sdks/python/apache_beam/metrics/cells.py

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements.  See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License.  You may obtain a copy of the License at
#
#    http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#

"""
This file contains metric cell classes. A metric cell is used to accumulate
in-memory changes to a metric. It represents a specific metric in a single
context.
"""

# pytype: skip-file

# ruff: noqa: UP006
import copy
import logging
import threading
import time
from datetime import datetime
from datetime import timezone
from typing import Iterable
from typing import Optional
from typing import Set

from apache_beam.portability.api import metrics_pb2
from apache_beam.utils.histogram import Histogram

try:
  import cython
except ImportError:

  class fake_cython:
    compiled = False

  globals()['cython'] = fake_cython

__all__ = [
    'MetricCell', 'MetricCellFactory', 'DistributionResult', 'GaugeResult'
]

_LOGGER = logging.getLogger(__name__)


class MetricCell(object):
  """For internal use only; no backwards-compatibility guarantees.

  Accumulates in-memory changes to a metric.

  A MetricCell represents a specific metric in a single context and bundle.
  All subclasses must be thread safe, as these are used in the pipeline runners,
  and may be subject to parallel/concurrent updates. Cells should only be used
  directly within a runner.
  """
  def __init__(self):
    self._lock = threading.Lock()
    self._start_time = None

  def update(self, value):
    raise NotImplementedError

  def get_cumulative(self):
    raise NotImplementedError

  def to_runner_api_monitoring_info(self, name, transform_id):
    if not self._start_time:
      self._start_time = datetime.now(timezone.utc)
    mi = self.to_runner_api_monitoring_info_impl(name, transform_id)
    mi.start_time.FromDatetime(self._start_time)
    return mi

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    raise NotImplementedError

  def reset(self):
    # type: () -> None
    raise NotImplementedError

  def __reduce__(self):
    raise NotImplementedError


class MetricCellFactory(object):
  def __call__(self):
    # type: () -> MetricCell
    raise NotImplementedError


class CounterCell(MetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value and delta of a counter metric.

  Each cell tracks the state of an integer metric independently per context
  per bundle. Therefore, each metric has a different cell in each bundle,
  cells are aggregated by the runner.

  This class is thread safe.
  """
  def __init__(self, *args):
    super().__init__(*args)
    self.value = 0

  def reset(self):
    # type: () -> None
    self.value = 0

  def combine(self, other):
    # type: (CounterCell) -> CounterCell
    result = CounterCell()
    result.inc(self.value + other.value)
    return result

  def inc(self, n=1):
    self.update(n)

  def dec(self, n=1):
    self.update(-n)

  def update(self, value):
    # type: (int) -> None
    if cython.compiled:
      ivalue = value
      # Since We hold the GIL, no need for another lock.
      # And because the C threads won't preempt and interleave
      # each other.
      # Assuming there is no code trying to access the counters
      # directly by circumventing the GIL.
      self.value += ivalue
    else:
      with self._lock:
        self.value += value

  def get_cumulative(self):
    # type: () -> int
    with self._lock:
      return self.value

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    from apache_beam.metrics import monitoring_infos
    if not name.urn:
      # User counter case.
      return monitoring_infos.int64_user_counter(
          name.namespace,
          name.name,
          self.get_cumulative(),
          ptransform=transform_id)
    else:
      # Arbitrary URN case.
      return monitoring_infos.int64_counter(
          name.urn, self.get_cumulative(), labels=name.labels)


class DistributionCell(MetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value and delta for a distribution metric.

  Each cell tracks the state of a metric independently per context per bundle.
  Therefore, each metric has a different cell in each bundle, that is later
  aggregated.

