sdk/cwl/arvados_cwl/__init__.py

   1 #!/usr/bin/env python
   2
   3 # Implement cwl-runner interface for submitting and running work on Arvados, using
   4 # either the Crunch jobs API or Crunch containers API.
   5
   6 import argparse
   7 import logging
   8 import os
   9 import sys
  10 import threading
  11 import hashlib
  12 import copy
  13 import json
  14 from functools import partial
  15 import pkg_resources  # part of setuptools
  16
  17 from cwltool.errors import WorkflowException
  18 import cwltool.main
  19 import cwltool.workflow
  20 import schema_salad
  21
  22 import arvados
  23 import arvados.config
  24
  25 from .arvcontainer import ArvadosContainer, RunnerContainer
  26 from .arvjob import ArvadosJob, RunnerJob, RunnerTemplate
  27 from. runner import Runner, upload_instance
  28 from .arvtool import ArvadosCommandTool
  29 from .arvworkflow import ArvadosWorkflow, upload_workflow
  30 from .fsaccess import CollectionFsAccess
  31 from .perf import Perf
  32 from .pathmapper import FinalOutputPathMapper
  33 from ._version import __version__
  34
  35 from cwltool.pack import pack
  36 from cwltool.process import shortname, UnsupportedRequirement, getListing
  37 from cwltool.pathmapper import adjustFileObjs, adjustDirObjs
  38 from cwltool.draft2tool import compute_checksums
  39 from arvados.api import OrderedJsonModel
  40
  41 logger = logging.getLogger('arvados.cwl-runner')
  42 metrics = logging.getLogger('arvados.cwl-runner.metrics')
  43 logger.setLevel(logging.INFO)
  44
  45
  46 class ArvCwlRunner(object):
  47     """Execute a CWL tool or workflow, submit work (using either jobs or
  48     containers API), wait for them to complete, and report output.
  49
  50     """
  51
  52     def __init__(self, api_client, work_api=None, keep_client=None, output_name=None):
  53         self.api = api_client
  54         self.processes = {}
  55         self.lock = threading.Lock()
  56         self.cond = threading.Condition(self.lock)
  57         self.final_output = None
  58         self.final_status = None
  59         self.uploaded = {}
  60         self.num_retries = 4
  61         self.uuid = None
  62         self.work_api = work_api
  63         self.stop_polling = threading.Event()
  64         self.poll_api = None
  65         self.pipeline = None
  66         self.final_output_collection = None
  67         self.output_name = output_name
  68         self.project_uuid = None
  69
  70         if keep_client is not None:
  71             self.keep_client = keep_client
  72         else:
  73             self.keep_client = arvados.keep.KeepClient(api_client=self.api, num_retries=self.num_retries)
  74
  75         if self.work_api is None:
  76             # todo: autodetect API to use.
  77             self.work_api = "jobs"
  78
  79         if self.work_api not in ("containers", "jobs"):
  80             raise Exception("Unsupported API '%s'" % self.work_api)
  81
  82     def arv_make_tool(self, toolpath_object, **kwargs):
  83         kwargs["work_api"] = self.work_api
  84         if "class" in toolpath_object and toolpath_object["class"] == "CommandLineTool":
  85             return ArvadosCommandTool(self, toolpath_object, **kwargs)
  86         elif "class" in toolpath_object and toolpath_object["class"] == "Workflow":
  87             return ArvadosWorkflow(self, toolpath_object, **kwargs)
  88         else:
  89             return cwltool.workflow.defaultMakeTool(toolpath_object, **kwargs)
  90
  91     def output_callback(self, out, processStatus):
  92         if processStatus == "success":
  93             logger.info("Overall process status is %s", processStatus)
  94             if self.pipeline:
  95                 self.api.pipeline_instances().update(uuid=self.pipeline["uuid"],
  96                                                      body={"state": "Complete"}).execute(num_retries=self.num_retries)
  97         else:
  98             logger.warn("Overall process status is %s", processStatus)
  99             if self.pipeline:
 100                 self.api.pipeline_instances().update(uuid=self.pipeline["uuid"],
 101                                                      body={"state": "Failed"}).execute(num_retries=self.num_retries)
 102         self.final_status = processStatus
 103         self.final_output = out
 104
 105     def on_message(self, event):
 106         if "object_uuid" in event:
 107             if event["object_uuid"] in self.processes and event["event_type"] == "update":
 108                 if event["properties"]["new_attributes"]["state"] == "Running" and self.processes[event["object_uuid"]].running is False:
 109                     uuid = event["object_uuid"]
 110                     with self.lock:
 111                         j = self.processes[uuid]
 112                         logger.info("Job %s (%s) is Running", j.name, uuid)
 113                         j.running = True
 114                         j.update_pipeline_component(event["properties"]["new_attributes"])
 115                 elif event["properties"]["new_attributes"]["state"] in ("Complete", "Failed", "Cancelled"):
 116                     uuid = event["object_uuid"]
 117                     try:
 118                         self.cond.acquire()
 119                         j = self.processes[uuid]
 120                         logger.info("Job %s (%s) is %s", j.name, uuid, event["properties"]["new_attributes"]["state"])
 121                         with Perf(metrics, "done %s" % j.name):
 122                             j.done(event["properties"]["new_attributes"])
 123                         self.cond.notify()
 124                     finally:
 125                         self.cond.release()
 126
 127     def poll_states(self):
 128         """Poll status of jobs or containers listed in the processes dict.
