALUE,
            )
        return token_value, is_ascending

    @classmethod
    def _get_value_for_sort(cls, run, key_type, key, ascending):
        """Returns a tuple suitable to be used as a sort key for runs."""
        sort_value = None
        key = SearchUtils.translate_key_alias(key)
        if key_type == cls._METRIC_IDENTIFIER:
            sort_value = run.data.metrics.get(key)
        elif key_type == cls._PARAM_IDENTIFIER:
            sort_value = run.data.params.get(key)
        elif key_type == cls._TAG_IDENTIFIER:
            sort_value = run.data.tags.get(key)
        elif key_type == cls._ATTRIBUTE_IDENTIFIER:
            sort_value = getattr(run.info, key)
        else:
            raise MlflowException(
                f"Invalid order_by entity type '{key_type}'", error_code=INVALID_PARAMETER_VALUE
            )

        # Return a key such that None values are always at the end.
        is_none = sort_value is None
        is_nan = isinstance(sort_value, float) and math.isnan(sort_value)
        fill_value = (1 if ascending else -1) * math.inf

        if is_none:
            sort_value = fill_value
        elif is_nan:
            sort_value = -fill_value

        is_none_or_nan = is_none or is_nan

        return (is_none_or_nan, sort_value) if ascending else (not is_none_or_nan, sort_value)

    @classmethod
    def sort(cls, runs, order_by_list):
        """Sorts a set of runs based on their natural ordering and an overriding set of order_bys.
        Runs are naturally ordered first by start time descending, then by run id for tie-breaking.
        """
        runs = sorted(runs, key=lambda run: (-run.info.start_time, run.info.run_uuid))
        if not order_by_list:
            return runs
        # NB: We rely on the stability of Python's sort function, so that we can apply
        # the ordering conditions in reverse order.
        for order_by_clause in reversed(order_by_list):
            (key_type, key, ascending) = cls.parse_order_by_for_search_runs(order_by_clause)

            runs = sorted(
                runs,
                key=lambda run: cls._get_value_for_sort(run, key_type, key, ascending),
                reverse=not ascending,
            )
        return runs

    @classmethod
    def parse_start_offset_from_page_token(cls, page_token):
        # Note: the page_token is expected to be a base64-encoded JSON that looks like
        # { "offset": xxx }. However, this format is not stable, so it should not be
        # relied upon outside of this method.
        if not page_token:
            return 0

        try:
            decoded_token = base64.b64decode(page_token)
        except TypeError:
            raise MlflowException(
                "Invalid page token, could not base64-decode", error_code=INVALID_PARAMETER_VALUE
            )
        except base64.binascii.Error:
            raise MlflowException(
                "Invalid page token, could not base64-decode", error_code=INVALID_PARAMETER_VALUE
            )

        try:
            parsed_token = json.loads(decoded_token)
        except ValueError:
            raise MlflowException(
                f"Invalid page token, decoded value={decoded_token}",
                error_code=INVALID_PARAMETER_VALUE,
            )

        offset_str = parsed_token.get("offset")
        if not offset_str:
            raise MlflowException(
                f"Invalid page token, parsed value={parsed_token}",
                error_code=INVALID_PARAMETER_VALUE,
            )

        try:
            offset = int(offset_str)
        except ValueError:
            raise MlflowException(
                f"Invalid page token, not stringable {offset_str}",
                error_code=INVALID_PARAMETER_VALUE,
            )

        return offset

    @classmethod
    def create_page_token(cls, offset):
        return base64.b64encode(json.dumps({"offset": offset}).encode("utf-8"))

    @classmethod
    def paginate(cls, runs, page_token, max_results):
        """Paginates a set of runs based on an offset encoded into the page_token and a max
        results limit. Returns a pair containing the set of paginated runs, followed by
        an optional next_page_token if there are further results that need to be returned.
        """
        start_offset = cls.parse_start_offset_from_page_token(page_token)
        final_offset = start_offset + max_results

        paginated_runs = runs[start_offset:final_offset]
        next_page_token = None
        if final_offset < len(runs):
            next_page_token = cls.create_page_token(final_offset)
        return (paginated_runs, next_page_token)

    # Model Registry specific parser
    # TODO: Tech debt. Refactor search code into common utils, tracking server, and model
    #       registry specific code.

