Implement work-stealing scheduler (#862)

Closes #858

src/xdist/dsession.py

@@ -8,6 +8,7 @@ from xdist.scheduler import (
     LoadScopeScheduling,
     LoadFileScheduling,
     LoadGroupScheduling,
+    WorkStealingScheduling,
 )
 
 
@@ -100,6 +101,7 @@ class DSession:
             "loadscope": LoadScopeScheduling,
             "loadfile": LoadFileScheduling,
             "loadgroup": LoadGroupScheduling,
+            "worksteal": WorkStealingScheduling,
         }
         return schedulers[dist](config, log)
 
@@ -282,6 +284,17 @@ class DSession:
         """
         self.sched.mark_test_complete(node, item_index, duration)
 
+    def worker_unscheduled(self, node, indices):
+        """
+        Emitted when a node fires the 'unscheduled' event, signalling that
+        some tests have been removed from the worker's queue and should be
+        sent to some worker again.
+
+        This should happen only in response to 'steal' command, so schedulers
+        not using 'steal' command don't have to implement it.
+        """
+        self.sched.remove_pending_tests_from_node(node, indices)
+
     def worker_collectreport(self, node, rep):
         """Emitted when a node calls the pytest_collectreport hook.
 

src/xdist/plugin.py

@@ -94,7 +94,15 @@ def pytest_addoption(parser):
         "--dist",
         metavar="distmode",
         action="store",
-        choices=["each", "load", "loadscope", "loadfile", "loadgroup", "no"],
+        choices=[
+            "each",
+            "load",
+            "loadscope",
+            "loadfile",
+            "loadgroup",
+            "worksteal",
+            "no",
+        ],
         dest="dist",
         default="no",
         help=(
@@ -107,6 +115,8 @@ def pytest_addoption(parser):
             "loadfile: load balance by sending test grouped by file"
             " to any available environment.\n\n"
             "loadgroup: like load, but sends tests marked with 'xdist_group' to the same worker.\n\n"
+            "worksteal: split the test suite between available environments,"
+            " then rebalance when any worker runs out of tests.\n\n"
             "(default) no: run tests inprocess, don't distribute."
         ),
     )

src/xdist/remote.py

@@ -6,6 +6,7 @@
     needs not to be installed in remote environments.
 """
 
+import contextlib
 import sys
 import os
 import time
@@ -56,14 +57,21 @@ def worker_title(title):
 
 
 class WorkerInteractor:
+    SHUTDOWN_MARK = object()
+
     def __init__(self, config, channel):
         self.config = config
         self.workerid = config.workerinput.get("workerid", "?")
         self.testrunuid = config.workerinput["testrunuid"]
         self.log = Producer(f"worker-{self.workerid}", enabled=config.option.debug)
         self.channel = channel
+        self.torun = self._make_queue()
+        self.nextitem_index = None
         config.pluginmanager.register(self)
 
+    def _make_queue(self):
+        return self.channel.gateway.execmodel.queue.Queue()
+
     def sendevent(self, name, **kwargs):
         self.log("sending", name, kwargs)
         self.channel.send((name, kwargs))
@@ -92,38 +100,60 @@ class WorkerInteractor:
     def pytest_collection(self, session):
         self.sendevent("collectionstart")
 