 129
 130         Runs in a separate thread.
 131         """
 132
 133         while True:
 134             self.stop_polling.wait(15)
 135             if self.stop_polling.is_set():
 136                 break
 137             with self.lock:
 138                 keys = self.processes.keys()
 139             if not keys:
 140                 continue
 141
 142             if self.work_api == "containers":
 143                 table = self.poll_api.containers()
 144             elif self.work_api == "jobs":
 145                 table = self.poll_api.jobs()
 146
 147             try:
 148                 proc_states = table.list(filters=[["uuid", "in", keys]]).execute(num_retries=self.num_retries)
 149             except Exception as e:
 150                 logger.warn("Error checking states on API server: %s", e)
 151                 continue
 152
 153             for p in proc_states["items"]:
 154                 self.on_message({
 155                     "object_uuid": p["uuid"],
 156                     "event_type": "update",
 157                     "properties": {
 158                         "new_attributes": p
 159                     }
 160                 })
 161
 162     def get_uploaded(self):
 163         return self.uploaded.copy()
 164
 165     def add_uploaded(self, src, pair):
 166         self.uploaded[src] = pair
 167
 168     def check_writable(self, obj):
 169         if isinstance(obj, dict):
 170             if obj.get("writable"):
 171                 raise UnsupportedRequirement("InitialWorkDir feature 'writable: true' not supported")
 172             for v in obj.itervalues():
 173                 self.check_writable(v)
 174         if isinstance(obj, list):
 175             for v in obj:
 176                 self.check_writable(v)
 177
 178     def make_output_collection(self, name, outputObj):
 179         outputObj = copy.deepcopy(outputObj)
 180
 181         files = []
 182         def capture(fileobj):
 183             files.append(fileobj)
 184
 185         adjustDirObjs(outputObj, capture)
 186         adjustFileObjs(outputObj, capture)
 187
 188         generatemapper = FinalOutputPathMapper(files, "", "", separateDirs=False)
 189
 190         final = arvados.collection.Collection(api_client=self.api,
 191                                               keep_client=self.keep_client,
 192                                               num_retries=self.num_retries)
 193
 194         srccollections = {}
 195         for k,v in generatemapper.items():
 196             sp = k.split("/")
 197             srccollection = sp[0][5:]
 198             if srccollection not in srccollections:
 199                 srccollections[srccollection] = arvados.collection.CollectionReader(
 200                     srccollection,
 201                     api_client=self.api,
 202                     keep_client=self.keep_client,
 203                     num_retries=self.num_retries)
 204             reader = srccollections[srccollection]
 205             try:
 206                 srcpath = "/".join(sp[1:]) if len(sp) > 1 else "."
