Split schedulers into their own submodule.
This commit is contained in:
4
setup.py
4
setup.py
@@ -1,4 +1,4 @@
|
||||
from setuptools import setup
|
||||
from setuptools import setup, find_packages
|
||||
|
||||
setup(
|
||||
name="pytest-xdist",
|
||||
@@ -11,7 +11,7 @@ setup(
|
||||
author_email='pytest-dev@python.org,holger@merlinux.eu',
|
||||
url='https://github.com/pytest-dev/pytest-xdist',
|
||||
platforms=['linux', 'osx', 'win32'],
|
||||
packages=['xdist'],
|
||||
packages=find_packages(exclude=['testing', 'example']),
|
||||
entry_points={
|
||||
'pytest11': [
|
||||
'xdist = xdist.plugin',
|
||||
|
||||
@@ -1,6 +1,10 @@
|
||||
from xdist.dsession import (
|
||||
DSession, LoadScheduling, EachScheduling, report_collection_diff,
|
||||
from xdist.dsession import DSession
|
||||
from xdist.report import report_collection_diff
|
||||
from xdist.scheduler import (
|
||||
EachScheduling,
|
||||
LoadScheduling,
|
||||
)
|
||||
|
||||
import py
|
||||
import pytest
|
||||
import execnet
|
||||
|
||||
@@ -1,445 +1,16 @@
|
||||
import difflib
|
||||
import itertools
|
||||
from _pytest.runner import CollectReport
|
||||
|
||||
import pytest
|
||||
import py
|
||||
from xdist.slavemanage import NodeManager, parse_spec_config
|
||||
import pytest
|
||||
|
||||
from xdist.slavemanage import NodeManager
|
||||
from xdist.scheduler import (
|
||||
EachScheduling,
|
||||
LoadScheduling,
|
||||
)
|
||||
|
||||
|
||||
queue = py.builtin._tryimport('queue', 'Queue')
|
||||
|
||||
|
||||
class EachScheduling:
|
||||
"""Implement scheduling of test items on all nodes
|
||||
|
||||
If a node gets added after the test run is started then it is
|
||||
assumed to replace a node which got removed before it finished
|
||||
its collection. In this case it will only be used if a node
|
||||
with the same spec got removed earlier.
|
||||
|
||||
Any nodes added after the run is started will only get items
|
||||
assigned if a node with a matching spec was removed before it
|
||||
finished all its pending items. The new node will then be
|
||||
assigned the remaining items from the removed node.
|
||||
"""
|
||||
|
||||
def __init__(self, config, log=None):
|
||||
self.config = config
|
||||
self.numnodes = len(parse_spec_config(config))
|
||||
self.node2collection = {}
|
||||
self.node2pending = {}
|
||||
self._started = []
|
||||
self._removed2pending = {}
|
||||
if log is None:
|
||||
self.log = py.log.Producer("eachsched")
|
||||
else:
|
||||
self.log = log.eachsched
|
||||
self.collection_is_completed = False
|
||||
|
||||
@property
|
||||
def nodes(self):
|
||||
"""A list of all nodes in the scheduler."""
|
||||
return list(self.node2pending.keys())
|
||||
|
||||
@property
|
||||
def tests_finished(self):
|
||||
if not self.collection_is_completed:
|
||||
return False
|
||||
if self._removed2pending:
|
||||
return False
|
||||
for pending in self.node2pending.values():
|
||||
if len(pending) >= 2:
|
||||
return False
|
||||
return True
|
||||
|
||||
@property
|
||||
def has_pending(self):
|
||||
"""Return True if there are pending test items
|
||||
|
||||
This indicates that collection has finished and nodes are
|
||||
still processing test items, so this can be thought of as
|
||||
"the scheduler is active".
|
||||
"""
|
||||
for pending in self.node2pending.values():
|
||||
if pending:
|
||||
return True
|
||||
return False
|
||||
|
||||
def add_node(self, node):
|
||||
assert node not in self.node2pending
|
||||
self.node2pending[node] = []
|
||||
|
||||
def add_node_collection(self, node, collection):
|
||||
"""Add the collected test items from a node
|
||||
|
||||
Collection is complete once all nodes have submitted their
|
||||
collection. In this case its pending list is set to an empty
|
||||
list. When the collection is already completed this
|
||||
submission is from a node which was restarted to replace a
|
||||
dead node. In this case we already assign the pending items
|
||||
here. In either case ``.schedule()`` will instruct the
|
||||
node to start running the required tests.
