Skip to content

Commit

Permalink
Add base implementation for efficient cross-cell instance listing
Browse files Browse the repository at this point in the history
This adds a new set of routines that can list instances efficiently
across cells while retaining a stable sort ordering. The theory here
is that we do the query to the cells in parallel, and merge sort the
results.

This is not yet plugged into the actual list instances api call
because it lacks some important features (limit, markers) which
are added in subsequent patches.

Change-Id: I3ffed6eb7008df9d728fe9728b368fec8d3434bc
  • Loading branch information
kk7ds committed Sep 21, 2017
1 parent 0aeaa2b commit bac8f5b
Show file tree
Hide file tree
Showing 3 changed files with 382 additions and 0 deletions.
119 changes: 119 additions & 0 deletions nova/compute/instance_list.py
@@ -0,0 +1,119 @@
# Licensed 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.

import heapq

from nova import context
from nova import db


class InstanceSortContext(object):
def __init__(self, sort_keys, sort_dirs):
self._sort_keys = sort_keys
self._sort_dirs = sort_dirs

def compare_instances(self, inst1, inst2):
"""Implements cmp(inst1, inst2) for the first key that is different
Adjusts for the requested sort direction by inverting the result
as needed.
"""
for skey, sdir in zip(self._sort_keys, self._sort_dirs):
resultflag = 1 if sdir == 'desc' else -1
if inst1[skey] < inst2[skey]:
return resultflag
elif inst1[skey] > inst2[skey]:
return resultflag * -1
return 0


class InstanceWrapper(object):
"""Wrap an instance object from the database so it is sortable.
We use heapq.merge() below to do the merge sort of things from the
cell databases. That routine assumes it can use regular python
operators (> and <) on the contents. Since that won't work with
instances from the database (and depends on the sort keys/dirs),
we need this wrapper class to provide that.
Implementing __lt__ is enough for heapq.merge() to do its work.
"""
def __init__(self, sort_ctx, db_instance):
self._sort_ctx = sort_ctx
self._db_instance = db_instance

def __lt__(self, other):
r = self._sort_ctx.compare_instances(self._db_instance,
other._db_instance)
# cmp(x, y) returns -1 if x < y
return r == -1


def get_instances_sorted(ctx, filters, limit, marker, columns_to_join,
sort_keys, sort_dirs):
"""Get a cross-cell list of instances matching filters.
This iterates cells in parallel generating a unified and sorted
list of instances as efficiently as possible. It takes care to
iterate the list as infrequently as possible. We wrap the results
in InstanceWrapper objects so that they are sortable by
heapq.merge(), which requires that the '<' operator just works. We
encapsulate our sorting requirements into an InstanceSortContext
which we pass to all of the wrappers so they behave the way we
want.
This function is a generator of instances from the database like what you
would get from instance_get_all_by_filters_sort() in the DB API.
FIXME: Make limit work
FIXME: Make marker work
"""

if not sort_keys:
# This is the default from the process_sort_params() method in
# the DB API. It doesn't really matter, as this only comes into
# play if the user didn't ask for a specific ordering, but we
# use the same scheme for consistency.
sort_keys = ['created_at', 'id']
sort_dirs = ['asc', 'asc']

sort_ctx = InstanceSortContext(sort_keys, sort_dirs)

def do_query(ctx):
"""Generate InstanceWrapper(Instance) objects from a cell.
We do this inside the thread (created by
scatter_gather_all_cells()) so that we return wrappers and
avoid having to iterate the combined result list in the caller
again. This is run against each cell by the scatter_gather
routine.
"""

return (InstanceWrapper(sort_ctx, inst) for inst in
db.instance_get_all_by_filters_sort(
ctx, filters,
limit=limit, marker=marker,
columns_to_join=columns_to_join,
sort_keys=sort_keys,
sort_dirs=sort_dirs))

# FIXME(danms): If we raise or timeout on a cell we need to handle
# that here gracefully. The below routine will provide sentinels
# to indicate that, which will crash the merge below, but we don't
# handle this anywhere yet anyway.
results = context.scatter_gather_all_cells(ctx, do_query)

# Generate results from heapq so we can return the inner
# instance instead of the wrapper. This is basically free
# as it works as our caller iterates the results.
for i in heapq.merge(*results.values()):
yield i._db_instance
113 changes: 113 additions & 0 deletions nova/tests/functional/compute/test_instance_list.py
@@ -0,0 +1,113 @@
# Licensed 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.

