File size: 14,695 Bytes
9368cc4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 | #!/usr/bin/env python3
"""Tasks built on the `flow` scheduler repo.
Instructions are written the way a colleague would report the problem: a
symptom, and how to see it. They never name the faulty function, never state
the expected values, and never describe the fix.
"""
import sys
from pathlib import Path
sys.path.insert(0, str(Path(__file__).parent))
from make_tasks import make, spec # noqa: E402
R = "python/scheduler"
# ---------------------------------------------------------------- retry never fires
make(R, "retry-never-fires",
spec("python", "logic", 4, """
A workflow with a retry policy hangs instead of retrying.
To see it: define one task with RetryPolicy(max_attempts=3), start it, and
report it failed. The run then never makes progress again -- the task is not
finished, but nothing the scheduler offers ever includes it again, even after
the clock is advanced well past any backoff.
Retries are supposed to become eligible again once their delay has elapsed.
Track down why they don't and fix it.
"""),
[("flow/scheduler.py",
""" # RETRYING counts as a candidate: its backoff is enforced below by
# ready_at, and leaving it out would mean a retry never fires.
if state.get(task_id) not in (st.PENDING, st.READY, st.RETRYING):
continue""",
""" if state.get(task_id) not in (st.PENDING, st.READY):
continue""")],
{"test_retry.py": '''
import unittest
from flow import ResourcePool, ResourceRequest, RetryPolicy, Scheduler, Task, TaskGraph
from flow import state as st
def one_task(**kw):
graph = TaskGraph([Task("job", **kw)])
return Scheduler(graph, ResourcePool({"cpu": 4}))
class TestRetryBecomesEligible(unittest.TestCase):
def test_retry_is_offered_after_backoff(self):
s = one_task(retry=RetryPolicy(max_attempts=3, base_delay=10.0))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
s.clock.advance(10.0)
self.assertEqual(s.next_batch(state), ["job"])
def test_retry_runs_to_success(self):
s = one_task(retry=RetryPolicy(max_attempts=3, base_delay=1.0))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
s.clock.advance(1.0)
s.start("job", state)
s.finish("job", state, True)
self.assertEqual(state.get("job"), st.SUCCEEDED)
self.assertEqual(state.attempts["job"], 2)
def test_run_reaches_completion(self):
s = one_task(retry=RetryPolicy(max_attempts=2, base_delay=1.0))
state = s.new_state()
for _ in range(6):
batch = s.next_batch(state)
if not batch:
s.clock.advance(5.0)
continue
s.start(batch[0], state)
s.fail(batch[0], state)
if state.is_complete():
break
self.assertTrue(state.is_complete())
self.assertEqual(state.get("job"), st.FAILED)
class TestBackoffStillEnforced(unittest.TestCase):
"""A retry must not become eligible before its delay has elapsed."""
def test_not_offered_before_backoff(self):
s = one_task(retry=RetryPolicy(max_attempts=3, base_delay=30.0))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
self.assertEqual(s.next_batch(state), [])
s.clock.advance(29.0)
self.assertEqual(s.next_batch(state), [])
def test_terminal_tasks_are_never_offered(self):
s = one_task()
state = s.new_state()
s.start("job", state)
s.finish("job", state, True)
self.assertEqual(s.next_batch(state), [])
def test_exhausted_retries_end_as_failed(self):
s = one_task(retry=RetryPolicy(max_attempts=1))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
self.assertEqual(state.get("job"), st.FAILED)
'''})
# ---------------------------------------------------------------- resource leak
make(R, "resource-leak-on-retry",
spec("python", "logic", 4, """
A long-running workflow gradually stops scheduling anything.
Our pipeline has a pool of 4 "cpu" units and several flaky tasks that fail and
retry. The first few retries behave, but after enough failures the scheduler
stops offering work entirely, even though plenty of tasks are still pending and
nothing is actually running. Restarting the process clears it.
It looks like capacity is going missing over the life of a run. Please find out
where and fix it.
"""),
[("flow/scheduler.py",
""" The pool is released on every failure, not only the last one: a task
that is going to be retried must not keep holding capacity while it
waits out its backoff.
\"\"\"
self.pool.release(task_id)
task = self.graph.get(task_id)""",
""" The pool is released on every failure, not only the last one: a task
that is going to be retried must not keep holding capacity while it
waits out its backoff.