|
||||
"""
|
||||
assert node in self.node2pending
|
||||
if not self.collection_is_completed:
|
||||
self.node2collection[node] = list(collection)
|
||||
self.node2pending[node] = []
|
||||
if len(self.node2collection) >= self.numnodes:
|
||||
self.collection_is_completed = True
|
||||
elif self._removed2pending:
|
||||
for deadnode in self._removed2pending:
|
||||
if deadnode.gateway.spec == node.gateway.spec:
|
||||
dead_collection = self.node2collection[deadnode]
|
||||
if collection != dead_collection:
|
||||
msg = report_collection_diff(dead_collection,
|
||||
collection,
|
||||
deadnode.gateway.id,
|
||||
node.gateway.id)
|
||||
self.log(msg)
|
||||
return
|
||||
pending = self._removed2pending.pop(deadnode)
|
||||
self.node2pending[node] = pending
|
||||
break
|
||||
|
||||
def mark_test_complete(self, node, item_index, duration=0):
|
||||
self.node2pending[node].remove(item_index)
|
||||
|
||||
def remove_node(self, node):
|
||||
# KeyError if we didn't get an add_node() yet
|
||||
pending = self.node2pending.pop(node)
|
||||
if not pending:
|
||||
return
|
||||
crashitem = self.node2collection[node][pending.pop(0)]
|
||||
if pending:
|
||||
self._removed2pending[node] = pending
|
||||
return crashitem
|
||||
|
||||
def schedule(self):
|
||||
"""Schedule the test items on the nodes
|
||||
|
||||
If the node's pending list is empty it is a new node which
|
||||
needs to run all the tests. If the pending list is already
|
||||
populated (by ``.add_node_collection()``) then it replaces a
|
||||
dead node and we only need to run those tests.
|
||||
"""
|
||||
assert self.collection_is_completed
|
||||
for node, pending in self.node2pending.items():
|
||||
if node in self._started:
|
||||
continue
|
||||
if not pending:
|
||||
pending[:] = range(len(self.node2collection[node]))
|
||||
node.send_runtest_all()
|
||||
else:
|
||||
node.send_runtest_some(pending)
|
||||
self._started.append(node)
|
||||
|
||||
|
||||
class LoadScheduling:
|
||||
"""Implement load scheduling across nodes.
|
||||
|
||||
This distributes the tests collected across all nodes so each test
|
||||
is run just once. All nodes collect and submit the test suite and
|
||||
when all collections are received it is verified they are
|
||||
identical collections. Then the collection gets divided up in
|
||||
chunks and chunks get submitted to nodes. Whenever a node finishes
|
||||
an item, it calls ``.mark_test_complete()`` which will trigger the
|
||||
scheduler to assign more tests if the number of pending tests for
|
||||
the node falls below a low-watermark.
|
||||
|
||||
When created, ``numnodes`` defines how many nodes are expected to
|
||||
submit a collection. This is used to know when all nodes have
|
||||
finished collection or how large the chunks need to be created.
|
||||
|
||||
Attributes:
|
||||
|
||||
:numnodes: The expected number of nodes taking part. The actual
|
||||
number of nodes will vary during the scheduler's lifetime as
|
||||
nodes are added by the DSession as they are brought up and
|
||||
removed either because of a dead node or normal shutdown. This
|
||||
number is primarily used to know when the initial collection is
|
||||
completed.
|
||||
|
||||
:node2collection: Map of nodes and their test collection. All
|
||||
collections should always be identical.
|
||||
|
||||
:node2pending: Map of nodes and the indices of their pending
|
||||
tests. The indices are an index into ``.pending`` (which is
|
||||
identical to their own collection stored in
|
||||
``.node2collection``).
|
||||
|
||||
:collection: The one collection once it is validated to be
|
||||
identical between all the nodes. It is initialised to None
|
||||
until ``.schedule()`` is called.
|
||||
|
||||
:pending: List of indices of globally pending tests. These are
|
||||
tests which have not yet been allocated to a chunk for a node
|
||||
to process.
|
||||
|
||||
:log: A py.log.Producer instance.
|
||||
|
||||
:config: Config object, used for handling hooks.
|
||||
"""
|
||||
|
||||
def __init__(self, config, log=None):
|
||||
self.numnodes = len(parse_spec_config(config))
|
||||
self.node2collection = {}
|
||||
self.node2pending = {}
|
||||
self.pending = []
|
||||
self.collection = None
|
||||
if log is None:
|
||||
self.log = py.log.Producer("loadsched")
|
||||
else:
|
||||
self.log = log.loadsched
|
||||
self.config = config
|
||||
|
||||
@property
|
||||
def nodes(self):
|
||||
"""A list of all nodes in the scheduler."""
|
||||
return list(self.node2pending.keys())
|
||||
|
||||
@property
|
||||
def collection_is_completed(self):
|
||||
"""Boolean indication initial test collection is complete.
|
||||
|
||||
This is a boolean indicating all initial participating nodes
|
||||
have finished collection. The required number of initial
|
||||
nodes is defined by ``.numnodes``.
|
||||
"""
|
||||
return len(self.node2collection) >= self.numnodes
|
||||
|
||||
@property
|
||||
def tests_finished(self):
|
||||
"""Return True if all tests have been executed by the nodes."""
