X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/991d037e7b4741450dff745d22b0758170fe4ed9..5a46895af3657b549264a68bd2450baf8dbf7215:/sdk/python/arvados/keep.py diff --git a/sdk/python/arvados/keep.py b/sdk/python/arvados/keep.py index 7a9b5bfbee..44f10e4fb9 100644 --- a/sdk/python/arvados/keep.py +++ b/sdk/python/arvados/keep.py @@ -1,14 +1,22 @@ +# Copyright (C) The Arvados Authors. All rights reserved. +# +# SPDX-License-Identifier: Apache-2.0 + from __future__ import absolute_import from __future__ import division +import copy from future import standard_library +from future.utils import native_str standard_library.install_aliases() from builtins import next from builtins import str from builtins import range from builtins import object -import io +import collections import datetime import hashlib +import errno +import io import logging import math import os @@ -19,8 +27,10 @@ import socket import ssl import sys import threading +import resource from . import timer import urllib.parse +import traceback if sys.version_info >= (3, 0): from io import BytesIO @@ -32,6 +42,7 @@ import arvados.config as config import arvados.errors import arvados.retry as retry import arvados.util +import arvados.diskcache _logger = logging.getLogger('arvados.keep') global_client_object = None @@ -72,8 +83,9 @@ class KeepLocator(object): def __str__(self): return '+'.join( - str(s) for s in [self.md5sum, self.size, - self.permission_hint()] + self.hints + native_str(s) + for s in [self.md5sum, self.size, + self.permission_hint()] + self.hints if s is not None) def stripped(self): @@ -151,7 +163,6 @@ class Keep(object): config.get('ARVADOS_API_TOKEN'), config.flag_is_true('ARVADOS_API_HOST_INSECURE'), config.get('ARVADOS_KEEP_PROXY'), - config.get('ARVADOS_EXTERNAL_CLIENT') == 'true', os.environ.get('KEEP_LOCAL_STORE')) if (global_client_object is None) or (cls._last_key != key): global_client_object = KeepClient() @@ -167,11 +178,50 @@ class Keep(object): return Keep.global_client_object().put(data, **kwargs) class KeepBlockCache(object): - # Default RAM cache is 256MiB - def __init__(self, cache_max=(256 * 1024 * 1024)): + def __init__(self, cache_max=0, max_slots=0, disk_cache=False, disk_cache_dir=None): self.cache_max = cache_max self._cache = [] self._cache_lock = threading.Lock() + self._max_slots = max_slots + self._disk_cache = disk_cache + self._disk_cache_dir = disk_cache_dir + + if self._disk_cache and self._disk_cache_dir is None: + self._disk_cache_dir = os.path.join(os.path.expanduser("~"), ".cache", "arvados", "keep") + os.makedirs(self._disk_cache_dir, mode=0o700, exist_ok=True) + + if self._max_slots == 0: + if self._disk_cache: + # default max slots to half of maximum file handles + # NOFILE typically defaults to 1024 on Linux so this + # will be 512 slots. + self._max_slots = resource.getrlimit(resource.RLIMIT_NOFILE)[0] / 2 + else: + # RAM cache slots + self._max_slots = 512 + + if self.cache_max == 0: + if self._disk_cache: + fs = os.statvfs(self._disk_cache_dir) + # Calculation of available space incorporates existing cache usage + existing_usage = arvados.diskcache.DiskCacheSlot.cache_usage(self._disk_cache_dir) + avail = (fs.f_bavail * fs.f_bsize + existing_usage) / 4 + maxdisk = int((fs.f_blocks * fs.f_bsize) * 0.10) + # pick smallest of: + # 10% of total disk size + # 25% of available space + # max_slots * 64 MiB + self.cache_max = min(min(maxdisk, avail), (self._max_slots * 64 * 1024 * 1024)) + else: + # 256 MiB in RAM + self.cache_max = (256 * 1024 * 1024) + + self.cache_max = max(self.cache_max, 64 * 1024 * 1024) + + if self._disk_cache: + self._cache = arvados.diskcache.DiskCacheSlot.init_cache(self._disk_cache_dir, self._max_slots) + self.cap_cache() + class