 207                 final.copy(srcpath, v.target, source_collection=reader, overwrite=False)
 208             except IOError as e:
 209                 logger.warn("While preparing output collection: %s", e)
 210
 211         def rewrite(fileobj):
 212             fileobj["location"] = generatemapper.mapper(fileobj["location"]).target
 213             for k in ("basename", "size", "listing"):
 214                 if k in fileobj:
 215                     del fileobj[k]
 216
 217         adjustDirObjs(outputObj, rewrite)
 218         adjustFileObjs(outputObj, rewrite)
 219
 220         with final.open("cwl.output.json", "w") as f:
 221             json.dump(outputObj, f, sort_keys=True, indent=4, separators=(',',': '))
 222
 223         final.save_new(name=name, owner_uuid=self.project_uuid, ensure_unique_name=True)
 224
 225         logger.info("Final output collection %s \"%s\" (%s)", final.portable_data_hash(),
 226                     final.api_response()["name"],
 227                     final.manifest_locator())
 228
 229         self.final_output_collection = final
 230
 231     def arv_executor(self, tool, job_order, **kwargs):
 232         self.debug = kwargs.get("debug")
 233
 234         tool.visit(self.check_writable)
 235
 236         useruuid = self.api.users().current().execute()["uuid"]
 237         self.project_uuid = kwargs.get("project_uuid") if kwargs.get("project_uuid") else useruuid
 238         self.pipeline = None
 239         make_fs_access = kwargs.get("make_fs_access") or partial(CollectionFsAccess,
 240                                                                  api_client=self.api,
 241                                                                  keep_client=self.keep_client)
 242         self.fs_access = make_fs_access(kwargs["basedir"])
 243
 244         if kwargs.get("create_template"):
 245             tmpl = RunnerTemplate(self, tool, job_order, kwargs.get("enable_reuse"))
 246             tmpl.save()
 247             # cwltool.main will write our return value to stdout.
 248             return tmpl.uuid
 249
 250         if kwargs.get("create_workflow") or kwargs.get("update_workflow"):
 251             return upload_workflow(self, tool, job_order, self.project_uuid, kwargs.get("update_workflow"))
 252
 253         self.ignore_docker_for_reuse = kwargs.get("ignore_docker_for_reuse")
 254
 255         kwargs["make_fs_access"] = make_fs_access
 256         kwargs["enable_reuse"] = kwargs.get("enable_reuse")
 257         kwargs["use_container"] = True
 258         kwargs["tmpdir_prefix"] = "tmp"
 259         kwargs["on_error"] = "continue"
 260         kwargs["compute_checksum"] = kwargs.get("compute_checksum")
 261
 262         if self.work_api == "containers":
 263             kwargs["outdir"] = "/var/spool/cwl"
 264             kwargs["docker_outdir"] = "/var/spool/cwl"
 265             kwargs["tmpdir"] = "/tmp"
 266             kwargs["docker_tmpdir"] = "/tmp"
 267         elif self.work_api == "jobs":
 268             kwargs["outdir"] = "$(task.outdir)"
 269             kwargs["docker_outdir"] = "$(task.outdir)"
 270             kwargs["tmpdir"] = "$(task.tmpdir)"
 271
 272         upload_instance(self, shortname(tool.tool["id"]), tool, job_order)
 273
 274         runnerjob = None
 275         if kwargs.get("submit"):
 276             if self.work_api == "containers":
 277                 if tool.tool["class"] == "CommandLineTool":
 278                     runnerjob = tool.job(job_order,
 279                                          self.output_callback,
 280                                          **kwargs).next()
 281                 else:
 282                     runnerjob = RunnerContainer(self, tool, job_order, kwargs.get("enable_reuse"), self.output_name)
 283             else:
 284                 runnerjob = RunnerJob(self, tool, job_order, kwargs.get("enable_reuse"), self.output_name)
 285
 286         if not kwargs.get("submit") and "cwl_runner_job" not in kwargs and not self.work_api == "containers":
 287             # Create pipeline for local run
 288             self.pipeline = self.api.pipeline_instances().create(
 289                 body={
 290                     "owner_uuid": self.project_uuid,
 291                     "name": shortname(tool.tool["id"]),
 292                     "components": {},
 293                     "state": "RunningOnClient"}).execute(num_retries=self.num_retries)
 294             logger.info("Pipeline instance %s", self.pipeline["uuid"])
 295
 296         if runnerjob and not kwargs.get("wait"):
 297             runnerjob.run(wait=kwargs.get("wait"))
 298             return runnerjob.uuid
 299
 300         self.poll_api = arvados.api('v1')
 301         self.polling_thread = threading.Thread(target=self.poll_states)
 302         self.polling_thread.start()
 303
 304         if runnerjob:
 305             jobiter = iter((runnerjob,))
 306         else:
 307             if "cwl_runner_job" in kwargs:
 308                 self.uuid = kwargs.get("cwl_runner_job").get('uuid')
 309             jobiter = tool.job(job_order,
 310                                self.output_callback,
 311                                **kwargs)
 312
 313         try:
 314             self.cond.acquire()
 315             # Will continue to hold the lock for the duration of this code
 316             # except when in cond.wait(), at which point on_message can update
 317             # job state and process output callbacks.