|
||||
if not self.collection_is_completed:
|
||||
return False
|
||||
if self.pending:
|
||||
return False
|
||||
for pending in self.node2pending.values():
|
||||
if len(pending) >= 2:
|
||||
return False
|
||||
return True
|
||||
|
||||
@property
|
||||
def has_pending(self):
|
||||
"""Return True if there are pending test items
|
||||
|
||||
This indicates that collection has finished and nodes are
|
||||
still processing test items, so this can be thought of as
|
||||
"the scheduler is active".
|
||||
"""
|
||||
if self.pending:
|
||||
return True
|
||||
for pending in self.node2pending.values():
|
||||
if pending:
|
||||
return True
|
||||
return False
|
||||
|
||||
def add_node(self, node):
|
||||
"""Add a new node to the scheduler.
|
||||
|
||||
From now on the node will be allocated chunks of tests to
|
||||
execute.
|
||||
|
||||
Called by the ``DSession.slave_slaveready`` hook when it
|
||||
successfully bootstraps a new node.
|
||||
"""
|
||||
assert node not in self.node2pending
|
||||
self.node2pending[node] = []
|
||||
|
||||
def add_node_collection(self, node, collection):
|
||||
"""Add the collected test items from a node
|
||||
|
||||
The collection is stored in the ``.node2collection`` map.
|
||||
Called by the ``DSession.slave_collectionfinish`` hook.
|
||||
"""
|
||||
assert node in self.node2pending
|
||||
if self.collection_is_completed:
|
||||
# A new node has been added later, perhaps an original one died.
|
||||
# .schedule() should have
|
||||
# been called by now
|
||||
assert self.collection
|
||||
if collection != self.collection:
|
||||
other_node = next(iter(self.node2collection.keys()))
|
||||
msg = report_collection_diff(self.collection,
|
||||
collection,
|
||||
other_node.gateway.id,
|
||||
node.gateway.id)
|
||||
self.log(msg)
|
||||
return
|
||||
self.node2collection[node] = list(collection)
|
||||
|
||||
def mark_test_complete(self, node, item_index, duration=0):
|
||||
"""Mark test item as completed by node
|
||||
|
||||
The duration it took to execute the item is used as a hint to
|
||||
the scheduler.
|
||||
|
||||
This is called by the ``DSession.slave_testreport`` hook.
|
||||
"""
|
||||
self.node2pending[node].remove(item_index)
|
||||
self.check_schedule(node, duration=duration)
|
||||
|
||||
def check_schedule(self, node, duration=0):
|
||||
"""Maybe schedule new items on the node
|
||||
|
||||
If there are any globally pending nodes left then this will
|
||||
check if the given node should be given any more tests. The
|
||||
``duration`` of the last test is optionally used as a
|
||||
heuristic to influence how many tests the node is assigned.
|
||||
"""
|
||||
if node.shutting_down:
|
||||
return
|
||||
|
||||
if self.pending:
|
||||
# how many nodes do we have?
|
||||
num_nodes = len(self.node2pending)
|
||||
# if our node goes below a heuristic minimum, fill it out to
|
||||
# heuristic maximum
|
||||
items_per_node_min = max(2, len(self.pending) // num_nodes // 4)
|
||||
items_per_node_max = max(2, len(self.pending) // num_nodes // 2)
|
||||
node_pending = self.node2pending[node]
|
||||
if len(node_pending) < items_per_node_min:
|
||||
if duration >= 0.1 and len(node_pending) >= 2:
|
||||
# seems the node is doing long-running tests
|
||||
# and has enough items to continue
|
||||
# so let's rather wait with sending new items
|
||||
return
|
||||
num_send = items_per_node_max - len(node_pending)
|
||||
self._send_tests(node, num_send)
|
||||
self.log("num items waiting for node:", len(self.pending))
|
||||
|
||||
def remove_node(self, node):
|
||||
"""Remove a node from the scheduler
|
||||
|
||||
This should be called either when the node crashed or at
|
||||
shutdown time. In the former case any pending items assigned
|
||||
to the node will be re-scheduled. Called by the
|
||||
``DSession.slave_slavefinished`` and
|
||||
``DSession.slave_errordown`` hooks.
|
||||
|
||||
Return the item which was being executing while the node
|
||||
crashed or None if the node has no more pending items.
|
||||
|
||||
"""
|
||||
pending = self.node2pending.pop(node)
|
||||
if not pending:
|
||||
return
|
||||
|
||||
# The node crashed, reassing pending items
|
||||
crashitem = self.collection[pending.pop(0)]
|
||||
self.pending.extend(pending)
|
||||
for node in self.node2pending:
|
||||
self.check_schedule(node)
|
||||
return crashitem
|
||||
|
||||
def schedule(self):
|
||||
"""Initiate distribution of the test collection
|
||||
|
||||
Initiate scheduling of the items across the nodes. If this
|
||||
gets called again later it behaves the same as calling
|
||||
``.check_schedule()`` on all nodes so that newly added nodes
|
||||
will start to be used.