  This class is thread safe.
  """
  def __init__(self, *args):
    super().__init__(*args)
    self.data = DistributionData.identity_element()

  def reset(self):
    # type: () -> None
    self.data = DistributionData.identity_element()

  def combine(self, other):
    # type: (DistributionCell) -> DistributionCell
    result = DistributionCell()
    result.data = self.data.combine(other.data)
    return result

  def update(self, value):
    if cython.compiled:
      # We will hold the GIL throughout the entire _update.
      self._update(value)
    else:
      with self._lock:
        self._update(value)

  def _update(self, value):
    if cython.compiled:
      ivalue = value
    else:
      ivalue = int(value)
    self.data.count = self.data.count + 1
    self.data.sum = self.data.sum + ivalue
    if ivalue < self.data.min:
      self.data.min = ivalue
    if ivalue > self.data.max:
      self.data.max = ivalue

  def get_cumulative(self):
    # type: () -> DistributionData
    with self._lock:
      return self.data.get_cumulative()

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    from apache_beam.metrics import monitoring_infos
    return monitoring_infos.int64_user_distribution(
        name.namespace,
        name.name,
        self.get_cumulative(),
        ptransform=transform_id)


class AbstractMetricCell(MetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value and delta for a metric with a data class.

  This class is thread safe.
  """
  def __init__(self, data_class):
    super().__init__()
    self.data_class = data_class
    self.data = self.data_class.identity_element()

  def reset(self):
    self.data = self.data_class.identity_element()

  def combine(self, other: 'AbstractMetricCell') -> 'AbstractMetricCell':
    result = type(self)()  # type: ignore[call-arg]
    result.data = self.data.combine(other.data)
    return result

  def set(self, value):
    with self._lock:
      self._update_locked(value)

  def update(self, value):
    with self._lock:
      self._update_locked(value)

  def _update_locked(self, value):
    raise NotImplementedError(type(self))

  def get_cumulative(self):
    with self._lock:
      return self.data.get_cumulative()

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    raise NotImplementedError(type(self))


class GaugeCell(AbstractMetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value and delta for a gauge metric.

  Each cell tracks the state of a metric independently per context per bundle.
  Therefore, each metric has a different cell in each bundle, that is later
  aggregated.

  This class is thread safe.
  """
  def __init__(self):
    super().__init__(GaugeData)

  def _update_locked(self, value):
    # Set the value directly without checking timestamp, because
    # this value is naturally the latest value.
    self.data.value = int(value)
    self.data.timestamp = time.time()

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    from apache_beam.metrics import monitoring_infos
    return monitoring_infos.int64_user_gauge(
        name.namespace,
        name.name,
        self.get_cumulative(),
        ptransform=transform_id)


class StringSetCell(AbstractMetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value for a StringSet metric.

  Each cell tracks the state of a metric independently per context per bundle.
  Therefore, each metric has a different cell in each bundle, that is later
  aggregated.

  This class is thread safe.
  """
  def __init__(self):
    super().__init__(StringSetData)

  def add(self, value):
    self.update(value)

  def _update_locked(self, value):
    self.data.add(value)

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    from apache_beam.metrics import monitoring_infos
    return monitoring_infos.user_set_string(
        name.namespace,
        name.name,
        self.get_cumulative(),
        ptransform=transform_id)


class BoundedTrieCell(AbstractMetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value for a BoundedTrie metric.

  Each cell tracks the state of a metric independently per context per bundle.
  Therefore, each metric has a different cell in each bundle, that is later
  aggregated.

  This class is thread safe.
  """
  def __init__(self):
    super().__init__(BoundedTrieData)

  def add(self, value):
    self.update(value)

  def _update_locked(self, value):
    self.data.add(value)

  def to_runner_api_monitoring_info_impl(self, name, transform_id):
    from apache_beam.metrics import monitoring_infos
    return monitoring_infos.user_bounded_trie(
        name.namespace,
        name.name,
        self.get_cumulative(),
        ptransform=transform_id)


class DistributionResult(object):
  """The result of a Distribution metric."""
  def __init__(self, data):
    # type: (DistributionData) -> None
    self.data = data

  def __eq__(self, other):
    # type: (object) -> bool
    if isinstance(other, DistributionResult):
      return self.data == other.data
    else:
      return False

  def __hash__(self):
    # type: () -> int
    return hash(self.data)

  def __repr__(self):
    # type: () -> str
    return (
        'DistributionResult(sum={}, count={}, min={}, max={}, '
        'mean={})'.format(self.sum, self.count, self.min, self.max, self.mean))

  @property
  def max(self):
    # type: () -> Optional[int]
    return self.data.max if self.data.count else None

  @property
  def min(self):
    # type: () -> Optional[int]
    return self.data.min if self.data.count else None

  @property
  def count(self):
    # type: () -> Optional[int]
    return self.data.count

  @property
  def sum(self):
    # type: () -> Optional[int]
    return self.data.sum

  @property
  def mean(self):
    # type: () -> Optional[float]

    """Returns the float mean of the distribution.