CacheSlot(object): __slots__ = ("locator", "ready", "content") @@ -195,6 +245,9 @@ class KeepBlockCache(object): else: return len(self.content) + def evict(self): + return True + def cap_cache(self): '''Cap the cache size to self.cache_max''' with self._cache_lock: @@ -202,12 +255,27 @@ class KeepBlockCache(object): # None (that means there was an error reading the block). self._cache = [c for c in self._cache if not (c.ready.is_set() and c.content is None)] sm = sum([slot.size() for slot in self._cache]) - while len(self._cache) > 0 and sm > self.cache_max: + while len(self._cache) > 0 and (sm > self.cache_max or len(self._cache) > self._max_slots): for i in range(len(self._cache)-1, -1, -1): + # start from the back, find a slot that is a candidate to evict if self._cache[i].ready.is_set(): + sz = self._cache[i].size() + + # If evict returns false it means the + # underlying disk cache couldn't lock the file + # for deletion because another process was using + # it. Don't count it as reducing the amount + # of data in the cache, find something else to + # throw out. + if self._cache[i].evict(): + sm -= sz + + # either way we forget about it. either the + # other process will delete it, or if we need + # it again and it is still there, we'll find + # it on disk. del self._cache[i] break - sm = sum([slot.size() for slot in self._cache]) def _get(self, locator): # Test if the locator is already in the cache @@ -219,6 +287,12 @@ class KeepBlockCache(object): del self._cache[i] self._cache.insert(0, n) return n + if self._disk_cache: + # see if it exists on disk + n = arvados.diskcache.DiskCacheSlot.get_from_disk(locator, self._disk_cache_dir) + if n is not None: + self._cache.insert(0, n) + return n return None def get(self, locator): @@ -234,10 +308,55 @@ class KeepBlockCache(object): return n, False else: # Add a new cache slot for the locator - n = KeepBlockCache.CacheSlot(locator) + if self._disk_cache: + n = arvados.diskcache.DiskCacheSlot(locator, self._disk_cache_dir) + else: + n = KeepBlockCache.CacheSlot(locator) self._cache.insert(0, n) return n, True + def set(self, slot, blob): + tryagain = False + + try: + slot.set(blob) + except OSError as e: + tryagain = True + if e.errno == errno.ENOMEM: + # Reduce max slots to current - 4, cap cache and retry + with self._cache_lock: + self._max_slots = max(4, len(self._cache) - 4) + elif e.errno == errno.ENOSPC: + # Reduce disk max space to current - 256 MiB, cap cache and retry + with self._cache_lock: + sm = sum([st.size() for st in self._cache]) + self.cache_max = max((256 * 1024 * 1024), sm - (256 * 1024 * 1024)) + elif e.errno == errno.ENODEV: + _logger.error("Unable to use disk cache: The underlying filesystem does not support memory mapping.") + except Exception as e: + tryagain = True + + # Check if we should evict things from the cache. Either + # because we added a new thing or we adjusted the limits down, + # so we might need to push something out. + self.cap_cache() + + if not tryagain: + # Done + return + + try: + # There was an error, we ran cap_cache so try one more time. + slot.set(blob) + except Exception as e: + # It failed again. Give up. + raise arvados.errors.KeepCacheError("Unable to save block %s to disk cache: %s" % (slot.locator, e)) + finally: + # Set the notice that that we are done with the cache + # slot one way or another. + slot.ready.set() + + class Counter(object): def __init__(self, v=0): self._lk = threading.Lock() @@ -284,17 +403,21 @@ class KeepClient(object): def __init__(self, root, user_agent_pool=queue.LifoQueue(), upload_counter=None, - download_counter=None, **headers): + download_counter=None, + headers={}, + insecure=False): self.root = root self._user_agent_pool = user_agent_pool self._result = {'error': None} self._usable = True self._session = None + self._socket = None self.get_headers = {'Accept': 'application/octet-stream'} self.get_headers.update(headers) self.put_headers = headers self.upload_counter = upload_counter self.download_counter = download_counter + self.insecure = insecure def usable(self): """Is it worth attempting a request?""" @@ -320,15 +443,28 @@ class KeepClient(object): except: ua.close() - @staticmethod - def _socket_open(family, socktype, protocol, address=None): + def _socket_open(self, *args, **kwargs): + if len(args) + len(kwargs) == 2: + return self._socket_open_pycurl_7_21_5(*args, **kwargs) + else: + return self._socket_open_pycurl_7_19_3(*args, **kwargs) + + def _socket_open_pycurl_7_19_3(self, family, socktype, protocol, address=None): + return self._socket_open_pycurl_7_21_5( + purpose=None, + address=collections.namedtuple( + 'Address', ['family', 'socktype', 'protocol', 'addr'], + )(family, socktype, protocol, address)) + + def _socket_open_pycurl_7_21_5(self, purpose, address): """Because pycurl doesn't have CURLOPT_TCP_KEEPALIVE""" - s = socket.socket(family, socktype, protocol) + s = socket.socket(address.family, address.socktype, address.protocol) s.setsockopt(socket.SOL_SOCKET, socket.SO_KEEPALIVE, 1) # Will throw invalid protocol error on mac. This test prevents that. if hasattr(socket, 'TCP_KEEPIDLE'): s.setsockopt(socket.IPPROTO_TCP, socket.TCP_KEEPIDLE, 75) s.setsockopt(socket.IPPROTO_TCP, socket.TCP_KEEPINTVL, 75) + self._socket = s return s def get(self, locator, method="GET", timeout=None): @@ -342,20 +478,30 @@ class KeepClient(object): self._headers = {} response_body = BytesIO() curl.setopt(pycurl.NOSIGNAL, 1) - curl.setopt(pycurl.OPENSOCKETFUNCTION, self._socket_open) + curl.setopt(pycurl.OPENSOCKETFUNCTION, + lambda *args, **kwargs: self._socket_open(*args, **kwargs)) curl.setopt(pycurl.URL, url.encode('utf-8')) curl.setopt(pycurl.HTTPHEADER, [ '{}: {}'.format(k,v) for k,v in self.get_headers.items()]) curl.setopt(pycurl.WRITEFUNCTION, response_body.write) curl.setopt(pycurl.HEADERFUNCTION, self._headerfunction) + if self.insecure: + curl.setopt(pycurl.SSL_VERIFYPEER, 0) + curl.setopt(pycurl.SSL_VERIFYHOST, 0) + else: + curl.setopt(pycurl.CAINFO, arvados.util.ca_certs_path()) if method == "HEAD": curl.setopt(pycurl.NOBODY, True) - self._setcurltimeouts(curl, timeout) + self._setcurltimeouts(curl, timeout, method=="HEAD") try: curl.perform() except Exception as e: raise arvados.errors.HttpError(0, str(e)) + finally: + if self._socket: + self._socket.close() + self._socket = None self._result = { 'status_code': curl.getinfo(pycurl.RESPONSE_CODE), 'body': response_body.getvalue(), @@ -390,6 +536,10 @@ class KeepClient(object): _logger.info("HEAD %s: %s bytes", self._result['status_code'], self._result.get('content-length')) + if self._result['headers'].get('x-keep-locator'): + # This is a response to a remote block copy request, return + # the local copy block locator. + return self._result['headers'].get('x-keep-locator') return True _logger.info("GET %s: %s bytes in %s msec (%.3f MiB/sec)", @@ -409,7 +559,9 @@ class KeepClient(object): return None return self._result['body'] - def put(self, hash_s, body, timeout=None): + def put(self, hash_s, body, timeout=None, headers={}): + put_headers = copy.copy(self.put_headers) + put_headers.update(headers) url = self.root + hash_s _logger.debug("Request: PUT %s", url) curl = self._get_user_agent() @@ -420,7 +572,8 @@ class KeepClient(object): body_reader = BytesIO(body) response_body = BytesIO() curl.setopt(pycurl.NOSIGNAL, 1) - curl.setopt(pycurl.OPENSOCKETFUNCTION, self._socket_open) + curl.setopt(pycurl.OPENSOCKETFUNCTION, + lambda *args, **kwargs: self._socket_open(*args, **kwargs)) curl.setopt(pycurl.URL, url.encode('utf-8')) # Using UPLOAD tells cURL to wait for a "go ahead" from the # Keep server (in the form of a HTTP/1.1 "100 Continue" @@ -432,14 +585,23 @@ class KeepClient(object): curl.setopt(pycurl.INFILESIZE, len(body)) curl.setopt(pycurl.READFUNCTION, body_reader.read) curl.setopt(pycurl.HTTPHEADER, [ - '{}: {}'.format(k,v) for k,v in self.put_headers.items()]) + '{}: {}'.format(k,v) for k,v in put_headers.items()]) curl.setopt(pycurl.WRITEFUNCTION, response_body.write) curl.setopt(pycurl.HEADERFUNCTION, self._headerfunction) + if self.insecure: + curl.setopt(pycurl.SSL_VERIFYPEER, 0) + curl.setopt(pycurl.SSL_VERIFYHOST, 0) + else: + curl.setopt(pycurl.CAINFO, arvados.util.ca_certs_path()) self._setcurltimeouts(curl, timeout) try: curl.perform() except Exception as e: raise arvados.errors.HttpError(0, str(e)) + finally: + if self._socket: + self._socket.close() + self._socket = None self._result = { 'status_code': curl.getinfo(pycurl.RESPONSE_CODE), 'body': response_body.getvalue().decode('utf-8'), @@ -474,7 +636,7 @@ class KeepClient(object): self.upload_counter.add(len(body)) return True - def _setcurltimeouts(self, curl, timeouts): + def _setcurltimeouts(self, curl, timeouts, ignore_bandwidth=False): if not timeouts: return elif isinstance(timeouts, tuple): @@ -487,8 +649,9 @@ class KeepClient(object): conn_t, xfer_t = (timeouts, timeouts) bandwidth_bps = KeepClient.DEFAULT_TIMEOUT[2] curl.setopt(pycurl.CONNECTTIMEOUT_MS, int(conn_t*1000)) - curl.setopt(pycurl.LOW_SPEED_TIME, int(math.ceil(xfer_t))) - curl.setopt(pycurl.LOW_SPEED_LIMIT, int(math.ceil(bandwidth_bps))) + if not ignore_bandwidth: + curl.setopt(pycurl.LOW_SPEED_TIME, int(math.ceil(xfer_t))) + curl.setopt(pycurl.LOW_SPEED_LIMIT, int(math.ceil(bandwidth_bps))) def _headerfunction(self, header_line): if isinstance(header_line, bytes): @@ -509,38 +672,68 @@ class KeepClient(object): self._lastheadername = name self._headers[name] = value # Returning None implies all bytes were written - + class KeepWriterQueue(queue.Queue): - def __init__(self, copies): + def __init__(self, copies, classes=[]): queue.Queue.__init__(self) # Old-style superclass self.wanted_copies = copies + self.wanted_storage_classes = classes self.successful_copies = 0 + self.confirmed_storage_classes = {} self.response = None - self.successful_copies_lock = threading.Lock() - self.pending_tries = copies + self.storage_classes_tracking = True + self.queue_data_lock = threading.RLock() + self.pending_tries = max(copies, len(classes)) self.pending_tries_notification = threading.Condition() - - def write_success(self, response, replicas_nr): - with self.successful_copies_lock: + + def write_success(self, response, replicas_nr, classes_confirmed): + with self.queue_data_lock: self.successful_copies += replicas_nr + if classes_confirmed is None: + self.storage_classes_tracking = False + elif self.storage_classes_tracking: + for st_class, st_copies in classes_confirmed.items(): + try: + self.confirmed_storage_classes[st_class] += st_copies + except KeyError: + self.confirmed_storage_classes[st_class] = st_copies + self.pending_tries = max(self.wanted_copies - self.successful_copies, len(self.pending_classes())) self.response = response with self.pending_tries_notification: self.pending_tries_notification.notify_all() - + def write_fail(self, ks): with self.pending_tries_notification: self.pending_tries += 1 self.pending_tries_notification.notify() - + def pending_copies(self): - with self.successful_copies_lock: + with self.queue_data_lock: return self.wanted_copies - self.successful_copies + def satisfied_classes(self): + with self.queue_data_lock: + if not self.storage_classes_tracking: + # Notifies disabled storage classes