|
||||
|
||||
This is called by the ``DSession.slave_collectionfinish`` hook
|
||||
if ``.collection_is_completed`` is True.
|
||||
"""
|
||||
assert self.collection_is_completed
|
||||
|
||||
# Initial distribution already happened, reschedule on all nodes
|
||||
if self.collection is not None:
|
||||
for node in self.nodes:
|
||||
self.check_schedule(node)
|
||||
return
|
||||
|
||||
# XXX allow nodes to have different collections
|
||||
if not self._check_nodes_have_same_collection():
|
||||
self.log('**Different tests collected, aborting run**')
|
||||
return
|
||||
|
||||
# Collections are identical, create the index of pending items.
|
||||
self.collection = list(self.node2collection.values())[0]
|
||||
self.pending[:] = range(len(self.collection))
|
||||
if not self.collection:
|
||||
return
|
||||
|
||||
# Send a batch of tests to run. If we don't have at least two
|
||||
# tests per node, we have to send them all so that we can send
|
||||
# shutdown signals and get all nodes working.
|
||||
initial_batch = max(len(self.pending) // 4,
|
||||
2 * len(self.nodes))
|
||||
|
||||
# distribute tests round-robin up to the batch size
|
||||
# (or until we run out)
|
||||
nodes = itertools.cycle(self.nodes)
|
||||
for i in range(initial_batch):
|
||||
self._send_tests(next(nodes), 1)
|
||||
|
||||
if not self.pending:
|
||||
# initial distribution sent all tests, start node shutdown
|
||||
for node in self.nodes:
|
||||
node.shutdown()
|
||||
|
||||
def _send_tests(self, node, num):
|
||||
tests_per_node = self.pending[:num]
|
||||
if tests_per_node:
|
||||
del self.pending[:num]
|
||||
self.node2pending[node].extend(tests_per_node)
|
||||
node.send_runtest_some(tests_per_node)
|
||||
|
||||
def _check_nodes_have_same_collection(self):
|
||||
"""Return True if all nodes have collected the same items.
|
||||
|
||||
If collections differ, this method returns False while logging
|
||||
the collection differences and posting collection errors to
|
||||
pytest_collectreport hook.
|
||||
"""
|
||||
node_collection_items = list(self.node2collection.items())
|
||||
first_node, col = node_collection_items[0]
|
||||
same_collection = True
|
||||
for node, collection in node_collection_items[1:]:
|
||||
msg = report_collection_diff(
|
||||
col,
|
||||
collection,
|
||||
first_node.gateway.id,
|
||||
node.gateway.id,
|
||||
)
|
||||
if msg:
|
||||
same_collection = False
|
||||
self.log(msg)
|
||||
if self.config is not None:
|
||||
rep = CollectReport(
|
||||
node.gateway.id, 'failed',
|
||||
longrepr=msg, result=[])
|
||||
self.config.hook.pytest_collectreport(report=rep)
|
||||
|
||||
return same_collection
|
||||
|
||||
|
||||
def report_collection_diff(from_collection, to_collection, from_id, to_id):
|
||||
"""Report the collected test difference between two nodes.
|
||||
|
||||
:returns: detailed message describing the difference between the given
|
||||
collections, or None if they are equal.
|
||||
"""
|
||||
if from_collection == to_collection:
|
||||
return None
|
||||
|
||||
diff = difflib.unified_diff(
|
||||
from_collection,
|
||||
to_collection,
|
||||
fromfile=from_id,
|
||||
tofile=to_id,
|
||||
)
|
||||
error_message = py.builtin._totext(
|
||||
'Different tests were collected between {from_id} and {to_id}. '
|
||||
'The difference is:\n'
|
||||
'{diff}'
|
||||
).format(from_id=from_id, to_id=to_id, diff='\n'.join(diff))
|
||||
msg = "\n".join([x.rstrip() for x in error_message.split("\n")])
|
||||
return msg
|
||||
|
||||
|
||||
class Interrupted(KeyboardInterrupt):
|
||||
""" signals an immediate interruption. """
|
||||
|
||||
@@ -457,6 +28,7 @@ class DSession:
|
||||
automatically starts collecting tests. Once tests are collected
|
||||
it will wait for instructions.
