from __future__ import annotations
import datetime
import decimal
from abc import ABC, abstractmethod
from collections import OrderedDict, namedtuple
from collections.abc import Iterable
from decimal import Decimal
from enum import Enum
from typing import Any, Dict, List, Union
import re
import lz4.frame
import pyarrow
from databricks.sql import OperationalError, exc
from databricks.sql.cloudfetch.download_manager import ResultFileDownloadManager
from databricks.sql.thrift_api.TCLIService.ttypes import (
TRowSet,
TSparkArrowResultLink,
TSparkRowSetType,
)
from databricks.sql.parameters.native import ParameterStructure, TDbsqlParameter
BIT_MASKS = [1, 2, 4, 8, 16, 32, 64, 128]
import logging
logger = logging.getLogger(__name__)
class ResultSetQueue(ABC):
@abstractmethod
def next_n_rows(self, num_rows: int) -> pyarrow.Table:
pass
@abstractmethod
def remaining_rows(self) -> pyarrow.Table:
pass
class ResultSetQueueFactory(ABC):
@staticmethod
def build_queue(
row_set_type: TSparkRowSetType,
t_row_set: TRowSet,
arrow_schema_bytes: bytes,
max_download_threads: int,
lz4_compressed: bool = True,
description: List[List[Any]] = None,
) -> ResultSetQueue:
"""
Factory method to build a result set queue.
Args:
row_set_type (enum): Row set type (Arrow, Column, or URL).
t_row_set (TRowSet): Result containing arrow batches, columns, or cloud fetch links.
arrow_schema_bytes (bytes): Bytes representing the arrow schema.
lz4_compressed (bool): Whether result data has been lz4 compressed.
description (List[List[Any]]): Hive table schema description.
max_download_threads (int): Maximum number of downloader thread pool threads.
Returns:
ResultSetQueue
"""
if row_set_type == TSparkRowSetType.ARROW_BASED_SET:
arrow_table, n_valid_rows = convert_arrow_based_set_to_arrow_table(
t_row_set.arrowBatches, lz4_compressed, arrow_schema_bytes
)
converted_arrow_table = convert_decimals_in_arrow_table(
arrow_table, description
)
return ArrowQueue(converted_arrow_table, n_valid_rows)
elif row_set_type == TSparkRowSetType.COLUMN_BASED_SET:
arrow_table, n_valid_rows = convert_column_based_set_to_arrow_table(
t_row_set.columns, description
)
converted_arrow_table = convert_decimals_in_arrow_table(
arrow_table, description
)
return ArrowQueue(converted_arrow_table, n_valid_rows)
elif row_set_type == TSparkRowSetType.URL_BASED_SET:
return CloudFetchQueue(
arrow_schema_bytes,
start_row_offset=t_row_set.startRowOffset,
result_links=t_row_set.resultLinks,
lz4_compressed=lz4_compressed,
description=description,
max_download_threads=max_download_threads,
)
else:
raise AssertionError("Row set type is not valid")
class ArrowQueue(ResultSetQueue):
def __init__(
self,
arrow_table: pyarrow.Table,
n_valid_rows: int,
start_row_index: int = 0,
):
"""
A queue-like wrapper over an Arrow table
:param arrow_table: The Arrow table from which we want to take rows
:param n_valid_rows: The index of the last valid row in the table
:param start_row_index: The first row in the table we should start fetching from
"""
self.cur_row_index = start_row_index
self.arrow_table = arrow_table
self.n_valid_rows = n_valid_rows
def next_n_rows(self, num_rows: int) -> pyarrow.Table:
"""Get upto the next n rows of the Arrow dataframe"""
length = min(num_rows, self.n_valid_rows - self.cur_row_index)
# Note that the table.slice API is not the same as Python's slice
# The second argument should be length, not end index
slice = self.arrow_table.slice(self.cur_row_index, length)
self.cur_row_index += slice.num_rows
return slice
def remaining_rows(self) -> pyarrow.Table:
slice = self.arrow_table.slice(
self.cur_row_index, self.n_valid_rows - self.cur_row_index
)
self.cur_row_index += slice.num_rows
return slice
class CloudFetchQueue(ResultSetQueue):
def __init__(
self,
schema_bytes,
max_download_threads: int,
start_row_offset: int = 0,
result_links: List[TSparkArrowResultLink] = None,
lz4_compressed: bool = True,
description: List[List[Any]] = None,
):
"""
A queue-like wrapper over CloudFetch arrow batches.