    VALID_SEARCH_KEYS_FOR_MODEL_VERSIONS = {"name", "run_id", "source_path"}
    VALID_SEARCH_KEYS_FOR_REGISTERED_MODELS = {"name"}

    @classmethod
    def _check_valid_identifier_list(cls, value_token):
        if len(value_token._groupable_tokens) == 0:
            raise MlflowException(
                "While parsing a list in the query,"
                " expected a non-empty list of string values, but got empty list",
                error_code=INVALID_PARAMETER_VALUE,
            )

        # Single element (e.g. `('x')`)
        if (
            len(value_token._groupable_tokens) == 1
            and value_token._groupable_tokens[0].ttype is TokenType.String.Single
        ):
            return

        # Multiple elements (e.g. `('x','y')`)
        if not isinstance(value_token._groupable_tokens[0], IdentifierList):
            raise MlflowException(
                "While parsing a list in the query,"
                " expected a non-empty list of string values, but got ill-formed list.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif not all(
            token.ttype
            in {*cls.STRING_VALUE_TYPES, *cls.DELIMITER_VALUE_TYPES, cls.WHITESPACE_VALUE_TYPE}
            for token in value_token._groupable_tokens[0].tokens
        ):
            raise MlflowException(
                "While parsing a list in the query, expected string value, punctuation, "
                f"or whitespace, but got different type in list: {value_token}",
                error_code=INVALID_PARAMETER_VALUE,
            )

    @classmethod
    def _parse_list_from_sql_token(cls, token):
        try:
            str_or_tuple = ast.literal_eval(token.value)
            return [str_or_tuple] if isinstance(str_or_tuple, str) else str_or_tuple
        except SyntaxError:
            raise MlflowException(
                "While parsing a list in the query,"
                " expected a non-empty list of string values, but got ill-formed list.",
                error_code=INVALID_PARAMETER_VALUE,
            )

    @classmethod
    def _parse_run_ids(cls, token):
        cls._check_valid_identifier_list(token)
        run_id_list = cls._parse_list_from_sql_token(token)
        # Because MySQL IN clause is case-insensitive, but all run_ids only contain lower
        # case letters, so that we filter out run_ids containing upper case letters here.
        return [run_id for run_id in run_id_list if run_id.islower()]


class SearchExperimentsUtils(SearchUtils):
    VALID_SEARCH_ATTRIBUTE_KEYS = {"name", "creation_time", "last_update_time"}
    VALID_ORDER_BY_ATTRIBUTE_KEYS = {"name", "experiment_id", "creation_time", "last_update_time"}
    NUMERIC_ATTRIBUTES = {"creation_time", "last_update_time"}

    @classmethod
    def _invalid_statement_token_search_experiments(cls, token):
        if (
            isinstance(token, Comparison)
            or token.is_whitespace
            or token.match(ttype=TokenType.Keyword, values=["AND"])
        ):
            return False
        return True

    @classmethod
    def _process_statement(cls, statement):
        tokens = _join_in_comparison_tokens(statement.tokens)
        invalids = list(filter(cls._invalid_statement_token_search_experiments, tokens))
        if len(invalids) > 0:
            invalid_clauses = ", ".join(map(str, invalids))
            raise MlflowException.invalid_parameter_value(
                f"Invalid clause(s) in filter string: {invalid_clauses}"
            )
        return [cls._get_comparison(t) for t in tokens if isinstance(t, Comparison)]