|
||||
"""
|
||||
|
||||
def __init__(self, config):
|
||||
self.config = config
|
||||
self.log = py.log.Producer("dsession")
|
||||
@@ -521,11 +93,12 @@ class DSession:
|
||||
|
||||
@pytest.mark.trylast
|
||||
def pytest_xdist_make_scheduler(self, config, log):
|
||||
dist = config.getvalue("dist")
|
||||
if dist == "load":
|
||||
return LoadScheduling(config, log)
|
||||
elif dist == "each":
|
||||
return EachScheduling(config, log)
|
||||
dist = config.getvalue('dist')
|
||||
schedulers = {
|
||||
'each': EachScheduling,
|
||||
'load': LoadScheduling,
|
||||
}
|
||||
return schedulers[dist](config, log)
|
||||
|
||||
def pytest_runtestloop(self):
|
||||
self.sched = self.config.hook.pytest_xdist_make_scheduler(
|
||||
@@ -746,7 +319,7 @@ class TerminalDistReporter:
|
||||
return " / ".join(parts)
|
||||
|
||||
def rewrite(self, line, newline=False):
|
||||
pline = line + " " * max(self._lastlen-len(line), 0)
|
||||
pline = line + " " * max(self._lastlen - len(line), 0)
|
||||
if newline:
|
||||
self._lastlen = 0
|
||||
pline += "\n"
|
||||
|
||||
@@ -31,11 +31,12 @@ def pytest_addoption(parser):
|
||||
"when crashed (set to zero to disable this feature)")
|
||||
group._addoption(
|
||||
'--dist', metavar="distmode",
|
||||
action="store", choices=['load', 'each', 'no'],
|
||||
action="store", choices=['each', 'load', 'no'],
|
||||
dest="dist", default="no",
|
||||
help=("set mode for distributing tests to exec environments.\n\n"
|
||||
"each: send each test to each available environment.\n\n"
|
||||
"load: send each test to available environment.\n\n"
|
||||
"each: send each test to all available environments.\n\n"
|
||||
"load: load balance by sending any pending test to any"
|
||||
" available environment.\n\n"
|
||||
"(default) no: run tests inprocess, don't distribute."))
|
||||
group._addoption(
|
||||
'--tx', dest="tx", action="append", default=[],
|
||||
|
||||
26
xdist/report.py
Normal file
26
xdist/report.py
Normal file
@@ -0,0 +1,26 @@
|
||||
import py
|
||||
from difflib import unified_diff
|
||||
|
||||
|
||||
def report_collection_diff(from_collection, to_collection, from_id, to_id):
|
||||
"""Report the collected test difference between two nodes.
|
||||
|
||||
:returns: detailed message describing the difference between the given
|
||||
collections, or None if they are equal.
|
||||
"""
|
||||
if from_collection == to_collection:
|
||||
return None
|
||||
|
||||
diff = unified_diff(
|
||||
from_collection,
|
||||
to_collection,
|
||||
fromfile=from_id,
|
||||
tofile=to_id,
|
||||
)
|
||||
error_message = py.builtin._totext(
|
||||
'Different tests were collected between {from_id} and {to_id}. '
|
||||
'The difference is:\n'
|
||||
'{diff}'
|
||||
).format(from_id=from_id, to_id=to_id, diff='\n'.join(diff))
|
||||
msg = "\n".join([x.rstrip() for x in error_message.split("\n")])
|
||||
return msg
|
||||
2
xdist/scheduler/__init__.py
Normal file
2
xdist/scheduler/__init__.py
Normal file
@@ -0,0 +1,2 @@
|
||||
from xdist.scheduler.each import EachScheduling # noqa
|
||||
from xdist.scheduler.load import LoadScheduling # noqa
|
||||
129
xdist/scheduler/each.py
Normal file
129
xdist/scheduler/each.py
Normal file
@@ -0,0 +1,129 @@
|
||||
from py.log import Producer
|
||||
|
||||
from xdist.slavemanage import parse_spec_config
|
||||
from xdist.report import report_collection_diff
|
||||
|
||||
|
||||
class EachScheduling:
|
||||
"""Implement scheduling of test items on all nodes
|
||||
|
||||
If a node gets added after the test run is started then it is
|
||||
assumed to replace a node which got removed before it finished
|
||||
its collection. In this case it will only be used if a node
|
||||
with the same spec got removed earlier.
|
||||
|
||||
Any nodes added after the run is started will only get items
|
||||
assigned if a node with a matching spec was removed before it
|
||||
finished all its pending items. The new node will then be
|
||||
assigned the remaining items from the removed node.
|
||||
"""
|
||||
|
||||
def __init__(self, config, log=None):
|
||||
self.config = config
|
||||
self.numnodes = len(parse_spec_config(config))
|
||||
self.node2collection = {}
|
||||
self.node2pending = {}
|
||||
self._started = []
|
||||
self._removed2pending = {}
|
||||
if log is None:
|
||||
self.log = Producer("eachsched")
|
||||
else:
|
||||
self.log = log.eachsched
|
||||
self.collection_is_completed = False
|
||||
|
||||
@property
|
||||
def nodes(self):
|
||||
"""A list of all nodes in the scheduler."""
|
||||
return list(self.node2pending.keys())
|
||||
|
||||
@property
|
||||
def tests_finished(self):
|
||||
if not self.collection_is_completed:
|
||||
return False
|
||||
if self._removed2pending:
|
||||
return False
|
||||
for pending in self.node2pending.values():
|
||||
if len(pending) >= 2:
|
||||
return False
|
||||
return True
|
||||
|
||||
@property
|
||||
def has_pending(self):
|
||||
"""Return True if there are pending test items
|
||||
|
||||
This indicates that collection has finished and nodes are
|
||||
still processing test items, so this can be thought of as
|
||||
"the scheduler is active".