 318
 319             loopperf = Perf(metrics, "jobiter")
 320             loopperf.__enter__()
 321             for runnable in jobiter:
 322                 loopperf.__exit__()
 323                 if runnable:
 324                     with Perf(metrics, "run"):
 325                         runnable.run(**kwargs)
 326                 else:
 327                     if self.processes:
 328                         self.cond.wait(1)
 329                     else:
 330                         logger.error("Workflow is deadlocked, no runnable jobs and not waiting on any pending jobs.")
 331                         break
 332                 loopperf.__enter__()
 333             loopperf.__exit__()
 334
 335             while self.processes:
 336                 self.cond.wait(1)
 337
 338         except UnsupportedRequirement:
 339             raise
 340         except:
 341             if sys.exc_info()[0] is KeyboardInterrupt:
 342                 logger.error("Interrupted, marking pipeline as failed")
 343             else:
 344                 logger.error("Caught unhandled exception, marking pipeline as failed.  Error was: %s", sys.exc_info()[1], exc_info=(sys.exc_info()[1] if self.debug else False))
 345             if self.pipeline:
 346                 self.api.pipeline_instances().update(uuid=self.pipeline["uuid"],
 347                                                      body={"state": "Failed"}).execute(num_retries=self.num_retries)
 348             if runnerjob and runnerjob.uuid and self.work_api == "containers":
 349                 self.api.container_requests().update(uuid=runnerjob.uuid,
 350                                                      body={"priority": "0"}).execute(num_retries=self.num_retries)
 351         finally:
 352             self.cond.release()
 353             self.stop_polling.set()
 354             self.polling_thread.join()
 355
 356         if self.final_status == "UnsupportedRequirement":
 357             raise UnsupportedRequirement("Check log for details.")
 358
 359         if self.final_status != "success":
 360             raise WorkflowException("Workflow failed.")
 361
 362         if self.final_output is None:
 363             raise WorkflowException("Workflow did not return a result.")
 364
 365         if kwargs.get("submit") and isinstance(runnerjob, Runner):
 366             logger.info("Final output collection %s", runnerjob.final_output)
 367         else:
 368             if self.output_name is None:
 369                 self.output_name = "Output of %s" % (shortname(tool.tool["id"]))
 370             self.make_output_collection(self.output_name, self.final_output)
 371
 372         if kwargs.get("compute_checksum"):
 373             adjustDirObjs(self.final_output, partial(getListing, self.fs_access))
 374             adjustFileObjs(self.final_output, partial(compute_checksums, self.fs_access))
 375
 376         return self.final_output
 377
 378
 379 def versionstring():
 380     """Print version string of key packages for provenance and debugging."""
 381
 382     arvcwlpkg = pkg_resources.require("arvados-cwl-runner")
 383     arvpkg = pkg_resources.require("arvados-python-client")
 384     cwlpkg = pkg_resources.require("cwltool")
 385
 386     return "%s %s %s, %s %s, %s %s" % (sys.argv[0], __version__, arvcwlpkg[0].version,
 387                                     "arvados-python-client", arvpkg[0].version,
 388                                     "cwltool", cwlpkg[0].version)
 389
 390
 391 def arg_parser():  # type: () -> argparse.ArgumentParser
 392     parser = argparse.ArgumentParser(description='Arvados executor for Common Workflow Language')
 393
 394     parser.add_argument("--basedir", type=str,
 395                         help="Base directory used to resolve relative references in the input, default to directory of input object file or current directory (if inputs piped/provided on command line).")
 396     parser.add_argument("--outdir", type=str, default=os.path.abspath('.'),
 397                         help="Output directory, default current directory")
 398
 399     parser.add_argument("--eval-timeout",
 400                         help="Time to wait for a Javascript expression to evaluate before giving an error, default 20s.",
 401                         type=float,
 402                         default=20)
 403     parser.add_argument("--version", action="store_true", help="Print version and exit")
 404
 405     exgroup = parser.add_mutually_exclusive_group()
 406     exgroup.add_argument("--verbose", action="store_true", help="Default logging")
 407     exgroup.add_argument("--quiet", action="store_true", help="Only print warnings and errors.")
 408     exgroup.add_argument("--debug", action="store_true", help="Print even more logging")
 409
 410     parser.add_argument("--metrics", action="store_true", help="Print timing metrics")
 411
 412     parser.add_argument("--tool-help", action="store_true", help="Print command line help for tool")
 413
 414     exgroup = parser.add_mutually_exclusive_group()
 415     exgroup.add_argument("--enable-reuse", action="store_true",
 416                         default=True, dest="enable_reuse",
 417                         help="")
 418     exgroup.add_argument("--disable-reuse", action="store_false",
 419                         default=True, dest="enable_reuse",
 420                         help="")
 421
 422     parser.add_argument("--project-uuid", type=str, metavar="UUID", help="Project that will own the workflow jobs, if not provided, will go to home project.")