    @classmethod
    def _get_identifier(cls, identifier, valid_attributes):
        tokens = identifier.split(".", maxsplit=1)
        if len(tokens) == 1:
            key = tokens[0]
            identifier = cls._ATTRIBUTE_IDENTIFIER
        else:
            entity_type, key = tokens
            valid_entity_types = ("attribute", "tag", "tags")
            if entity_type not in valid_entity_types:
                raise MlflowException.invalid_parameter_value(
                    f"Invalid entity type '{entity_type}'. "
                    f"Valid entity types are {valid_entity_types}"
                )
            identifier = cls._valid_entity_type(entity_type)

        key = cls._trim_backticks(cls._strip_quotes(key))
        if identifier == cls._ATTRIBUTE_IDENTIFIER and key not in valid_attributes:
            raise MlflowException.invalid_parameter_value(
                f"Invalid attribute key '{key}' specified. Valid keys are '{valid_attributes}'"
            )
        return {"type": identifier, "key": key}

    @classmethod
    def _get_comparison(cls, comparison):
        stripped_comparison = [token for token in comparison.tokens if not token.is_whitespace]
        cls._validate_comparison(stripped_comparison)
        left, comparator, right = stripped_comparison
        comp = cls._get_identifier(left.value, cls.VALID_SEARCH_ATTRIBUTE_KEYS)
        comp["comparator"] = comparator.value
        comp["value"] = cls._get_value(comp.get("type"), comp.get("key"), right)
        return comp

    @classmethod
    def parse_order_by_for_search_experiments(cls, order_by):
        token_value, is_ascending = cls._parse_order_by_string(order_by)
        identifier = cls._get_identifier(token_value.strip(), cls.VALID_ORDER_BY_ATTRIBUTE_KEYS)
        return identifier["type"], identifier["key"], is_ascending

    @classmethod
    def is_attribute(cls, key_type, comparator):
        if key_type == cls._ATTRIBUTE_IDENTIFIER:
            if comparator not in cls.VALID_STRING_ATTRIBUTE_COMPARATORS:
                raise MlflowException(
                    f"Invalid comparator '{comparator}' not one of "
                    f"'{cls.VALID_STRING_ATTRIBUTE_COMPARATORS}'"
                )
            return True
        return False

    @classmethod
    def _does_experiment_match_clause(cls, experiment, sed):
        key_type = sed.get("type")
        key = sed.get("key")
        value = sed.get("value")
        comparator = sed.get("comparator").upper()

        if cls.is_string_attribute(key_type, key, comparator):
            lhs = getattr(experiment, key)
        elif cls.is_numeric_attribute(key_type, key, comparator):
            lhs = getattr(experiment, key)
            value = float(value)
        elif cls.is_tag(key_type, comparator):
            if key not in experiment.tags:
                return False
            lhs = experiment.tags.get(key, None)
            if lhs is None:
                return experiment
        else:
            raise MlflowException(
                f"Invalid search expression type '{key_type}'", error_code=INVALID_PARAMETER_VALUE
            )

        return SearchUtils.get_comparison_func(comparator)(lhs, value)

    @classmethod
    def filter(cls, experiments, filter_string):
        if not filter_string:
            return experiments
        parsed = cls.parse_search_filter(filter_string)

        def experiment_matches(experiment):
            return all(cls._does_experiment_match_clause(experiment, s) for s in parsed)

        return list(filter(experiment_matches, experiments))

    @classmethod
    def _get_sort_key(cls, order_by_list):
        order_by = []
        parsed_order_by = map(cls.parse_order_by_for_search_experiments, order_by_list)
        for type_, key, ascending in parsed_order_by:
            if type_ == "attribute":
                order_by.append((key, ascending))
            else:
                raise MlflowException.invalid_parameter_value(f"Invalid order_by entity: {type_}")

        # Add a tie-breaker
        if not any(key == "experiment_id" for key, _ in order_by):
            order_by.append(("experiment_id", False))

        # https://stackoverflow.com/a/56842689
        class _Sorter:
            def __init__(self, obj, ascending):
                self.obj = obj
                self.ascending = ascending