    If the distribution contains no elements, it returns None.
    """
    if self.data.count == 0:
      return None
    return self.data.sum / self.data.count


class GaugeResult(object):
  def __init__(self, data):
    # type: (GaugeData) -> None
    self.data = data

  def __eq__(self, other):
    # type: (object) -> bool
    if isinstance(other, GaugeResult):
      return self.data == other.data
    else:
      return False

  def __hash__(self):
    # type: () -> int
    return hash(self.data)

  def __repr__(self):
    return '<GaugeResult(value={}, timestamp={})>'.format(
        self.value, self.timestamp)

  @property
  def value(self):
    # type: () -> Optional[int]
    return self.data.value

  @property
  def timestamp(self):
    # type: () -> Optional[int]
    return self.data.timestamp


class GaugeData(object):
  """For internal use only; no backwards-compatibility guarantees.

  The data structure that holds data about a gauge metric.

  Gauge metrics are restricted to integers only.

  This object is not thread safe, so it's not supposed to be modified
  by other than the GaugeCell that contains it.
  """
  def __init__(self, value, timestamp=None):
    # type: (Optional[int], Optional[int]) -> None
    self.value = value
    self.timestamp = timestamp if timestamp is not None else 0

  def __eq__(self, other):
    # type: (object) -> bool
    if isinstance(other, GaugeData):
      return self.value == other.value and self.timestamp == other.timestamp
    else:
      return False

  def __hash__(self):
    # type: () -> int
    return hash((self.value, self.timestamp))

  def __repr__(self):
    # type: () -> str
    return '<GaugeData(value={}, timestamp={})>'.format(
        self.value, self.timestamp)

  def get_cumulative(self):
    # type: () -> GaugeData
    return GaugeData(self.value, timestamp=self.timestamp)

  def get_result(self):
    # type: () -> GaugeResult
    return GaugeResult(self.get_cumulative())

  def combine(self, other):
    # type: (Optional[GaugeData]) -> GaugeData
    if other is None:
      return self

    if other.timestamp > self.timestamp:
      return other
    else:
      return self

  @staticmethod
  def singleton(value, timestamp=None):
    # type: (Optional[int], Optional[int]) -> GaugeData
    return GaugeData(value, timestamp=timestamp)

  @staticmethod
  def identity_element():
    # type: () -> GaugeData
    return GaugeData(0, timestamp=0)


class DistributionData(object):
  """For internal use only; no backwards-compatibility guarantees.

  The data structure that holds data about a distribution metric.

  Distribution metrics are restricted to distributions of integers only.

  This object is not thread safe, so it's not supposed to be modified
  by other than the DistributionCell that contains it.
  """
  def __init__(self, sum, count, min, max):
    # type: (int, int, int, int) -> None
    if count:
      self.sum = sum
      self.count = count
      self.min = min
      self.max = max
    else:
      self.sum = self.count = 0
      self.min = 2**63 - 1
      # Avoid Wimplicitly-unsigned-literal caused by -2**63.
      self.max = -self.min - 1

  def __eq__(self, other):
    # type: (object) -> bool
    if isinstance(other, DistributionData):
      return (
          self.sum == other.sum and self.count == other.count and
          self.min == other.min and self.max == other.max)
    else:
      return False

  def __hash__(self):
    # type: () -> int
    return hash((self.sum, self.count, self.min, self.max))

  def __repr__(self):
    # type: () -> str
    return 'DistributionData(sum={}, count={}, min={}, max={})'.format(
        self.sum, self.count, self.min, self.max)

  def get_cumulative(self):
    # type: () -> DistributionData
    return DistributionData(self.sum, self.count, self.min, self.max)

  def get_result(self) -> DistributionResult:
    return DistributionResult(self.get_cumulative())

  def combine(self, other):
    # type: (Optional[DistributionData]) -> DistributionData
    if other is None:
      return self

    return DistributionData(
        self.sum + other.sum,
        self.count + other.count,
        self.min if self.min < other.min else other.min,
        self.max if self.max > other.max else other.max)

  @staticmethod
  def singleton(value):
    # type: (int) -> DistributionData
    return DistributionData(value, 1, value, value)

  @staticmethod
  def identity_element():
    # type: () -> DistributionData
    return DistributionData(0, 0, 2**63 - 1, -2**63)


class StringSetData(object):
  """For internal use only; no backwards-compatibility guarantees.