+    def handle_command(self, command):
+        if command is self.SHUTDOWN_MARK:
+            self.torun.put(self.SHUTDOWN_MARK)
+            return
+
+        name, kwargs = command
+
+        self.log("received command", name, kwargs)
+        if name == "runtests":
+            for i in kwargs["indices"]:
+                self.torun.put(i)
+        elif name == "runtests_all":
+            for i in range(len(self.session.items)):
+                self.torun.put(i)
+        elif name == "shutdown":
+            self.torun.put(self.SHUTDOWN_MARK)
+        elif name == "steal":
+            self.steal(kwargs["indices"])
+
+    def steal(self, indices):
+        indices = set(indices)
+        stolen = []
+
+        old_queue, self.torun = self.torun, self._make_queue()
+
+        def old_queue_get_nowait_noraise():
+            with contextlib.suppress(self.channel.gateway.execmodel.queue.Empty):
+                return old_queue.get_nowait()
+
+        for i in iter(old_queue_get_nowait_noraise, None):
+            if i in indices:
+                stolen.append(i)
+            else:
+                self.torun.put(i)
+
+        self.sendevent("unscheduled", indices=stolen)
+
     @pytest.hookimpl
     def pytest_runtestloop(self, session):
         self.log("entering main loop")
-        torun = []
-        while 1:
-            try:
-                name, kwargs = self.channel.receive()
-            except EOFError:
-                return True
-            self.log("received command", name, kwargs)
-            if name == "runtests":
-                torun.extend(kwargs["indices"])
-            elif name == "runtests_all":
-                torun.extend(range(len(session.items)))
-            self.log("items to run:", torun)
-            # only run if we have an item and a next item
-            while len(torun) >= 2:
-                self.run_one_test(torun)
-            if name == "shutdown":
-                if torun:
-                    self.run_one_test(torun)
-                break
+        self.channel.setcallback(self.handle_command, endmarker=self.SHUTDOWN_MARK)
+        self.nextitem_index = self.torun.get()
+        while self.nextitem_index is not self.SHUTDOWN_MARK:
+            self.run_one_test()
         return True
 
-    def run_one_test(self, torun):
+    def run_one_test(self):
         items = self.session.items
-        self.item_index = torun.pop(0)
+        self.item_index, self.nextitem_index = self.nextitem_index, self.torun.get()
         item = items[self.item_index]
-        if torun:
-            nextitem = items[torun[0]]
-        else:
+        if self.nextitem_index is self.SHUTDOWN_MARK:
             nextitem = None
+        else:
+            nextitem = items[self.nextitem_index]
 
         worker_title("[pytest-xdist running] %s" % item.nodeid)
 

src/xdist/scheduler/__init__.py

@@ -3,3 +3,4 @@ from xdist.scheduler.load import LoadScheduling  # noqa
 from xdist.scheduler.loadfile import LoadFileScheduling  # noqa
 from xdist.scheduler.loadscope import LoadScopeScheduling  # noqa
 from xdist.scheduler.loadgroup import LoadGroupScheduling  # noqa
+from xdist.scheduler.worksteal import WorkStealingScheduling  # noqa