            # Only need < and == are needed for use as a key parameter in the sorted function
            def __eq__(self, other):
                return other.obj == self.obj

            def __lt__(self, other):
                if self.obj is None:
                    return False
                elif other.obj is None:
                    return True
                elif self.ascending:
                    return self.obj < other.obj
                else:
                    return other.obj < self.obj

        def _apply_sorter(experiment, key, ascending):
            attr = getattr(experiment, key)
            return _Sorter(attr, ascending)

        return lambda experiment: tuple(_apply_sorter(experiment, k, asc) for (k, asc) in order_by)

    @classmethod
    def sort(cls, experiments, order_by_list):
        return sorted(experiments, key=cls._get_sort_key(order_by_list))


# https://stackoverflow.com/a/56842689
class _Reversor:
    def __init__(self, obj):
        self.obj = obj

    # Only need < and == are needed for use as a key parameter in the sorted function
    def __eq__(self, other):
        return other.obj == self.obj

    def __lt__(self, other):
        return other.obj < self.obj


def _apply_reversor(model, key, ascending):
    attr = getattr(model, key)
    return attr if ascending else _Reversor(attr)


class SearchModelUtils(SearchUtils):
    NUMERIC_ATTRIBUTES = {"creation_timestamp", "last_updated_timestamp"}
    VALID_SEARCH_ATTRIBUTE_KEYS = {"name"}
    VALID_ORDER_BY_KEYS_REGISTERED_MODELS = {"name", "creation_timestamp", "last_updated_timestamp"}

    @classmethod
    def _does_registered_model_match_clauses(cls, model, sed):
        key_type = sed.get("type")
        key = sed.get("key")
        value = sed.get("value")
        comparator = sed.get("comparator").upper()

        # what comparators do we support here?
        if cls.is_string_attribute(key_type, key, comparator):
            lhs = getattr(model, key)
        elif cls.is_numeric_attribute(key_type, key, comparator):
            lhs = getattr(model, key)
            value = int(value)
        elif cls.is_tag(key_type, comparator):
            # if the filter doesn't apply, do we return False or?
            lhs = model.tags.get(key, None)
        else:
            raise MlflowException(
                f"Invalid search expression type '{key_type}'", error_code=INVALID_PARAMETER_VALUE
            )
        if lhs is None:
            return False

        return SearchUtils.get_comparison_func(comparator)(lhs, value)

    @classmethod
    def filter(cls, registered_models, filter_string):
        """Filters a set of registered models based on a search filter string."""
        if not filter_string:
            return registered_models
        parsed = cls.parse_search_filter(filter_string)

        def registered_model_matches(model):
            return all(cls._does_registered_model_match_clauses(model, s) for s in parsed)

        return [
            registered_model
            for registered_model in registered_models
            if registered_model_matches(registered_model)
        ]

    @classmethod
    def parse_order_by_for_search_registered_models(cls, order_by):
        token_value, is_ascending = cls._parse_order_by_string(order_by)
        identifier = SearchExperimentsUtils._get_identifier(
            token_value.strip(), cls.VALID_ORDER_BY_KEYS_REGISTERED_MODELS
        )
        return identifier["type"], identifier["key"], is_ascending

    @classmethod
    def _get_sort_key(cls, order_by_list):
        order_by = []
        parsed_order_by = map(cls.parse_order_by_for_search_registered_models, order_by_list or [])
        for type_, key, ascending in parsed_order_by:
            if type_ == "attribute":
                order_by.append((key, ascending))
            else:
                raise MlflowException.invalid_parameter_value(f"Invalid order_by entity: {type_}")

        # Add a tie-breaker
        if not any(key == "name" for key, _ in order_by):
            order_by.append(("name", True))

        return lambda model: tuple(_apply_reversor(model, k, asc) for (k, asc) in order_by)

    @classmethod
    def sort(cls, models, order_by_list):
        return sorted(models, key=cls._get_sort_key(order_by_list))