|
||||
"""
|
||||
for pending in self.node2pending.values():
|
||||
if pending:
|
||||
return True
|
||||
return False
|
||||
|
||||
def add_node(self, node):
|
||||
assert node not in self.node2pending
|
||||
self.node2pending[node] = []
|
||||
|
||||
def add_node_collection(self, node, collection):
|
||||
"""Add the collected test items from a node
|
||||
|
||||
Collection is complete once all nodes have submitted their
|
||||
collection. In this case its pending list is set to an empty
|
||||
list. When the collection is already completed this
|
||||
submission is from a node which was restarted to replace a
|
||||
dead node. In this case we already assign the pending items
|
||||
here. In either case ``.schedule()`` will instruct the
|
||||
node to start running the required tests.
|
||||
"""
|
||||
assert node in self.node2pending
|
||||
if not self.collection_is_completed:
|
||||
self.node2collection[node] = list(collection)
|
||||
self.node2pending[node] = []
|
||||
if len(self.node2collection) >= self.numnodes:
|
||||
self.collection_is_completed = True
|
||||
elif self._removed2pending:
|
||||
for deadnode in self._removed2pending:
|
||||
if deadnode.gateway.spec == node.gateway.spec:
|
||||
dead_collection = self.node2collection[deadnode]
|
||||
if collection != dead_collection:
|
||||
msg = report_collection_diff(dead_collection,
|
||||
collection,
|
||||
deadnode.gateway.id,
|
||||
node.gateway.id)
|
||||
self.log(msg)
|
||||
return
|
||||
pending = self._removed2pending.pop(deadnode)
|
||||
self.node2pending[node] = pending
|
||||
break
|
||||
|
||||
def mark_test_complete(self, node, item_index, duration=0):
|
||||
self.node2pending[node].remove(item_index)
|
||||
|
||||
def remove_node(self, node):
|
||||
# KeyError if we didn't get an add_node() yet
|
||||
pending = self.node2pending.pop(node)
|
||||
if not pending:
|
||||
return
|
||||
crashitem = self.node2collection[node][pending.pop(0)]
|
||||
if pending:
|
||||
self._removed2pending[node] = pending
|
||||
return crashitem
|
||||
|
||||
def schedule(self):
|
||||
"""Schedule the test items on the nodes
|
||||
|
||||
If the node's pending list is empty it is a new node which
|
||||
needs to run all the tests. If the pending list is already
|
||||
populated (by ``.add_node_collection()``) then it replaces a
|
||||
dead node and we only need to run those tests.
|
||||
"""
|
||||
assert self.collection_is_completed
|
||||
for node, pending in self.node2pending.items():
|
||||
if node in self._started:
|
||||
continue
|
||||
if not pending:
|
||||
pending[:] = range(len(self.node2collection[node]))
|
||||
node.send_runtest_all()
|
||||
else:
|
||||
node.send_runtest_some(pending)
|
||||
self._started.append(node)
|
||||
288
xdist/scheduler/load.py
Normal file
288
xdist/scheduler/load.py
Normal file
@@ -0,0 +1,288 @@
|
||||
from itertools import cycle
|
||||
|
||||
from py.log import Producer
|
||||
from _pytest.runner import CollectReport
|
||||
|
||||
from xdist.slavemanage import parse_spec_config
|
||||
from xdist.report import report_collection_diff
|
||||
|
||||
|
||||
class LoadScheduling:
|
||||
"""Implement load scheduling across nodes.
|
||||
|
||||
This distributes the tests collected across all nodes so each test
|
||||
is run just once. All nodes collect and submit the test suite and
|
||||
when all collections are received it is verified they are
|
||||
identical collections. Then the collection gets divided up in
|
||||
chunks and chunks get submitted to nodes. Whenever a node finishes
|
||||
an item, it calls ``.mark_test_complete()`` which will trigger the
|
||||
scheduler to assign more tests if the number of pending tests for
|
||||
the node falls below a low-watermark.
|
||||
|
||||
When created, ``numnodes`` defines how many nodes are expected to
|
||||
submit a collection. This is used to know when all nodes have
|
||||
finished collection or how large the chunks need to be created.
|
||||
|
||||
Attributes:
|
||||
|
||||
:numnodes: The expected number of nodes taking part. The actual
|
||||
number of nodes will vary during the scheduler's lifetime as
|
||||
nodes are added by the DSession as they are brought up and
|
||||
removed either because of a dead node or normal shutdown. This
|
||||
number is primarily used to know when the initial collection is
|
||||
completed.
|
||||
|
||||
:node2collection: Map of nodes and their test collection. All
|
||||
collections should always be identical.