import datetime

from nova.compute import instance_list
from nova import context
from nova import objects
from nova import test


class InstanceListTestCase(test.TestCase):
NUMBER_OF_CELLS = 3

def setUp(self):
super(InstanceListTestCase, self).setUp()

self.context = context.RequestContext('fake', 'fake')
self.num_instances = 3
self.instances = []

dt = datetime.datetime(1985, 10, 25, 1, 21, 0)
spread = datetime.timedelta(minutes=10)

cells = objects.CellMappingList.get_all(self.context)
# Create three instances in each of the real cells. Leave the
# first cell empty to make sure we don't break with an empty
# one.
for cell in cells[1:]:
for i in range(0, self.num_instances):
with context.target_cell(self.context, cell) as cctx:
inst = objects.Instance(
context=cctx,
project_id=self.context.project_id,
user_id=self.context.user_id,
launched_at=dt,
instance_type_id=i,
hostname='%s-inst%i' % (cell.name, i))
inst.create()
self.instances.append(inst)
im = objects.InstanceMapping(context=self.context,
project_id=inst.project_id,
user_id=inst.user_id,
instance_uuid=inst.uuid,
cell_mapping=cell)
im.create()
dt += spread

def test_get_sorted(self):
filters = {}
limit = None
marker = None
columns = []
sort_keys = ['uuid']
sort_dirs = ['asc']
insts = instance_list.get_instances_sorted(self.context, filters,
limit, marker, columns,
sort_keys, sort_dirs)
uuids = [inst['uuid'] for inst in insts]
self.assertEqual(sorted(uuids), uuids)
self.assertEqual(len(self.instances), len(uuids))

def test_get_sorted_descending(self):
filters = {}
limit = None
marker = None
columns = []
sort_keys = ['uuid']
sort_dirs = ['desc']
insts = instance_list.get_instances_sorted(self.context, filters,
limit, marker, columns,
sort_keys, sort_dirs)
uuids = [inst['uuid'] for inst in insts]
self.assertEqual(list(reversed(sorted(uuids))), uuids)
self.assertEqual(len(self.instances), len(uuids))

def test_get_sorted_with_filter(self):
filters = {'instance_type_id': 1}
limit = None
marker = None
columns = []
sort_keys = ['uuid']
sort_dirs = ['asc']
insts = instance_list.get_instances_sorted(self.context, filters,
limit, marker, columns,
sort_keys, sort_dirs)
uuids = [inst['uuid'] for inst in insts]
expected = [inst['uuid'] for inst in self.instances
if inst['instance_type_id'] == 1]
self.assertEqual(list(sorted(expected)), uuids)

def test_get_sorted_by_defaults(self):
filters = {}
limit = None
marker = None
columns = []
sort_keys = None
sort_dirs = None
insts = instance_list.get_instances_sorted(self.context, filters,
limit, marker, columns,
sort_keys, sort_dirs)
uuids = set([inst['uuid'] for inst in insts])
expected = set([inst['uuid'] for inst in self.instances])
self.assertEqual(expected, uuids)
150 changes: 150 additions & 0 deletions nova/tests/unit/compute/test_instance_list.py
@@ -0,0 +1,150 @@
# Licensed 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.

import datetime
import mock

from nova.compute import instance_list
from nova import objects
from nova import test
from nova.tests import fixtures
from nova.tests import uuidsentinel as uuids


class TestUtils(test.NoDBTestCase):
def test_compare_simple(self):
dt1 = datetime.datetime(2015, 11, 5, 20, 30, 00)
dt2 = datetime.datetime(1955, 10, 25, 1, 21, 00)

inst1 = {'key0': 'foo', 'key1': 'd', 'key2': 456, 'key4': dt1}
inst2 = {'key0': 'foo', 'key1': 's', 'key2': 123, 'key4': dt2}

# Equal key0, inst == inst2
ctx = instance_list.InstanceSortContext(['key0'], ['asc'])
self.assertEqual(0, ctx.compare_instances(inst1, inst2))

# Equal key0, inst == inst2 (direction should not matter)
ctx = instance_list.InstanceSortContext(['key0'], ['desc'])
self.assertEqual(0, ctx.compare_instances(inst1, inst2))

# Ascending by key1, inst1 < inst2
ctx = instance_list.InstanceSortContext(['key1'], ['asc'])
self.assertEqual(-1, ctx.compare_instances(inst1, inst2))