src/xdist/scheduler/worksteal.py

@@ -0,0 +1,325 @@
+from collections import namedtuple
+
+from _pytest.runner import CollectReport
+
+from xdist.remote import Producer
+from xdist.workermanage import parse_spec_config
+from xdist.report import report_collection_diff
+
+
+NodePending = namedtuple("NodePending", ["node", "pending"])
+
+# Every worker needs at least 2 tests in queue - the current and the next one.
+MIN_PENDING = 2
+
+
+class WorkStealingScheduling:
+    """Implement work-stealing scheduling.
+
+    Initially, tests are distributed evenly among all nodes.
+
+    When some node completes most of its assigned tests (when only one pending
+    test remains), an attempt is made to reassign ("steal") some tests from
+    other nodes to this node.
+
+    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.
+
+    :steal_requested_from_node: The node to which the current "steal" request
+       was sent. ``None`` if there is no request in progress. Only one request
+       can be in progress at any time, the scheduler doesn't send multiple
+       simultaneous requests.
+    """
+
+    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("workstealsched")
+        else:
+            self.log = log.workstealsched
+        self.config = config
+        self.steal_requested_from_node = None
+
+    @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
+        if self.steal_requested_from_node is not None:
+            return False
+        for pending in self.node2pending.values():
+            if len(pending) >= MIN_PENDING:
+                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=None):
+        """Mark test item as completed by node
+
+        This is called by the ``DSession.worker_testreport`` hook.
+        """
+        self.node2pending[node].remove(item_index)
+        self.check_schedule()
+
+    def mark_test_pending(self, item):
+        self.pending.insert(
+            0,
+            self.collection.index(item),
+        )
+        self.check_schedule()
+
+    def remove_pending_tests_from_node(self, node, indices):
+        """Node returned some test indices back in response to 'steal' command.
+
+        This is called by ``DSession.worker_unscheduled``.
+        """
+        assert node is self.steal_requested_from_node
+        self.steal_requested_from_node = None
+
+        indices_set = set(indices)
+        self.node2pending[node] = [
+            i for i in self.node2pending[node] if i not in indices_set
+        ]
+        self.pending.extend(indices)
+        self.check_schedule()
+
+    def check_schedule(self):
+        """Reschedule tests/perform load balancing."""
+        nodes_up = [
+            NodePending(node, pending)
+            for node, pending in self.node2pending.items()
+            if not node.shutting_down
+        ]
+
+        def get_idle_nodes():
+            return [node for node, pending in nodes_up if len(pending) < MIN_PENDING]
+
+        idle_nodes = get_idle_nodes()
+        if not idle_nodes:
+            return
+
+        if self.pending:
+            # Distribute pending tests evenly among idle nodes
+            for i, node in enumerate(idle_nodes):
+                nodes_remaining = len(idle_nodes) - i
+                num_send = len(self.pending) // nodes_remaining
+                self._send_tests(node, num_send)
+
+            idle_nodes = get_idle_nodes()
+            # No need to steal anything if all nodes have enough work to continue
+            if not idle_nodes:
+                return
+
+        # Only one active stealing request is allowed
+        if self.steal_requested_from_node is not None:
+            return
+
+        # Find the node that has the longest test queue
+        steal_from = max(
+            nodes_up, key=lambda node_pending: len(node_pending.pending), default=None
+        )
+
+        if steal_from is None:
+            num_steal = 0
+        else:
+            # Steal half of the test queue - but keep that node running too.
+            # If the node has 2 or less tests queued, stealing will fail
+            # anyway.
+            max_steal = max(0, len(steal_from.pending) - MIN_PENDING)
+            num_steal = min(len(steal_from.pending) // 2, max_steal)
+
+        if num_steal == 0:
+            # Can't get more work - shutdown idle nodes. This will force them
+            # to run the last test now instead of waiting for more tests.
+            for node in idle_nodes:
+                node.shutdown()
+            return
+
+        steal_from.node.send_steal(steal_from.pending[-num_steal:])
+        self.steal_requested_from_node = steal_from.node
+
+    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 node was removed without completing its assigned tests - it crashed
+        if pending:
+            crashitem = self.collection[pending.pop(0)]
+        else:
+            crashitem = None
+
+        self.pending.extend(pending)
+
+        # Dead node won't respond to "steal" request
+        if self.steal_requested_from_node is node:
+            self.steal_requested_from_node = None
+
+        self.check_schedule()
+        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:
+            self.check_schedule()
+            return
+
+        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
+
+        self.check_schedule()
+
+    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

src/xdist/workermanage.py

@@ -300,6 +300,9 @@ class WorkerController:
     def send_runtest_all(self):
         self.sendcommand("runtests_all")
 
+    def send_steal(self, indices):
+        self.sendcommand("steal", indices=indices)
+
     def shutdown(self):
         if not self._down:
             try:
@@ -359,6 +362,8 @@ class WorkerController:
                 self.notify_inproc(eventname, node=self, ids=kwargs["ids"])
             elif eventname == "runtest_protocol_complete":
                 self.notify_inproc(eventname, node=self, **kwargs)
+            elif eventname == "unscheduled":
+                self.notify_inproc(eventname, node=self, **kwargs)
             elif eventname == "logwarning":
                 self.notify_inproc(
                     eventname,