  The data structure that holds data about a StringSet metric.

  StringSet metrics are restricted to set of strings only.

  This object is not thread safe, so it's not supposed to be modified
  by other than the StringSetCell that contains it.

  The summation of all string length for a StringSetData cannot exceed 1 MB.
  Further addition of elements are dropped.
  """

  _STRING_SET_SIZE_LIMIT = 1_000_000

  def __init__(self, string_set: Optional[Set] = None, string_size: int = 0):
    self.string_set = string_set or set()
    if not string_size:
      string_size = 0
      for s in self.string_set:
        string_size += len(s)
    self.string_size = string_size

  def __eq__(self, other: object) -> bool:
    if isinstance(other, StringSetData):
      return (
          self.string_size == other.string_size and
          self.string_set == other.string_set)
    else:
      return False

  def __hash__(self) -> int:
    return hash(self.string_set)

  def __repr__(self) -> str:
    return 'StringSetData{}:{}'.format(self.string_set, self.string_size)

  def get_cumulative(self) -> "StringSetData":
    return StringSetData(set(self.string_set), self.string_size)

  def get_result(self) -> Set[str]:
    return set(self.string_set)

  def add(self, *strings):
    """
    Add strings into this StringSetData and return the result StringSetData.
    Reuse the original StringSetData's set.
    """
    self.string_size = self.add_until_capacity(
        self.string_set, self.string_size, strings)
    return self

  def combine(self, other: "StringSetData") -> "StringSetData":
    """
    Combines this StringSetData with other, both original StringSetData are left
    intact.
    """
    if other is None:
      return self

    if not other.string_set:
      return self
    elif not self.string_set:
      return other

    combined = set(self.string_set)
    string_size = self.add_until_capacity(
        combined, self.string_size, other.string_set)
    return StringSetData(combined, string_size)

  @classmethod
  def add_until_capacity(
      cls, combined: set, current_size: int, others: Iterable[str]):
    """
    Add strings into set until reach capacity. Return the all string size of
    added set.
    """
    if current_size > cls._STRING_SET_SIZE_LIMIT:
      return current_size

    for string in others:
      if string not in combined:
        combined.add(string)
        current_size += len(string)
        if current_size > cls._STRING_SET_SIZE_LIMIT:
          _LOGGER.warning(
              "StringSet metrics reaches capacity. Further incoming elements "
              "won't be recorded. Current size: %d, last element size: %d.",
              current_size,
              len(string))
          break
    return current_size

  @staticmethod
  def singleton(value: str) -> "StringSetData":
    return StringSetData({value})

  @staticmethod
  def identity_element() -> "StringSetData":
    return StringSetData()


class _BoundedTrieNode(object):
  def __init__(self):
    # invariant: size = len(self.flattened()) = min(1, sum(size of children))
    self._size = 1
    self._children: Optional[dict[str, '_BoundedTrieNode']] = {}
    self._truncated = False

  def to_proto(self) -> metrics_pb2.BoundedTrieNode:
    return metrics_pb2.BoundedTrieNode(
        truncated=self._truncated,
        children={
            name: child.to_proto()
            for name, child in self._children.items()
        } if self._children else None)

  @staticmethod
  def from_proto(proto: metrics_pb2.BoundedTrieNode) -> '_BoundedTrieNode':
    node = _BoundedTrieNode()
    if proto.truncated:
      node._truncated = True
      node._children = None
    else:
      node._children = {
          name: _BoundedTrieNode.from_proto(child)
          for name, child in proto.children.items()
      }
      node._size = max(1, sum(child._size for child in node._children.values()))
    return node

  def size(self):
    return self._size

  def contains(self, segments):
    if self._truncated or not segments:
      return True
    head, *tail = segments
    return head in self._children and self._children[head].contains(tail)