expectation to + # the outer loop. + return None + return list(set(self.wanted_storage_classes) - set(self.pending_classes())) + + def pending_classes(self): + with self.queue_data_lock: + if (not self.storage_classes_tracking) or (self.wanted_storage_classes is None): + return [] + unsatisfied_classes = copy.copy(self.wanted_storage_classes) + for st_class, st_copies in self.confirmed_storage_classes.items(): + if st_class in unsatisfied_classes and st_copies >= self.wanted_copies: + unsatisfied_classes.remove(st_class) + return unsatisfied_classes + def get_next_task(self): with self.pending_tries_notification: while True: - if self.pending_copies() < 1: + if self.pending_copies() < 1 and len(self.pending_classes()) == 0: # This notify_all() is unnecessary -- # write_success() already called notify_all() # when pending<1 became true, so it's not @@ -567,41 +760,40 @@ class KeepClient(object): class KeepWriterThreadPool(object): - def __init__(self, data, data_hash, copies, max_service_replicas, timeout=None): + def __init__(self, data, data_hash, copies, max_service_replicas, timeout=None, classes=[]): self.total_task_nr = 0 - self.wanted_copies = copies if (not max_service_replicas) or (max_service_replicas >= copies): num_threads = 1 else: num_threads = int(math.ceil(1.0*copies/max_service_replicas)) _logger.debug("Pool max threads is %d", num_threads) self.workers = [] - self.queue = KeepClient.KeepWriterQueue(copies) + self.queue = KeepClient.KeepWriterQueue(copies, classes) # Create workers for _ in range(num_threads): w = KeepClient.KeepWriterThread(self.queue, data, data_hash, timeout) self.workers.append(w) - + def add_task(self, ks, service_root): self.queue.put((ks, service_root)) self.total_task_nr += 1 - + def done(self): - return self.queue.successful_copies - + return self.queue.successful_copies, self.queue.satisfied_classes() + def join(self): # Start workers for worker in self.workers: worker.start() # Wait for finished work self.queue.join() - + def response(self): return self.queue.response - - + + class KeepWriterThread(threading.Thread): - TaskFailed = RuntimeError() + class TaskFailed(RuntimeError): pass def __init__(self, queue, data, data_hash, timeout=None): super(KeepClient.KeepWriterThread, self).__init__() @@ -618,29 +810,35 @@ class KeepClient(object): except queue.Empty: return try: - locator, copies = self.do_task(service, service_root) + locator, copies, classes = self.do_task(service, service_root) except Exception as e: - if e is not self.TaskFailed: + if not isinstance(e, self.TaskFailed): _logger.exception("Exception in KeepWriterThread") self.queue.write_fail(service) else: - self.queue.write_success(locator, copies) + self.queue.write_success(locator, copies, classes) finally: self.queue.task_done() def do_task(self, service, service_root): + classes = self.queue.pending_classes() + headers = {} + if len(classes) > 0: + classes.sort() + headers['X-Keep-Storage-Classes'] = ', '.join(classes) success = bool(service.put(self.data_hash, self.data, - timeout=self.timeout)) + timeout=self.timeout, + headers=headers)) result = service.last_result() if not success: - if result.get('status_code', None): + if result.get('status_code'): _logger.debug("Request fail: PUT %s => %s %s", self.data_hash, - result['status_code'], - result['body']) - raise self.TaskFailed + result.get('status_code'), + result.get('body')) + raise self.TaskFailed() _logger.debug("KeepWriterThread %s succeeded %s+%i %s", str(threading.current_thread()), @@ -652,7 +850,18 @@ class KeepClient(object): except (KeyError, ValueError): replicas_stored = 1 - return result['body'].strip(), replicas_stored + classes_confirmed = {} + try: + scch = result['headers']['x-keep-storage-classes-confirmed'] + for