Attributes:
schema_bytes (bytes): Table schema in bytes.
max_download_threads (int): Maximum number of downloader thread pool threads.
start_row_offset (int): The offset of the first row of the cloud fetch links.
result_links (List[TSparkArrowResultLink]): Links containing the downloadable URL and metadata.
lz4_compressed (bool): Whether the files are lz4 compressed.
description (List[List[Any]]): Hive table schema description.
"""
self.schema_bytes = schema_bytes
self.max_download_threads = max_download_threads
self.start_row_index = start_row_offset
self.result_links = result_links
self.lz4_compressed = lz4_compressed
self.description = description
self.download_manager = ResultFileDownloadManager(
self.max_download_threads, self.lz4_compressed
)
self.download_manager.add_file_links(result_links)
self.table = self._create_next_table()
self.table_row_index = 0
def next_n_rows(self, num_rows: int) -> pyarrow.Table:
"""
Get up to the next n rows of the cloud fetch Arrow dataframes.
Args:
num_rows (int): Number of rows to retrieve.
Returns:
pyarrow.Table
"""
if not self.table:
# Return empty pyarrow table to cause retry of fetch
return self._create_empty_table()
results = self.table.slice(0, 0)
while num_rows > 0 and self.table:
# Get remaining of num_rows or the rest of the current table, whichever is smaller
length = min(num_rows, self.table.num_rows - self.table_row_index)
table_slice = self.table.slice(self.table_row_index, length)
results = pyarrow.concat_tables([results, table_slice])
self.table_row_index += table_slice.num_rows
# Replace current table with the next table if we are at the end of the current table
if self.table_row_index == self.table.num_rows:
self.table = self._create_next_table()
self.table_row_index = 0
num_rows -= table_slice.num_rows
return results
def remaining_rows(self) -> pyarrow.Table:
"""
Get all remaining rows of the cloud fetch Arrow dataframes.
Returns:
pyarrow.Table
"""
if not self.table:
# Return empty pyarrow table to cause retry of fetch
return self._create_empty_table()
results = self.table.slice(0, 0)
while self.table:
table_slice = self.table.slice(
self.table_row_index, self.table.num_rows - self.table_row_index
)
results = pyarrow.concat_tables([results, table_slice])
self.table_row_index += table_slice.num_rows
self.table = self._create_next_table()
self.table_row_index = 0
return results
def _create_next_table(self) -> Union[pyarrow.Table, None]:
# Create next table by retrieving the logical next downloaded file, or return None to signal end of queue
downloaded_file = self.download_manager.get_next_downloaded_file(
self.start_row_index
)
if not downloaded_file:
# None signals no more Arrow tables can be built from the remaining handlers if any remain
return None
arrow_table = create_arrow_table_from_arrow_file(
downloaded_file.file_bytes, self.description
)
# The server rarely prepares the exact number of rows requested by the client in cloud fetch.
# Subsequently, we drop the extraneous rows in the last file if more rows are retrieved than requested
if arrow_table.num_rows > downloaded_file.row_count:
self.start_row_index += downloaded_file.row_count
return arrow_table.slice(0, downloaded_file.row_count)
# At this point, whether the file has extraneous rows or not, the arrow table should have the correct num rows
assert downloaded_file.row_count == arrow_table.num_rows
self.start_row_index += arrow_table.num_rows
return arrow_table
def _create_empty_table(self) -> pyarrow.Table:
# Create a 0-row table with just the schema bytes
return create_arrow_table_from_arrow_file(self.schema_bytes, self.description)
ExecuteResponse = namedtuple(
"ExecuteResponse",
"status has_been_closed_server_side has_more_rows description lz4_compressed is_staging_operation "
"command_handle arrow_queue arrow_schema_bytes",
)
def _bound(min_x, max_x, x):
"""Bound x by [min_x, max_x]
min_x or max_x being None means unbounded in that respective side.