  def add(self, segments) -> int:
    if self._truncated or not segments:
      return 0
    head, *tail = segments
    was_empty = not self._children
    child = self._children.get(head, None)  # type: ignore[union-attr]
    if child is None:
      child = self._children[head] = _BoundedTrieNode()  # type: ignore[index]
      delta = 0 if was_empty else 1
    else:
      delta = 0
    if tail:
      delta += child.add(tail)
    self._size += delta
    return delta

  def add_all(self, segments_iter):
    return sum(self.add(segments) for segments in segments_iter)

  def trim(self) -> int:
    if not self._children:
      return 0
    max_child = max(self._children.values(), key=lambda child: child._size)
    if max_child._size == 1:
      delta = 1 - self._size
      self._truncated = True
      self._children = None
    else:
      delta = max_child.trim()
    self._size += delta
    return delta

  def merge(self, other: '_BoundedTrieNode') -> int:
    if self._truncated:
      delta = 0
    elif other._truncated:
      delta = 1 - self._size
      self._truncated = True
      self._children = None
    elif not other._children:
      delta = 0
    elif not self._children:
      self._children = other._children
      delta = other._size - self._size
    else:
      delta = 0
      other_child: '_BoundedTrieNode'
      self_child: Optional['_BoundedTrieNode']
      for prefix, other_child in other._children.items():
        self_child = self._children.get(prefix, None)
        if self_child is None:
          self._children[prefix] = other_child
          delta += other_child._size
        else:
          delta += self_child.merge(other_child)
    self._size += delta
    return delta

  def flattened(self):
    if self._truncated:
      yield (True, )
    elif not self._children:
      yield (False, )
    else:
      for prefix, child in sorted(self._children.items()):
        for flattened in child.flattened():
          yield (prefix, ) + flattened

  def __hash__(self):
    return self._truncated or hash(sorted(self._children.items()))

  def __eq__(self, other):
    if isinstance(other, _BoundedTrieNode):
      return (
          self._truncated == other._truncated and
          self._children == other._children)
    else:
      return False

  def __repr__(self):
    return repr(set(''.join(str(s) for s in t) for t in self.flattened()))


class BoundedTrieData(object):
  _DEFAULT_BOUND = 100

  def __init__(self, *, root=None, singleton=None, bound=_DEFAULT_BOUND):
    self._singleton = singleton
    self._root = root
    self._bound = bound
    assert singleton is None or root is None

  def size(self):
    if self._singleton is not None:
      return 1
    elif self._root is not None:
      return self._root.size()
    else:
      return 0

  def contains(self, value):
    if self._singleton is not None:
      return tuple(value) == self._singleton
    elif self._root is not None:
      return self._root.contains(value)
    else:
      return False

  def flattened(self):
    return self.as_trie().flattened()

  def to_proto(self) -> metrics_pb2.BoundedTrie:
    return metrics_pb2.BoundedTrie(
        bound=self._bound,
        singleton=self._singleton if self._singleton else None,
        root=self._root.to_proto() if self._root else None)

  @staticmethod
  def from_proto(proto: metrics_pb2.BoundedTrie) -> 'BoundedTrieData':
    return BoundedTrieData(
        bound=proto.bound,
        singleton=tuple(proto.singleton) if proto.singleton else None,
        root=(
            _BoundedTrieNode.from_proto(proto.root)
            if proto.HasField('root') else None))

  def as_trie(self):
    if self._root is not None:
      return self._root
    else:
      root = _BoundedTrieNode()
      if self._singleton is not None:
        root.add(self._singleton)
      return root

  def __eq__(self, other: object) -> bool:
    if isinstance(other, BoundedTrieData):
      return self.as_trie() == other.as_trie()
    else:
      return False

  def __hash__(self) -> int:
    return hash(self.as_trie())

  def __repr__(self) -> str:
    return 'BoundedTrieData({})'.format(self.as_trie())

  def get_cumulative(self) -> "BoundedTrieData":
    return copy.deepcopy(self)

  def get_result(self) -> Set[tuple]:
    if self._root is None:
      if self._singleton is None:
        return set()
      else:
        return set([self._singleton + (False, )])
    else:
      return set(self._root.flattened())

  def add(self, segments):
    if self._root is None and self._singleton is None:
      self._singleton = segments
    elif self._singleton is not None and self._singleton == segments:
      # Optimize for the common case of re-adding the same value.
      return
    else:
      if self._root is None:
        self._root = self.as_trie()
        self._singleton = None
      self._root.add(segments)
      if self._root._size > self._bound:
        self._root.trim()