    @classmethod
    def _process_statement(cls, statement):
        tokens = _join_in_comparison_tokens(statement.tokens)
        invalids = list(filter(cls._invalid_statement_token_search_model_registry, tokens))
        if len(invalids) > 0:
            invalid_clauses = ", ".join(map(str, invalids))
            raise MlflowException.invalid_parameter_value(
                f"Invalid clause(s) in filter string: {invalid_clauses}"
            )
        return [cls._get_comparison(t) for t in tokens if isinstance(t, Comparison)]

    @classmethod
    def _get_model_search_identifier(cls, identifier, valid_attributes):
        tokens = identifier.split(".", maxsplit=1)
        if len(tokens) == 1:
            key = tokens[0]
            identifier = cls._ATTRIBUTE_IDENTIFIER
        else:
            entity_type, key = tokens
            valid_entity_types = ("attribute", "tag", "tags")
            if entity_type not in valid_entity_types:
                raise MlflowException.invalid_parameter_value(
                    f"Invalid entity type '{entity_type}'. "
                    f"Valid entity types are {valid_entity_types}"
                )
            identifier = (
                cls._TAG_IDENTIFIER if entity_type in ("tag", "tags") else cls._ATTRIBUTE_IDENTIFIER
            )

        if identifier == cls._ATTRIBUTE_IDENTIFIER and key not in valid_attributes:
            raise MlflowException.invalid_parameter_value(
                f"Invalid attribute key '{key}' specified. Valid keys are '{valid_attributes}'"
            )

        key = cls._trim_backticks(cls._strip_quotes(key))
        return {"type": identifier, "key": key}

    @classmethod
    def _get_comparison(cls, comparison):
        stripped_comparison = [token for token in comparison.tokens if not token.is_whitespace]
        cls._validate_comparison(stripped_comparison)
        left, comparator, right = stripped_comparison
        comp = cls._get_model_search_identifier(left.value, cls.VALID_SEARCH_ATTRIBUTE_KEYS)
        comp["comparator"] = comparator.value.upper()
        comp["value"] = cls._get_value(comp.get("type"), comp.get("key"), right)
        return comp

    @classmethod
    def _get_value(cls, identifier_type, key, token):
        if identifier_type == cls._TAG_IDENTIFIER:
            if token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            raise MlflowException(
                "Expected a quoted string value for "
                f"{identifier_type} (e.g. 'my-value'). Got value "
                f"{token.value}",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif identifier_type == cls._ATTRIBUTE_IDENTIFIER:
            if token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            elif isinstance(token, Parenthesis):
                if key != "run_id":
                    raise MlflowException(
                        "Only the 'run_id' attribute supports comparison with a list of quoted "
                        "string values.",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                return cls._parse_run_ids(token)
            else:
                raise MlflowException(
                    "Expected a quoted string value or a list of quoted string values for "
                    f"attributes. Got value {token.value}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
        else:
            # Expected to be either "param" or "metric".
            raise MlflowException(
                "Invalid identifier type. Expected one of "
                f"{[cls._ATTRIBUTE_IDENTIFIER, cls._TAG_IDENTIFIER]}.",
                error_code=INVALID_PARAMETER_VALUE,
            )

    @classmethod
    def _invalid_statement_token_search_model_registry(cls, token):
        if (
            isinstance(token, Comparison)
            or token.is_whitespace
            or token.match(ttype=TokenType.Keyword, values=["AND"])
        ):
            return False
        return True


class SearchModelVersionUtils(SearchUtils):
    NUMERIC_ATTRIBUTES = {"version_number", "creation_timestamp", "last_updated_timestamp"}
    VALID_SEARCH_ATTRIBUTE_KEYS = {
        "name",
        "version_number",
        "run_id",
        "source_path",
    }
    VALID_ORDER_BY_ATTRIBUTE_KEYS = {
        "name",
        "version_number",
        "creation_timestamp",
        "last_updated_timestamp",
    }
    VALID_STRING_ATTRIBUTE_COMPARATORS = {"!=", "=", "LIKE", "ILIKE", "IN"}