\"\"\"
task = self.graph.get(task_id)""")],
{"test_resource_lifecycle.py": '''
import unittest
from flow import (ResourcePool, ResourceRequest, RetryPolicy, Scheduler, Task,
TaskGraph)
def scheduler(capacity=4, **kw):
graph = TaskGraph([Task("job", resources=(ResourceRequest("cpu", 2),), **kw)])
return Scheduler(graph, ResourcePool({"cpu": capacity}))
class TestCapacityIsReturned(unittest.TestCase):
def test_failure_returns_capacity(self):
s = scheduler(retry=RetryPolicy(max_attempts=3, base_delay=1.0))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
self.assertEqual(s.pool.free("cpu"), 4)
def test_capacity_is_stable_across_many_retries(self):
s = scheduler(retry=RetryPolicy(max_attempts=5, base_delay=1.0))
state = s.new_state()
for _ in range(4):
s.clock.advance(60.0)
batch = s.next_batch(state)
if not batch:
break
s.start("job", state)
s.fail("job", state)
self.assertEqual(s.pool.free("cpu"), 4)
def test_nothing_is_still_held_after_failure(self):
s = scheduler(retry=RetryPolicy(max_attempts=3, base_delay=1.0))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
self.assertEqual(s.pool.held_by("job"), {})
self.assertEqual(s.pool.in_use(), {"cpu": 0})
class TestNormalPathUnaffected(unittest.TestCase):
def test_success_returns_capacity(self):
s = scheduler()
state = s.new_state()
s.start("job", state)
s.finish("job", state, True)
self.assertEqual(s.pool.free("cpu"), 4)
def test_capacity_is_held_while_running(self):
s = scheduler()
state = s.new_state()
s.start("job", state)
self.assertEqual(s.pool.free("cpu"), 2)
def test_final_failure_returns_capacity(self):
s = scheduler(retry=RetryPolicy(max_attempts=1))
state = s.new_state()
s.start("job", state)
s.fail("job", state)
self.assertEqual(s.pool.free("cpu"), 4)
'''})
# ---------------------------------------------------------------- batch overcommit
make(R, "batch-overcommit",
spec("python", "logic", 4, """
The scheduler hands us more work than our resource pool can support.
With a pool of 3 "cpu" units and two tasks that each request 2, a single call to
next_batch comes back with both of them. Starting both then blows up with a
ResourceExhausted from deep inside the pool.
Each task on its own fits, so the pool's accounting looks right to us -- it's
the batch that seems wrong. Please fix it so a batch is always startable.
"""),
[("flow/scheduler.py",
""" chosen: List[str] = []
reserved: Dict[str, int] = {}
in_flight = self.running_count(state)
for task_id in self.eligible(state):
if self.max_parallel and in_flight + len(chosen) >= self.max_parallel:
break
task = self.graph.get(task_id)
fits = True
for request in task.resources:
already = reserved.get(request.name, 0)
if self.pool.free(request.name) - already < request.amount:
fits = False
break
if not fits:
continue
for request in task.resources:
reserved[request.name] = reserved.get(request.name, 0) + request.amount
chosen.append(task_id)
return chosen""",
""" chosen: List[str] = []
in_flight = self.running_count(state)
for task_id in self.eligible(state):
if self.max_parallel and in_flight + len(chosen) >= self.max_parallel:
break
task = self.graph.get(task_id)
if not self.pool.can_admit(task.resources):
continue
chosen.append(task_id)
return chosen""")],
{"test_batching.py": '''
import unittest
from flow import ResourcePool, ResourceRequest, Scheduler, Task, TaskGraph
def sched(capacity, tasks, **kw):
return Scheduler(TaskGraph(tasks), ResourcePool(capacity), **kw)
class TestBatchIsStartable(unittest.TestCase):
def test_batch_fits_within_capacity(self):
s = sched({"cpu": 3},
[Task("a", resources=(ResourceRequest("cpu", 2),)),
Task("b", resources=(ResourceRequest("cpu", 2),))])
self.assertEqual(s.next_batch(s.new_state()), ["a"])
def test_whole_batch_can_actually_start(self):
s = sched({"cpu": 5},
[Task("a", resources=(ResourceRequest("cpu", 2),)),
Task("b", resources=(ResourceRequest("cpu", 2),)),
Task("c", resources=(ResourceRequest("cpu", 2),))])
state = s.new_state()
batch = s.next_batch(state)