"""
if min_x is None and max_x is None:
return x
if min_x is None:
return min(max_x, x)
if max_x is None:
return max(min_x, x)
return min(max_x, max(min_x, x))
class NoRetryReason(Enum):
OUT_OF_TIME = "out of time"
OUT_OF_ATTEMPTS = "out of attempts"
NOT_RETRYABLE = "non-retryable error"
class RequestErrorInfo(
namedtuple(
"RequestErrorInfo_", "error error_message retry_delay http_code method request"
)
):
@property
def request_session_id(self):
if hasattr(self.request, "sessionHandle"):
return self.request.sessionHandle.sessionId.guid
else:
return None
@property
def request_query_id(self):
if hasattr(self.request, "operationHandle"):
return self.request.operationHandle.operationId.guid
else:
return None
def full_info_logging_context(
self, no_retry_reason, attempt, max_attempts, elapsed, max_duration
):
log_base_data_dict = OrderedDict(
[
("method", self.method),
("session-id", self.request_session_id),
("query-id", self.request_query_id),
("http-code", self.http_code),
("error-message", self.error_message),
("original-exception", str(self.error)),
]
)
log_base_data_dict["no-retry-reason"] = (
no_retry_reason and no_retry_reason.value
)
log_base_data_dict["bounded-retry-delay"] = self.retry_delay
log_base_data_dict["attempt"] = "{}/{}".format(attempt, max_attempts)
log_base_data_dict["elapsed-seconds"] = "{}/{}".format(elapsed, max_duration)
return log_base_data_dict
def user_friendly_error_message(self, no_retry_reason, attempt, elapsed):
# This should be kept at the level that is appropriate to return to a Redash user
user_friendly_error_message = "Error during request to server"
if self.error_message:
user_friendly_error_message = "{}: {}".format(
user_friendly_error_message, self.error_message
)
return user_friendly_error_message
# Taken from PyHive
class ParamEscaper:
_DATE_FORMAT = "%Y-%m-%d"
_TIME_FORMAT = "%H:%M:%S.%f"
_DATETIME_FORMAT = "{} {}".format(_DATE_FORMAT, _TIME_FORMAT)
def escape_args(self, parameters):
if isinstance(parameters, dict):
return {k: self.escape_item(v) for k, v in parameters.items()}
elif isinstance(parameters, (list, tuple)):
return tuple(self.escape_item(x) for x in parameters)
else:
raise exc.ProgrammingError(
"Unsupported param format: {}".format(parameters)
)
def escape_number(self, item):
return item
def escape_string(self, item):
# Need to decode UTF-8 because of old sqlalchemy.
# Newer SQLAlchemy checks dialect.supports_unicode_binds before encoding Unicode strings
# as byte strings. The old version always encodes Unicode as byte strings, which breaks
# string formatting here.
if isinstance(item, bytes):
item = item.decode("utf-8")
# This is good enough when backslashes are literal, newlines are just followed, and the way
# to escape a single quote is to put two single quotes.
# (i.e. only special character is single quote)
return "'{}'".format(item.replace("\\", "\\\\").replace("'", "\\'"))
def escape_sequence(self, item):
l = map(str, map(self.escape_item, item))
return "(" + ",".join(l) + ")"
def escape_datetime(self, item, format, cutoff=0):
dt_str = item.strftime(format)
formatted = dt_str[:-cutoff] if cutoff and format.endswith(".%f") else dt_str
return "'{}'".format(formatted)
def escape_decimal(self, item):
return str(item)
def escape_item(self, item):
if item is None:
return "NULL"
elif isinstance(item, (int, float)):
return self.escape_number(item)
elif isinstance(item, str):
return self.escape_string(item)
elif isinstance(item, Iterable):
return self.escape_sequence(item)
elif isinstance(item, datetime.datetime):
return self.escape_datetime(item, self._DATETIME_FORMAT)
elif isinstance(item, datetime.date):
return self.escape_datetime(item, self._DATE_FORMAT)
elif isinstance(item, decimal.Decimal):
return self.escape_decimal(item)
else:
raise exc.ProgrammingError("Unsupported object {}".format(item))
def inject_parameters(operation: str, parameters: Dict[str, str]):
return operation % parameters
def _dbsqlparameter_names(params: List[TDbsqlParameter]) -> list[str]:
return [p.name if p.name else "" for p in params]
def _generate_named_interpolation_values(
params: List[TDbsqlParameter],
) -> dict[str, str]:
"""Returns a dictionary of the form {name: ":name"} for each parameter in params"""
names = _dbsqlparameter_names(params)
return {name: f":{name}" for name in names}
def _may_contain_inline_positional_markers(operation: str) -> bool:
"""Check for the presence of `%s` in the operation string."""
interpolated = operation.replace("%s", "?")
return interpolated != operation
def _interpolate_named_markers(
operation: str, parameters: List[TDbsqlParameter]
) -> str:
"""Replace all instances of `%(param)s` in `operation` with `:param`.
If `operation` contains no instances of `%(param)s` then the input string is returned unchanged.