    @classmethod
    def _does_model_version_match_clauses(cls, mv, sed):
        key_type = sed.get("type")
        key = sed.get("key")
        value = sed.get("value")
        comparator = sed.get("comparator").upper()

        if cls.is_string_attribute(key_type, key, comparator):
            lhs = getattr(mv, "source" if key == "source_path" else key)
        elif cls.is_numeric_attribute(key_type, key, comparator):
            if key == "version_number":
                key = "version"
            lhs = getattr(mv, key)
            value = int(value)
        elif cls.is_tag(key_type, comparator):
            lhs = mv.tags.get(key, None)
        else:
            raise MlflowException(
                f"Invalid search expression type '{key_type}'", error_code=INVALID_PARAMETER_VALUE
            )
        if lhs is None:
            return False

        if comparator == "IN" and isinstance(value, (set, list)):
            return lhs in set(value)

        return SearchUtils.get_comparison_func(comparator)(lhs, value)

    @classmethod
    def filter(cls, model_versions, filter_string):
        """Filters a set of model versions based on a search filter string."""
        model_versions = [mv for mv in model_versions if mv.current_stage != STAGE_DELETED_INTERNAL]
        if not filter_string:
            return model_versions
        parsed = cls.parse_search_filter(filter_string)

        def model_version_matches(mv):
            return all(cls._does_model_version_match_clauses(mv, s) for s in parsed)

        return [mv for mv in model_versions if model_version_matches(mv)]

    @classmethod
    def parse_order_by_for_search_model_versions(cls, order_by):
        token_value, is_ascending = cls._parse_order_by_string(order_by)
        identifier = SearchExperimentsUtils._get_identifier(
            token_value.strip(), cls.VALID_ORDER_BY_ATTRIBUTE_KEYS
        )
        return identifier["type"], identifier["key"], is_ascending

    @classmethod
    def _get_sort_key(cls, order_by_list):
        order_by = []
        parsed_order_by = map(cls.parse_order_by_for_search_model_versions, order_by_list or [])
        for type_, key, ascending in parsed_order_by:
            if type_ == "attribute":
                # Need to add this mapping because version is a keyword in sql
                if key == "version_number":
                    key = "version"
                order_by.append((key, ascending))
            else:
                raise MlflowException.invalid_parameter_value(f"Invalid order_by entity: {type_}")

        # Add a tie-breaker
        if not any(key == "name" for key, _ in order_by):
            order_by.append(("name", True))
        if not any(key == "version_number" for key, _ in order_by):
            order_by.append(("version", False))

        return lambda model_version: tuple(
            _apply_reversor(model_version, k, asc) for (k, asc) in order_by
        )

    @classmethod
    def sort(cls, model_versions, order_by_list):
        return sorted(model_versions, key=cls._get_sort_key(order_by_list))

    @classmethod
    def _get_model_version_search_identifier(cls, identifier, valid_attributes):
        tokens = identifier.split(".", maxsplit=1)
        if len(tokens) == 1:
            key = tokens[0]
            identifier = cls._ATTRIBUTE_IDENTIFIER
        else:
            entity_type, key = tokens
            valid_entity_types = ("attribute", "tag", "tags")
            if entity_type not in valid_entity_types:
                raise MlflowException.invalid_parameter_value(
                    f"Invalid entity type '{entity_type}'. "
                    f"Valid entity types are {valid_entity_types}"
                )
            identifier = (
                cls._TAG_IDENTIFIER if entity_type in ("tag", "tags") else cls._ATTRIBUTE_IDENTIFIER
            )

        if identifier == cls._ATTRIBUTE_IDENTIFIER and key not in valid_attributes:
            raise MlflowException.invalid_parameter_value(
                f"Invalid attribute key '{key}' specified. Valid keys are '{valid_attributes}'"
            )

        key = cls._trim_backticks(cls._strip_quotes(key))
        return {"type": identifier, "key": key}

