IQ.Lvbs/pytest-xdist
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Adopt 'src' layout and add 'testing' extras

Browse Source
This commit is contained in:
Bruno Oliveira
2019-06-06 19:35:13 -03:00
parent cdf422dec8
commit 7d41db3685
16 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
src/xdist/scheduler/__init__.py Normal file
View File

@@ -0,0 +1,4 @@
from xdist.scheduler.each import EachScheduling # noqa
from xdist.scheduler.load import LoadScheduling # noqa
from xdist.scheduler.loadfile import LoadFileScheduling # noqa
from xdist.scheduler.loadscope import LoadScopeScheduling # noqa

132
src/xdist/scheduler/each.py Normal file
View File

@@ -0,0 +1,132 @@
from py.log import Producer
from xdist.workermanage import parse_spec_config
from xdist.report import report_collection_diff
class EachScheduling(object):
"""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()
node.shutdown()
else:
node.send_runtest_some(pending)
self._started.append(node)

286
src/xdist/scheduler/load.py Normal file
View File

@@ -0,0 +1,286 @@
from itertools import cycle
from py.log import Producer
from _pytest.runner import CollectReport
from xdist.workermanage import parse_spec_config
from xdist.report import report_collection_diff
class LoadScheduling(object):
"""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.worker_workerready`` 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.worker_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.worker_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)
else:
node.shutdown()
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.worker_workerfinished`` and
``DSession.worker_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.worker_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

52
src/xdist/scheduler/loadfile.py Normal file
View File

@@ -0,0 +1,52 @@
from .loadscope import LoadScopeScheduling
from py.log import Producer
class LoadFileScheduling(LoadScopeScheduling):
"""Implement load scheduling across nodes, but grouping test test file.
This distributes the tests collected across all nodes so each test is run
just once. All nodes collect and submit the list of tests and when all
collections are received it is verified they are identical collections.
Then the collection gets divided up in work units, grouped by test file,
and those work units get submitted to nodes. Whenever a node finishes an
item, it calls ``.mark_test_complete()`` which will trigger the scheduler
to assign more work units 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.
This class behaves very much like LoadScopeScheduling, but with a file-level scope.
"""
def __init__(self, config, log=None):
super(LoadFileScheduling, self).__init__(config, log)
if log is None:
self.log = Producer("loadfilesched")
else:
self.log = log.loadfilesched
def _split_scope(self, nodeid):
"""Determine the scope (grouping) of a nodeid.
There are usually 3 cases for a nodeid::
example/loadsuite/test/test_beta.py::test_beta0
example/loadsuite/test/test_delta.py::Delta1::test_delta0
example/loadsuite/epsilon/__init__.py::epsilon.epsilon
#. Function in a test module.
#. Method of a class in a test module.
#. Doctest in a function in a package.
This function will group tests with the scope determined by splitting
the first ``::`` from the left. That is, test will be grouped in a
single work unit when they reside in the same file.
In the above example, scopes will be::
example/loadsuite/test/test_beta.py
example/loadsuite/test/test_delta.py
example/loadsuite/epsilon/__init__.py
"""
return nodeid.split("::", 1)[0]