```
"SELECT * FROM table WHERE field = %(field)s and other_field = %(other_field)s"
```
Yields
```
SELECT * FROM table WHERE field = :field and other_field = :other_field
```
"""
_output_operation = operation
PYFORMAT_PARAMSTYLE_REGEX = r"%\((\w+)\)s"
pat = re.compile(PYFORMAT_PARAMSTYLE_REGEX)
NAMED_PARAMSTYLE_FMT = ":{}"
PYFORMAT_PARAMSTYLE_FMT = "%({})s"
pyformat_markers = pat.findall(operation)
for marker in pyformat_markers:
pyformat_marker = PYFORMAT_PARAMSTYLE_FMT.format(marker)
named_marker = NAMED_PARAMSTYLE_FMT.format(marker)
_output_operation = _output_operation.replace(pyformat_marker, named_marker)
return _output_operation
def transform_paramstyle(
operation: str,
parameters: List[TDbsqlParameter],
param_structure: ParameterStructure,
) -> str:
"""
Performs a Python string interpolation such that any occurence of `%(param)s` will be replaced with `:param`
This utility function is built to assist users in the transition between the default paramstyle in
this connector prior to version 3.0.0 (`pyformat`) and the new default paramstyle (`named`).
Args:
operation: The operation or SQL text to transform.
parameters: The parameters to use for the transformation.
Returns:
str
"""
output = operation
if (
param_structure == ParameterStructure.POSITIONAL
and _may_contain_inline_positional_markers(operation)
):
logger.warning(
"It looks like this query may contain un-named query markers like `%s`"
" This format is not supported when use_inline_params=False."
" Use `?` instead or set use_inline_params=True"
)
elif param_structure == ParameterStructure.NAMED:
output = _interpolate_named_markers(operation, parameters)
return output
def create_arrow_table_from_arrow_file(file_bytes: bytes, description) -> pyarrow.Table:
arrow_table = convert_arrow_based_file_to_arrow_table(file_bytes)
return convert_decimals_in_arrow_table(arrow_table, description)
def convert_arrow_based_file_to_arrow_table(file_bytes: bytes):
try:
return pyarrow.ipc.open_stream(file_bytes).read_all()
except Exception as e:
raise RuntimeError("Failure to convert arrow based file to arrow table", e)
def convert_arrow_based_set_to_arrow_table(arrow_batches, lz4_compressed, schema_bytes):
ba = bytearray()
ba += schema_bytes
n_rows = 0
for arrow_batch in arrow_batches:
n_rows += arrow_batch.rowCount
ba += (
lz4.frame.decompress(arrow_batch.batch)
if lz4_compressed
else arrow_batch.batch
)
arrow_table = pyarrow.ipc.open_stream(ba).read_all()
return arrow_table, n_rows
def convert_decimals_in_arrow_table(table, description) -> pyarrow.Table:
for i, col in enumerate(table.itercolumns()):
if description[i][1] == "decimal":
decimal_col = col.to_pandas().apply(
lambda v: v if v is None else Decimal(v)
)
precision, scale = description[i][4], description[i][5]
assert scale is not None
assert precision is not None
# Spark limits decimal to a maximum scale of 38,
# so 128 is guaranteed to be big enough
dtype = pyarrow.decimal128(precision, scale)
col_data = pyarrow.array(decimal_col, type=dtype)
field = table.field(i).with_type(dtype)
table = table.set_column(i, field, col_data)
return table
def convert_column_based_set_to_arrow_table(columns, description):
arrow_table = pyarrow.Table.from_arrays(
[_convert_column_to_arrow_array(c) for c in columns],
# Only use the column names from the schema, the types are determined by the
# physical types used in column based set, as they can differ from the
# mapping used in _hive_schema_to_arrow_schema.
names=[c[0] for c in description],
)
return arrow_table, arrow_table.num_rows
def _convert_column_to_arrow_array(t_col):
"""
Return a pyarrow array from the values in a TColumn instance.
Note that ColumnBasedSet has no native support for complex types, so they will be converted
to strings server-side.
"""
field_name_to_arrow_type = {
"boolVal": pyarrow.bool_(),
"byteVal": pyarrow.int8(),
"i16Val": pyarrow.int16(),
"i32Val": pyarrow.int32(),
"i64Val": pyarrow.int64(),
"doubleVal": pyarrow.float64(),
"stringVal": pyarrow.string(),
"binaryVal": pyarrow.binary(),
}
for field in field_name_to_arrow_type.keys():
wrapper = getattr(t_col, field)
if wrapper:
return _create_arrow_array(wrapper, field_name_to_arrow_type[field])
raise OperationalError("Empty TColumn instance {}".format(t_col))
def _create_arrow_array(t_col_value_wrapper, arrow_type):
result = t_col_value_wrapper.values
nulls = t_col_value_wrapper.nulls # bitfield describing which values are null
assert isinstance(nulls, bytes)
# The number of bits in nulls can be both larger or smaller than the number of
# elements in result, so take the minimum of both to iterate over.
length = min(len(result), len(nulls) * 8)
for i in range(length):
if nulls[i >> 3] & BIT_MASKS[i & 0x7]:
result[i] = None
return pyarrow.array(result, type=arrow_type)