|
||||
|
||||
:node2pending: Map of nodes and the indices of their pending
|
||||
tests. The indices are an index into ``.pending`` (which is
|
||||
identical to their own collection stored in
|
||||
``.node2collection``).
|
||||
|
||||
:collection: The one collection once it is validated to be
|
||||
identical between all the nodes. It is initialised to None
|
||||
until ``.schedule()`` is called.
|
||||
|
||||
:pending: List of indices of globally pending tests. These are
|
||||
tests which have not yet been allocated to a chunk for a node
|
||||
to process.
|
||||
|
||||
:log: A py.log.Producer instance.
|
||||
|
||||
:config: Config object, used for handling hooks.
|
||||
"""
|
||||
|
||||
def __init__(self, config, log=None):
|
||||
self.numnodes = len(parse_spec_config(config))
|
||||
self.node2collection = {}
|
||||
self.node2pending = {}
|
||||
self.pending = []
|
||||
self.collection = None
|
||||
if log is None:
|
||||
self.log = Producer("loadsched")
|
||||
else:
|
||||
self.log = log.loadsched
|
||||
self.config = config
|
||||
|
||||
@property
|
||||
def nodes(self):
|
||||
"""A list of all nodes in the scheduler."""
|
||||
return list(self.node2pending.keys())
|
||||
|
||||
@property
|
||||
def collection_is_completed(self):
|
||||
"""Boolean indication initial test collection is complete.
|
||||
|
||||
This is a boolean indicating all initial participating nodes
|
||||
have finished collection. The required number of initial
|
||||
nodes is defined by ``.numnodes``.
|
||||
"""
|
||||
return len(self.node2collection) >= self.numnodes
|
||||
|
||||
@property
|
||||
def tests_finished(self):
|
||||
"""Return True if all tests have been executed by the nodes."""
|
||||
if not self.collection_is_completed:
|
||||
return False
|
||||
if self.pending:
|
||||
return False
|
||||
for pending in self.node2pending.values():
|
||||
if len(pending) >= 2:
|
||||
return False
|
||||
return True
|
||||
|
||||
@property
|
||||
def has_pending(self):
|
||||
"""Return True if there are pending test items
|
||||
|
||||
This indicates that collection has finished and nodes are
|
||||
still processing test items, so this can be thought of as
|
||||
"the scheduler is active".
|
||||
"""
|
||||
if self.pending:
|
||||
return True
|
||||
for pending in self.node2pending.values():
|
||||
if pending:
|
||||
return True
|
||||
return False
|
||||
|
||||
def add_node(self, node):
|
||||
"""Add a new node to the scheduler.
|
||||
|
||||
From now on the node will be allocated chunks of tests to
|
||||
execute.
|
||||
|
||||
Called by the ``DSession.slave_slaveready`` hook when it
|
||||
successfully bootstraps a new node.
|
||||
"""
|
||||
assert node not in self.node2pending
|
||||
self.node2pending[node] = []
|
||||
|
||||
def add_node_collection(self, node, collection):
|
||||
"""Add the collected test items from a node
|
||||
|
||||
The collection is stored in the ``.node2collection`` map.
|
||||
Called by the ``DSession.slave_collectionfinish`` hook.
|
||||
"""
|
||||
assert node in self.node2pending
|
||||
if self.collection_is_completed:
|
||||
# A new node has been added later, perhaps an original one died.
|
||||
# .schedule() should have
|
||||
# been called by now
|
||||
assert self.collection
|
||||
if collection != self.collection:
|
||||
other_node = next(iter(self.node2collection.keys()))
|
||||
msg = report_collection_diff(self.collection,
|
||||
collection,
|
||||
other_node.gateway.id,
|
||||
node.gateway.id)
|
||||
self.log(msg)
|
||||
return
|
||||
self.node2collection[node] = list(collection)
|
||||
|
||||
def mark_test_complete(self, node, item_index, duration=0):
|
||||
"""Mark test item as completed by node
|
||||
|
||||
The duration it took to execute the item is used as a hint to
|
||||
the scheduler.
|
||||
|
||||
This is called by the ``DSession.slave_testreport`` hook.
|
||||
"""
|
||||
self.node2pending[node].remove(item_index)
|
||||
self.check_schedule(node, duration=duration)
|
||||
|
||||
def check_schedule(self, node, duration=0):
|
||||
"""Maybe schedule new items on the node
|
||||
|
||||
If there are any globally pending nodes left then this will
|
||||
check if the given node should be given any more tests. The
|
||||
``duration`` of the last test is optionally used as a
|
||||
heuristic to influence how many tests the node is assigned.