  def combine(self, other: "BoundedTrieData") -> "BoundedTrieData":
    if self._root is None and self._singleton is None:
      return other
    elif other._root is None and other._singleton is None:
      return self
    else:
      if self._root is None and other._root is not None:
        self, other = other, self
      combined = copy.deepcopy(self.as_trie())
      if other._root is not None:
        combined.merge(other._root)
      else:
        combined.add(other._singleton)
      self._bound = min(self._bound, other._bound)
      while combined._size > self._bound:
        combined.trim()
      return BoundedTrieData(root=combined)

  @staticmethod
  def singleton(value: str) -> "BoundedTrieData":
    s = BoundedTrieData()
    s.add(value)
    return s

  @staticmethod
  def identity_element() -> "BoundedTrieData":
    return BoundedTrieData()


class HistogramCell(MetricCell):
  """For internal use only; no backwards-compatibility guarantees.

  Tracks the current value and delta for a histogram metric.

  Each cell tracks the state of a metric independently per context per bundle.
  Therefore, each metric has a different cell in each bundle, that is later
  aggregated.

  This class is thread safe since underlying histogram object is thread safe.
  """
  def __init__(self, bucket_type):
    self._bucket_type = bucket_type
    self.data = HistogramData.identity_element(bucket_type)

  def reset(self):
    self.data = HistogramData.identity_element(self._bucket_type)

  def combine(self, other: 'HistogramCell') -> 'HistogramCell':
    result = HistogramCell(self._bucket_type)
    result.data = self.data.combine(other.data)
    return result

  def update(self, value):
    self.data.histogram.record(value)

  def get_cumulative(self) -> 'HistogramData':
    return self.data.get_cumulative()

  def to_runner_api_monitoring_info(self, name, transform_id):
    # Histogram metric is currently worker-local and internal
    # use only. This method should be implemented when runners
    # support Histogram metric reporting.
    return None


class HistogramCellFactory(MetricCellFactory):
  def __init__(self, bucket_type):
    self._bucket_type = bucket_type

  def __call__(self):
    return HistogramCell(self._bucket_type)

  def __eq__(self, other):
    if not isinstance(other, HistogramCellFactory):
      return False
    return self._bucket_type == other._bucket_type

  def __hash__(self):
    return hash(self._bucket_type)


class HistogramResult(object):
  def __init__(self, data: 'HistogramData') -> None:
    self.data = data

  def __eq__(self, other):
    if isinstance(other, HistogramResult):
      return self.data == other.data
    else:
      return False

  def __hash__(self):
    return hash(self.data)

  def __repr__(self):
    return '<HistogramResult({})>'.format(
        self.data.histogram.get_percentile_info())

  @property
  def p99(self):
    return self.data.histogram.p99()

  @property
  def p95(self):
    return self.data.histogram.p95()

  @property
  def p90(self):
    return self.data.histogram.p90()

  @property
  def histogram(self):
    return self.data.histogram


class HistogramData(object):
  """For internal use only; no backwards-compatibility guarantees.

  The data structure that holds data about a histogram metric.

  This object is not thread safe, so it's not supposed to be modified
  outside the HistogramCell.
  """
  def __init__(self, histogram):
    self.histogram = histogram

  def __eq__(self, other):
    return self.histogram == other.histogram

  def __hash__(self):
    return hash(self.histogram)

  def __repr__(self):
    return 'HistogramData({})'.format(self.histogram.get_percentile_info())

  def get_cumulative(self) -> 'HistogramData':
    return HistogramData(self.histogram)

  def combine(self, other: Optional['HistogramData']) -> 'HistogramData':
    if other is None:
      return self

    return HistogramData(self.histogram.combine(other.histogram))

  @staticmethod
  def identity_element(bucket_type) -> 'HistogramData':
    return HistogramData(Histogram(bucket_type))

  def get_result(self) -> 'HistogramResult':
    return HistogramResult(self.get_cumulative())

  def to_proto(self) -> metrics_pb2.HistogramValue:
    return self.histogram.to_runner_api()

  @classmethod
  def from_proto(cls, proto: metrics_pb2.HistogramValue):
    return cls(Histogram.from_runner_api(proto))