409
src/xdist/scheduler/loadscope.py Normal file
View File

@@ -0,0 +1,409 @@
from collections import OrderedDict
from _pytest.runner import CollectReport
from py.log import Producer
from xdist.report import report_collection_diff
from xdist.workermanage import parse_spec_config
class LoadScopeScheduling(object):
"""Implement load scheduling across nodes, but grouping test by scope.
This distributes the tests collected across all nodes so each test is run
just once. All nodes collect and submit the list of tests and when all
collections are received it is verified they are identical collections.
Then the collection gets divided up in work units, grouped by test scope,
and those work units get submitted to nodes. Whenever a node finishes an
item, it calls ``.mark_test_complete()`` which will trigger the scheduler
to assign more work units 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.
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.
:collection: The final list of tests collected by all nodes once it is
validated to be identical between all the nodes. It is initialised to
None until ``.schedule()`` is called.
:workqueue: Ordered dictionary that maps all available scopes with their
associated tests (nodeid). Nodeids are in turn associated with their
completion status. One entry of the workqueue is called a work unit.
In turn, a collection of work unit is called a workload.
::
workqueue = {
'<full>/<path>/<to>/test_module.py': {
'<full>/<path>/<to>/test_module.py::test_case1': False,
'<full>/<path>/<to>/test_module.py::test_case2': False,
(...)
},
(...)
}
:assigned_work: Ordered dictionary that maps worker nodes with their
assigned work units.
::
assigned_work = {
'<worker node A>': {
'<full>/<path>/<to>/test_module.py': {
'<full>/<path>/<to>/test_module.py::test_case1': False,
'<full>/<path>/<to>/test_module.py::test_case2': False,
(...)
},
(...)
},
(...)
}
:registered_collections: Ordered dictionary that maps worker nodes with
their collection of tests gathered during test discovery.
::
registered_collections = {
'<worker node A>': [
'<full>/<path>/<to>/test_module.py::test_case1',
'<full>/<path>/<to>/test_module.py::test_case2',
],
(...)
}
: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.collection = None
self.workqueue = OrderedDict()
self.assigned_work = OrderedDict()
self.registered_collections = OrderedDict()
if log is None:
self.log = Producer("loadscopesched")
else:
self.log = log.loadscopesched
self.config = config
@property
def nodes(self):
"""A list of all active nodes in the scheduler."""
return list(self.assigned_work.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.registered_collections) >= 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.workqueue:
return False
for assigned_unit in self.assigned_work.values():
if self._pending_of(assigned_unit) >= 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.workqueue:
return True
for assigned_unit in self.assigned_work.values():
if self._pending_of(assigned_unit) > 0:
return True
return False
def add_node(self, node):
"""Add a new node to the scheduler.
From now on the node will be assigned work units to be executed.
Called by the ``DSession.worker_workerready`` hook when it successfully
bootstraps a new node.
"""
assert node not in self.assigned_work
self.assigned_work[node] = OrderedDict()
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 hooks:
- ``DSession.worker_workerfinished``.
- ``DSession.worker_errordown``.
Return the item being executed while the node crashed or None if the
node has no more pending items.
"""
workload = self.assigned_work.pop(node)
if not self._pending_of(workload):
return None
# The node crashed, identify test that crashed
for work_unit in workload.values():
for nodeid, completed in work_unit.items():
if not completed:
crashitem = nodeid
break
else:
continue
break
else:
raise RuntimeError(
"Unable to identify crashitem on a workload with pending items"
)
# Made uncompleted work unit available again
self.workqueue.update(workload)
for node in self.assigned_work:
self._reschedule(node)
return crashitem
def add_node_collection(self, node, collection):
"""Add the collected test items from a node.
The collection is stored in the ``.registered_collections`` dictionary.
Called by the hook:
- ``DSession.worker_collectionfinish``.
"""
# Check that add_node() was called on the node before
assert node in self.assigned_work
# A new node has been added later, perhaps an original one died.
if self.collection_is_completed:
# Assert that .schedule() should have been called by now
assert self.collection
# Check that the new collection matches the official collection
if collection != self.collection:
other_node = next(iter(self.registered_collections.keys()))
msg = report_collection_diff(
self.collection, collection, other_node.gateway.id, node.gateway.id
)
self.log(msg)
return
self.registered_collections[node] = list(collection)
def mark_test_complete(self, node, item_index, duration=0):
"""Mark test item as completed by node.
Called by the hook:
- ``DSession.worker_testreport``.
"""
nodeid = self.registered_collections[node][item_index]
scope = self._split_scope(nodeid)
self.assigned_work[node][scope][nodeid] = True
self._reschedule(node)
def _assign_work_unit(self, node):
"""Assign a work unit to a node."""
assert self.workqueue
# Grab a unit of work
scope, work_unit = self.workqueue.popitem(last=False)
# Keep track of the assigned work
assigned_to_node = self.assigned_work.setdefault(node, default=OrderedDict())
assigned_to_node[scope] = work_unit
# Ask the node to execute the workload
worker_collection = self.registered_collections[node]
nodeids_indexes = [
worker_collection.index(nodeid)
for nodeid, completed in work_unit.items()
if not completed
]
node.send_runtest_some(nodeids_indexes)
def _split_scope(self, nodeid):
"""Determine the scope (grouping) of a nodeid.
There are usually 3 cases for a nodeid::
example/loadsuite/test/test_beta.py::test_beta0
example/loadsuite/test/test_delta.py::Delta1::test_delta0
example/loadsuite/epsilon/__init__.py::epsilon.epsilon
#. Function in a test module.
#. Method of a class in a test module.
#. Doctest in a function in a package.
This function will group tests with the scope determined by splitting
the first ``::`` from the right. That is, classes will be grouped in a
single work unit, and functions from a test module will be grouped by
their module. In the above example, scopes will be::
example/loadsuite/test/test_beta.py
example/loadsuite/test/test_delta.py::Delta1
example/loadsuite/epsilon/__init__.py
"""
return nodeid.rsplit("::", 1)[0]
def _pending_of(self, workload):
"""Return the number of pending tests in a workload."""
pending = sum(list(scope.values()).count(False) for scope in workload.values())
return pending
def _reschedule(self, node):
"""Maybe schedule new items on the node.
If there are any globally pending work units left then this will check
if the given node should be given any more tests.
"""
# Do not add more work to a node shutting down
if node.shutting_down:
return
# Check that more work is available
if not self.workqueue:
node.shutdown()
return
self.log("Number of units waiting for node:", len(self.workqueue))
# Check that the node is almost depleted of work
# 2: Heuristic of minimum tests to enqueue more work
if self._pending_of(self.assigned_work[node]) > 2:
return
# Pop one unit of work and assign it
self._assign_work_unit(node)
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 ``._reschedule()`` on all
nodes so that newly added nodes will start to be used.
If ``.collection_is_completed`` is True, this is called by the hook:
- ``DSession.worker_collectionfinish``.
"""
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._reschedule(node)
return
# Check that all nodes collected the same tests
if not self._check_nodes_have_same_collection():
self.log("**Different tests collected, aborting run**")
return
# Collections are identical, create the final list of items
self.collection = list(next(iter(self.registered_collections.values())))
if not self.collection:
return
# Determine chunks of work (scopes)
for nodeid in self.collection:
scope = self._split_scope(nodeid)
work_unit = self.workqueue.setdefault(scope, default=OrderedDict())
work_unit[nodeid] = False
# Avoid having more workers than work
extra_nodes = len(self.nodes) - len(self.workqueue)
if extra_nodes > 0:
self.log("Shuting down {0} nodes".format(extra_nodes))
for _ in range(extra_nodes):
unused_node, assigned = self.assigned_work.popitem(last=True)
self.log("Shuting down unused node {0}".format(unused_node))
unused_node.shutdown()
# Assign initial workload
for node in self.nodes:
self._assign_work_unit(node)
# Ensure nodes start with at least two work units if possible (#277)
for node in self.nodes:
self._reschedule(node)
# Initial distribution sent all tests, start node shutdown
if not self.workqueue:
for node in self.nodes:
node.shutdown()
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.registered_collections.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 not msg:
continue
same_collection = False
self.log(msg)
if self.config is None:
continue
rep = CollectReport(node.gateway.id, "failed", longrepr=msg, result=[])
self.config.hook.pytest_collectreport(report=rep)
return same_collection