    @classmethod
    def _get_comparison(cls, comparison):
        stripped_comparison = [token for token in comparison.tokens if not token.is_whitespace]
        cls._validate_comparison(stripped_comparison)
        left, comparator, right = stripped_comparison
        comp = cls._get_model_version_search_identifier(left.value, cls.VALID_SEARCH_ATTRIBUTE_KEYS)
        comp["comparator"] = comparator.value.upper()
        comp["value"] = cls._get_value(comp.get("type"), comp.get("key"), right)
        return comp

    @classmethod
    def _get_value(cls, identifier_type, key, token):
        if identifier_type == cls._TAG_IDENTIFIER:
            if token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            raise MlflowException(
                "Expected a quoted string value for "
                f"{identifier_type} (e.g. 'my-value'). Got value "
                f"{token.value}",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif identifier_type == cls._ATTRIBUTE_IDENTIFIER:
            if token.ttype in cls.STRING_VALUE_TYPES or isinstance(token, Identifier):
                return cls._strip_quotes(token.value, expect_quoted_value=True)
            elif isinstance(token, Parenthesis):
                if key != "run_id":
                    raise MlflowException(
                        "Only the 'run_id' attribute supports comparison with a list of quoted "
                        "string values.",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                return cls._parse_run_ids(token)
            elif token.ttype in cls.NUMERIC_VALUE_TYPES:
                if key not in cls.NUMERIC_ATTRIBUTES:
                    raise MlflowException(
                        f"Only the '{cls.NUMERIC_ATTRIBUTES}' attributes support comparison with "
                        "numeric values.",
                        error_code=INVALID_PARAMETER_VALUE,
                    )
                if token.ttype == TokenType.Literal.Number.Integer:
                    return int(token.value)
                elif token.ttype == TokenType.Literal.Number.Float:
                    return float(token.value)
            else:
                raise MlflowException(
                    "Expected a quoted string value or a list of quoted string values for "
                    f"attributes. Got value {token.value}",
                    error_code=INVALID_PARAMETER_VALUE,
                )
        else:
            # Expected to be either "param" or "metric".
            raise MlflowException(
                "Invalid identifier type. Expected one of "
                f"{[cls._ATTRIBUTE_IDENTIFIER, cls._TAG_IDENTIFIER]}.",
                error_code=INVALID_PARAMETER_VALUE,
            )

    @classmethod
    def _process_statement(cls, statement):
        tokens = _join_in_comparison_tokens(statement.tokens)
        invalids = list(filter(cls._invalid_statement_token_search_model_version, tokens))
        if len(invalids) > 0:
            invalid_clauses = ", ".join(map(str, invalids))
            raise MlflowException.invalid_parameter_value(
                f"Invalid clause(s) in filter string: {invalid_clauses}"
            )
        return [cls._get_comparison(t) for t in tokens if isinstance(t, Comparison)]

    @classmethod
    def _invalid_statement_token_search_model_version(cls, token):
        if (
            isinstance(token, Comparison)
            or token.is_whitespace
            or token.match(ttype=TokenType.Keyword, values=["AND"])
        ):
            return False
        return True

    @classmethod
    def parse_search_filter(cls, filter_string):
        if not filter_string:
            return []
        try:
            parsed = sqlparse.parse(filter_string)
        except Exception:
            raise MlflowException(
                f"Error on parsing filter '{filter_string}'", error_code=INVALID_PARAMETER_VALUE
            )
        if len(parsed) == 0 or not isinstance(parsed[0], Statement):
            raise MlflowException(
                f"Invalid filter '{filter_string}'. Could not be parsed.",
                error_code=INVALID_PARAMETER_VALUE,
            )
        elif len(parsed) > 1:
            raise MlflowException(
                "Search filter contained multiple expression '%s'. "
                "Provide AND-ed expression list." % filter_string,
                error_code=INVALID_PARAMETER_VALUE,
            )
        return cls._process_statement(parsed[0])