confirmation in scch.replace(' ', '').split(','): + if '=' in confirmation: + stored_class, stored_copies = confirmation.split('=')[:2] + classes_confirmed[stored_class] = int(stored_copies) + except (KeyError, ValueError): + # Storage classes confirmed header missing or corrupt + classes_confirmed = None + + return result['body'].strip(), replicas_stored, classes_confirmed def __init__(self, api_client=None, proxy=None, @@ -730,6 +939,11 @@ class KeepClient(object): if local_store is None: local_store = os.environ.get('KEEP_LOCAL_STORE') + if api_client is None: + self.insecure = config.flag_is_true('ARVADOS_API_HOST_INSECURE') + else: + self.insecure = api_client.insecure + self.block_cache = block_cache if block_cache else KeepBlockCache() self.timeout = timeout self.proxy_timeout = proxy_timeout @@ -740,9 +954,12 @@ class KeepClient(object): self.get_counter = Counter() self.hits_counter = Counter() self.misses_counter = Counter() + self._storage_classes_unsupported_warning = False + self._default_classes = [] if local_store: self.local_store = local_store + self.head = self.local_store_head self.get = self.local_store_get self.put = self.local_store_put else: @@ -779,6 +996,12 @@ class KeepClient(object): self._writable_services = None self.using_proxy = None self._static_services_list = False + try: + self._default_classes = [ + k for k, v in self.api_client.config()['StorageClasses'].items() if v['Default']] + except KeyError: + # We're talking to an old cluster + pass def current_timeout(self, attempt_number): """Return the appropriate timeout to use for this client. @@ -889,7 +1112,7 @@ class KeepClient(object): _logger.debug("{}: {}".format(locator, sorted_roots)) return sorted_roots - def map_new_services(self, roots_map, locator, force_rebuild, need_writable, **headers): + def map_new_services(self, roots_map, locator, force_rebuild, need_writable, headers): # roots_map is a dictionary, mapping Keep service root strings # to KeepService objects. Poll for Keep services, and add any # new ones to roots_map. Return the current list of local @@ -902,7 +1125,8 @@ class KeepClient(object): root, self._user_agent_pool, upload_counter=self.upload_counter, download_counter=self.download_counter, - **headers) + headers=headers, + insecure=self.insecure) return local_roots @staticmethod @@ -923,23 +1147,29 @@ class KeepClient(object): else: return None - def get_from_cache(self, loc): + def get_from_cache(self, loc_s): """Fetch a block only if is in the cache, otherwise return None.""" - slot = self.block_cache.get(loc) + locator = KeepLocator(loc_s) + slot = self.block_cache.get(locator.md5sum) if slot is not None and slot.ready.is_set(): return slot.get() else: return None + def refresh_signature(self, loc): + """Ask Keep to get the remote block and return its local signature""" + now = datetime.datetime.utcnow().isoformat("T") + 'Z' + return self.head(loc, headers={'X-Keep-Signature': 'local, {}'.format(now)}) + @retry.retry_method - def head(self, loc_s, num_retries=None): - return self._get_or_head(loc_s, method="HEAD", num_retries=num_retries) + def head(self, loc_s, **kwargs): + return self._get_or_head(loc_s, method="HEAD", **kwargs) @retry.retry_method - def get(self, loc_s, num_retries=None): - return self._get_or_head(loc_s, method="GET", num_retries=num_retries) + def get(self, loc_s, **kwargs): + return self._get_or_head(loc_s, method="GET", **kwargs) - def _get_or_head(self, loc_s, method="GET", num_retries=None): + def _get_or_head(self, loc_s, method="GET", num_retries=None, request_id=None, headers=None, prefetch=False): """Get data from Keep. This method fetches one or more blocks of data from Keep. It @@ -964,76 +1194,95 @@ class KeepClient(object): self.get_counter.add(1) - locator = KeepLocator(loc_s) - if method == "GET": - slot, first = self.block_cache.reserve_cache(locator.md5sum) - if not first: - self.hits_counter.add(1) - v = slot.get() - return v - - self.misses_counter.add(1) - - # If the locator has hints specifying a prefix (indicating a - # remote keepproxy) or the UUID of a local gateway service, - # read data from the indicated service(s) instead of the usual - # list of local disk services. - hint_roots = ['http://keep.