 423     parser.add_argument("--output-name", type=str, help="Name to use for collection that stores the final output.", default=None)
 424     parser.add_argument("--ignore-docker-for-reuse", action="store_true",
 425                         help="Ignore Docker image version when deciding whether to reuse past jobs.",
 426                         default=False)
 427
 428     exgroup = parser.add_mutually_exclusive_group()
 429     exgroup.add_argument("--submit", action="store_true", help="Submit workflow to run on Arvados.",
 430                         default=True, dest="submit")
 431     exgroup.add_argument("--local", action="store_false", help="Run workflow on local host (submits jobs to Arvados).",
 432                         default=True, dest="submit")
 433     exgroup.add_argument("--create-template", action="store_true", help="Create an Arvados pipeline template.")
 434     exgroup.add_argument("--create-workflow", action="store_true", help="Create an Arvados workflow.")
 435     exgroup.add_argument("--update-workflow", type=str, metavar="UUID", help="Update existing Arvados workflow with uuid.")
 436
 437     exgroup = parser.add_mutually_exclusive_group()
 438     exgroup.add_argument("--wait", action="store_true", help="After submitting workflow runner job, wait for completion.",
 439                         default=True, dest="wait")
 440     exgroup.add_argument("--no-wait", action="store_false", help="Submit workflow runner job and exit.",
 441                         default=True, dest="wait")
 442
 443     parser.add_argument("--api", type=str,
 444                         default=None, dest="work_api",
 445                         help="Select work submission API, one of 'jobs' or 'containers'.")
 446
 447     parser.add_argument("--compute-checksum", action="store_true", default=False,
 448                         help="Compute checksum of contents while collecting outputs",
 449                         dest="compute_checksum")
 450
 451     parser.add_argument("workflow", type=str, nargs="?", default=None, help="The workflow to execute")
 452     parser.add_argument("job_order", nargs=argparse.REMAINDER, help="The input object to the workflow.")
 453
 454     return parser
 455
 456 def add_arv_hints():
 457     cache = {}
 458     res = pkg_resources.resource_stream(__name__, 'arv-cwl-schema.yml')
 459     cache["http://arvados.org/cwl"] = res.read()
 460     res.close()
 461     document_loader, cwlnames, _, _ = cwltool.process.get_schema("v1.0")
 462     _, extnames, _, _ = schema_salad.schema.load_schema("http://arvados.org/cwl", cache=cache)
 463     for n in extnames.names:
 464         if not cwlnames.has_name("http://arvados.org/cwl#"+n, ""):
 465             cwlnames.add_name("http://arvados.org/cwl#"+n, "", extnames.get_name(n, ""))
 466         document_loader.idx["http://arvados.org/cwl#"+n] = {}
 467
 468 def main(args, stdout, stderr, api_client=None, keep_client=None):
 469     parser = arg_parser()
 470
 471     job_order_object = None
 472     arvargs = parser.parse_args(args)
 473     if (arvargs.create_template or arvargs.create_workflow or arvargs.update_workflow) and not arvargs.job_order:
 474         job_order_object = ({}, "")
 475
 476     add_arv_hints()
 477
 478     try:
 479         if api_client is None:
 480             api_client=arvados.api('v1', model=OrderedJsonModel())
 481         runner = ArvCwlRunner(api_client, work_api=arvargs.work_api, keep_client=keep_client, output_name=arvargs.output_name)
 482     except Exception as e:
 483         logger.error(e)
 484         return 1
 485
 486     if arvargs.debug:
 487         logger.setLevel(logging.DEBUG)
 488
 489     if arvargs.quiet:
 490         logger.setLevel(logging.WARN)
 491         logging.getLogger('arvados.arv-run').setLevel(logging.WARN)
 492
 493     if arvargs.metrics:
 494         metrics.setLevel(logging.DEBUG)
 495         logging.getLogger("cwltool.metrics").setLevel(logging.DEBUG)
 496
 497     arvargs.conformance_test = None
 498     arvargs.use_container = True
 499
 500     return cwltool.main.main(args=arvargs,
 501                              stdout=stdout,
 502                              stderr=stderr,
 503                              executor=runner.arv_executor,
 504                              makeTool=runner.arv_make_tool,
 505                              versionfunc=versionstring,
 506                              job_order_object=job_order_object,
 507                              make_fs_access=partial(CollectionFsAccess, api_client=api_client))