|
||||
"""
|
||||
if node.shutting_down:
|
||||
return
|
||||
|
||||
if self.pending:
|
||||
# how many nodes do we have?
|
||||
num_nodes = len(self.node2pending)
|
||||
# if our node goes below a heuristic minimum, fill it out to
|
||||
# heuristic maximum
|
||||
items_per_node_min = max(2, len(self.pending) // num_nodes // 4)
|
||||
items_per_node_max = max(2, len(self.pending) // num_nodes // 2)
|
||||
node_pending = self.node2pending[node]
|
||||
if len(node_pending) < items_per_node_min:
|
||||
if duration >= 0.1 and len(node_pending) >= 2:
|
||||
# seems the node is doing long-running tests
|
||||
# and has enough items to continue
|
||||
# so let's rather wait with sending new items
|
||||
return
|
||||
num_send = items_per_node_max - len(node_pending)
|
||||
self._send_tests(node, num_send)
|
||||
self.log("num items waiting for node:", len(self.pending))
|
||||
|
||||
def remove_node(self, node):
|
||||
"""Remove a node from the scheduler
|
||||
|
||||
This should be called either when the node crashed or at
|
||||
shutdown time. In the former case any pending items assigned
|
||||
to the node will be re-scheduled. Called by the
|
||||
``DSession.slave_slavefinished`` and
|
||||
``DSession.slave_errordown`` hooks.
|
||||
|
||||
Return the item which was being executing while the node
|
||||
crashed or None if the node has no more pending items.
|
||||
|
||||
"""
|
||||
pending = self.node2pending.pop(node)
|
||||
if not pending:
|
||||
return
|
||||
|
||||
# The node crashed, reassing pending items
|
||||
crashitem = self.collection[pending.pop(0)]
|
||||
self.pending.extend(pending)
|
||||
for node in self.node2pending:
|
||||
self.check_schedule(node)
|
||||
return crashitem
|
||||
|
||||
def schedule(self):
|
||||
"""Initiate distribution of the test collection
|
||||
|
||||
Initiate scheduling of the items across the nodes. If this
|
||||
gets called again later it behaves the same as calling
|
||||
``.check_schedule()`` on all nodes so that newly added nodes
|
||||
will start to be used.
|
||||
|
||||
This is called by the ``DSession.slave_collectionfinish`` hook
|
||||
if ``.collection_is_completed`` is True.
|
||||
"""
|
||||
assert self.collection_is_completed
|
||||
|
||||
# Initial distribution already happened, reschedule on all nodes
|
||||
if self.collection is not None:
|
||||
for node in self.nodes:
|
||||
self.check_schedule(node)
|
||||
return
|
||||
|
||||
# XXX allow nodes to have different collections
|
||||
if not self._check_nodes_have_same_collection():
|
||||
self.log('**Different tests collected, aborting run**')
|
||||
return
|
||||
|
||||
# Collections are identical, create the index of pending items.
|
||||
self.collection = list(self.node2collection.values())[0]
|
||||
self.pending[:] = range(len(self.collection))
|
||||
if not self.collection:
|
||||
return
|
||||
|
||||
# Send a batch of tests to run. If we don't have at least two
|
||||
# tests per node, we have to send them all so that we can send
|
||||
# shutdown signals and get all nodes working.
|
||||
initial_batch = max(len(self.pending) // 4,
|
||||
2 * len(self.nodes))
|
||||
|
||||
# distribute tests round-robin up to the batch size
|
||||
# (or until we run out)
|
||||
nodes = cycle(self.nodes)
|
||||
for i in range(initial_batch):
|
||||
self._send_tests(next(nodes), 1)
|
||||
|
||||
if not self.pending:
|
||||
# initial distribution sent all tests, start node shutdown
|
||||
for node in self.nodes:
|
||||
node.shutdown()
|
||||
|
||||
def _send_tests(self, node, num):
|
||||
tests_per_node = self.pending[:num]
|
||||
if tests_per_node:
|
||||
del self.pending[:num]
|
||||
self.node2pending[node].extend(tests_per_node)
|
||||
node.send_runtest_some(tests_per_node)
|
||||
|
||||
def _check_nodes_have_same_collection(self):
|
||||
"""Return True if all nodes have collected the same items.
|
||||
|
||||
If collections differ, this method returns False while logging
|
||||
the collection differences and posting collection errors to
|
||||
pytest_collectreport hook.
|
||||
"""
|
||||
node_collection_items = list(self.node2collection.items())
|
||||
first_node, col = node_collection_items[0]
|
||||
same_collection = True
|
||||
for node, collection in node_collection_items[1:]:
|
||||
msg = report_collection_diff(
|
||||
col,
|
||||
collection,
|
||||
first_node.gateway.id,
|
||||
node.gateway.id,
|
||||
)
|
||||
if msg:
|
||||
same_collection = False
|
||||
self.log(msg)
|
||||
if self.config is not None:
|
||||
rep = CollectReport(
|
||||
node.gateway.id, 'failed',
|
||||
longrepr=msg, result=[])
|
||||
self.config.hook.pytest_collectreport(report=rep)
|
||||
|
||||
return same_collection
|
||||
Reference in New Issue
Block a user