# Descending by key1, inst2 < inst1
ctx = instance_list.InstanceSortContext(['key1'], ['desc'])
self.assertEqual(1, ctx.compare_instances(inst1, inst2))

# Ascending by key2, inst2 < inst1
ctx = instance_list.InstanceSortContext(['key2'], ['asc'])
self.assertEqual(1, ctx.compare_instances(inst1, inst2))

# Descending by key2, inst1 < inst2
ctx = instance_list.InstanceSortContext(['key2'], ['desc'])
self.assertEqual(-1, ctx.compare_instances(inst1, inst2))

# Ascending by key4, inst1 > inst2
ctx = instance_list.InstanceSortContext(['key4'], ['asc'])
self.assertEqual(1, ctx.compare_instances(inst1, inst2))

# Descending by key4, inst1 < inst2
ctx = instance_list.InstanceSortContext(['key4'], ['desc'])
self.assertEqual(-1, ctx.compare_instances(inst1, inst2))

def test_compare_multiple(self):
# key0 should not affect ordering, but key1 should

inst1 = {'key0': 'foo', 'key1': 'd', 'key2': 456}
inst2 = {'key0': 'foo', 'key1': 's', 'key2': 123}

# Should be equivalent to ascending by key1
ctx = instance_list.InstanceSortContext(['key0', 'key1'],
['asc', 'asc'])
self.assertEqual(-1, ctx.compare_instances(inst1, inst2))

# Should be equivalent to descending by key1
ctx = instance_list.InstanceSortContext(['key0', 'key1'],
['asc', 'desc'])
self.assertEqual(1, ctx.compare_instances(inst1, inst2))

def test_wrapper(self):
inst1 = {'key0': 'foo', 'key1': 'd', 'key2': 456}
inst2 = {'key0': 'foo', 'key1': 's', 'key2': 123}

# Should sort by key1
ctx = instance_list.InstanceSortContext(['key0', 'key1'],
['asc', 'asc'])
iw1 = instance_list.InstanceWrapper(ctx, inst1)
iw2 = instance_list.InstanceWrapper(ctx, inst2)
# Check this both ways to make sure we're comparing against -1
# and not just nonzero return from cmp()
self.assertTrue(iw1 < iw2)
self.assertFalse(iw2 < iw1)

# Should sort reverse by key1
ctx = instance_list.InstanceSortContext(['key0', 'key1'],
['asc', 'desc'])
iw1 = instance_list.InstanceWrapper(ctx, inst1)
iw2 = instance_list.InstanceWrapper(ctx, inst2)
# Check this both ways to make sure we're comparing against -1
# and not just nonzero return from cmp()
self.assertTrue(iw1 > iw2)
self.assertFalse(iw2 > iw1)


class TestInstanceList(test.NoDBTestCase):
def setUp(self):
super(TestInstanceList, self).setUp()

cells = [objects.CellMapping(uuid=getattr(uuids, 'cell%i' % i),
name='cell%i' % i,
transport_url='fake:///',
database_connection='fake://')
for i in range(0, 3)]

insts = {}
for cell in cells:
insts[cell.uuid] = list([
dict(
uuid=getattr(uuids, '%s-inst%i' % (cell.name, i)),
hostname='%s-inst%i' % (cell.name, i))
for i in range(0, 3)])

self.cells = cells
self.insts = insts
self.context = mock.sentinel.context
self.useFixture(fixtures.SpawnIsSynchronousFixture())

@mock.patch('nova.db.instance_get_all_by_filters_sort')
@mock.patch('nova.objects.CellMappingList.get_all')
def test_get_instances_sorted(self, mock_cells, mock_inst):
mock_cells.return_value = self.cells
insts_by_cell = self.insts.values()

mock_inst.side_effect = insts_by_cell
insts = instance_list.get_instances_sorted(self.context, {},
None, None,
[], ['hostname'], ['asc'])
insts_one = [inst['hostname'] for inst in insts]

# Reverse the order that we get things from the cells so we can
# make sure that the result is still sorted the same way
insts_by_cell = list(reversed(list(insts_by_cell)))
mock_inst.reset_mock()
mock_inst.side_effect = insts_by_cell

insts = instance_list.get_instances_sorted(self.context, {},
None, None,
[], ['hostname'], ['asc'])
insts_two = [inst['hostname'] for inst in insts]

self.assertEqual(insts_one, insts_two)

0 comments on commit bac8f5b

Please sign in to comment.