for task_id in batch:
s.start(task_id, state) # must not raise
self.assertEqual(len(batch), 2)
def test_three_way_split(self):
s = sched({"slots": 4},
[Task(name, resources=(ResourceRequest("slots", 3),))
for name in ("a", "b", "c")])
self.assertEqual(s.next_batch(s.new_state()), ["a"])
class TestBatchingOtherwiseUnchanged(unittest.TestCase):
def test_resourceless_tasks_all_admitted(self):
s = sched({}, [Task("a"), Task("b"), Task("c")])
self.assertEqual(s.next_batch(s.new_state()), ["a", "b", "c"])
def test_max_parallel_still_caps(self):
s = sched({}, [Task("a"), Task("b"), Task("c")], max_parallel=2)
self.assertEqual(len(s.next_batch(s.new_state())), 2)
def test_priority_order_preserved(self):
s = sched({}, [Task("low", priority=0), Task("high", priority=9)])
self.assertEqual(s.next_batch(s.new_state()), ["high", "low"])
def test_everything_fits_when_capacity_is_ample(self):
s = sched({"cpu": 10},
[Task("a", resources=(ResourceRequest("cpu", 1),)),
Task("b", resources=(ResourceRequest("cpu", 1),))])
self.assertEqual(s.next_batch(s.new_state()), ["a", "b"])
'''})
# ---------------------------------------------------------------- cascade depth
make(R, "cascade-shallow",
spec("python", "logic", 4, """
Runs containing a failure never finish.
Our deploy graph is a chain: fetch -> build -> test -> package -> ship. When
build fails, the run is left sitting there forever. Inspecting the state, test
has been skipped as we'd expect, but package and ship are still pending and the
scheduler will not offer them (correctly -- their dependency never succeeded),
so the run is simply stuck and is_complete() never becomes true.
Anything that can no longer run should end up in a terminal state. Please fix.
"""),
[("flow/scheduler.py",
""" skipped: List[str] = []
for downstream in sorted(self.graph.descendants_of(task_id)):""",
""" skipped: List[str] = []
for downstream in sorted(self.graph.dependents_of(task_id)):""")],
{"test_cascade.py": '''
import unittest
from flow import ResourcePool, Scheduler, Task, TaskGraph
from flow import state as st
def chain(*names):
tasks = []
previous = None
for name in names:
deps = frozenset({previous}) if previous else frozenset()
tasks.append(Task(name, depends_on=deps))
previous = name
return Scheduler(TaskGraph(tasks), ResourcePool({}))
class TestFailurePropagatesFully(unittest.TestCase):
def test_whole_chain_below_a_failure_is_skipped(self):
s = chain("fetch", "build", "test", "package", "ship")
state = s.new_state()
s.start("fetch", state)
s.finish("fetch", state, True)
s.start("build", state)
s.fail("build", state)
for name in ("test", "package", "ship"):
self.assertEqual(state.get(name), st.SKIPPED, name)
def test_run_completes_after_a_failure(self):
s = chain("a", "b", "c", "d")
state = s.new_state()
s.start("a", state)
s.fail("a", state)
self.assertTrue(state.is_complete())
def test_diamond_below_a_failure(self):
tasks = [Task("root"),
Task("left", depends_on=frozenset({"root"})),
Task("right", depends_on=frozenset({"root"})),
Task("join", depends_on=frozenset({"left", "right"})),
Task("after", depends_on=frozenset({"join"}))]
s = Scheduler(TaskGraph(tasks), ResourcePool({}))
state = s.new_state()
s.start("root", state)
s.fail("root", state)
self.assertTrue(state.is_complete())
self.assertEqual(state.get("after"), st.SKIPPED)
class TestUnrelatedWorkSurvives(unittest.TestCase):
def test_independent_branch_is_untouched(self):
tasks = [Task("a"), Task("b", depends_on=frozenset({"a"})),
Task("x"), Task("y", depends_on=frozenset({"x"}))]
s = Scheduler(TaskGraph(tasks), ResourcePool({}))
state = s.new_state()
s.start("a", state)
s.fail("a", state)
self.assertEqual(state.get("x"), st.PENDING)
self.assertEqual(state.get("y"), st.PENDING)
def test_successful_run_skips_nothing(self):
s = chain("a", "b", "c")
state = s.new_state()
for name in ("a", "b", "c"):
s.start(name, state)
s.finish(name, state, True)
self.assertEqual(state.counts().get("skipped", 0), 0)
'''})
print("done")
|