{}.arvadosapi.com/'.format(hint[2:]) - for hint in locator.hints if hint.startswith('K@') and len(hint) == 7] - hint_roots.extend([self._gateway_services[hint[2:]]['_service_root'] - for hint in locator.hints if ( - hint.startswith('K@') and - len(hint) == 29 and - self._gateway_services.get(hint[2:]) - )]) - # Map root URLs to their KeepService objects. - roots_map = { - root: self.KeepService(root, self._user_agent_pool, - upload_counter=self.upload_counter, - download_counter=self.download_counter) - for root in hint_roots - } + request_id = (request_id or + (hasattr(self, 'api_client') and self.api_client.request_id) or + arvados.util.new_request_id()) + if headers is None: + headers = {} + headers['X-Request-Id'] = request_id - # See #3147 for a discussion of the loop implementation. Highlights: - # * Refresh the list of Keep services after each failure, in case - # it's being updated. - # * Retry until we succeed, we're out of retries, or every available - # service has returned permanent failure. - sorted_roots = [] - roots_map = {} + slot = None blob = None - loop = retry.RetryLoop(num_retries, self._check_loop_result, - backoff_start=2) - for tries_left in loop: - try: - sorted_roots = self.map_new_services( - roots_map, locator, - force_rebuild=(tries_left < num_retries), - need_writable=False) - except Exception as error: - loop.save_result(error) - continue + try: + locator = KeepLocator(loc_s) + if method == "GET": + slot, first = self.block_cache.reserve_cache(locator.md5sum) + if not first: + if prefetch: + # this is request for a prefetch, if it is + # already in flight, return immediately. + # clear 'slot' to prevent finally block from + # calling slot.set() + slot = None + return None + self.hits_counter.add(1) + blob = slot.get() + if blob is None: + raise arvados.errors.KeepReadError( + "failed to read {}".format(loc_s)) + return blob + + self.misses_counter.add(1) + + # If the locator has hints specifying a prefix (indicating a + # remote keepproxy) or the UUID of a local gateway service, + # read data from the indicated service(s) instead of the usual + # list of local disk services. + hint_roots = ['http://keep.{}.arvadosapi.com/'.format(hint[2:]) + for hint in locator.hints if hint.startswith('K@') and len(hint) == 7] + hint_roots.extend([self._gateway_services[hint[2:]]['_service_root'] + for hint in locator.hints if ( + hint.startswith('K@') and + len(hint) == 29 and + self._gateway_services.get(hint[2:]) + )]) + # Map root URLs to their KeepService objects. + roots_map = { + root: self.KeepService(root, self._user_agent_pool, + upload_counter=self.upload_counter, + download_counter=self.download_counter, + headers=headers, + insecure=self.insecure) + for root in hint_roots + } + + # See #3147 for a discussion of the loop implementation. Highlights: + # * Refresh the list of Keep services after each failure, in case + # it's being updated. + # * Retry until we succeed, we're out of retries, or every available + # service has returned permanent failure. + sorted_roots = [] + roots_map = {} + loop = retry.RetryLoop(num_retries, self._check_loop_result, + backoff_start=2) + for tries_left in loop: + try: + sorted_roots = self.map_new_services( + roots_map, locator, + force_rebuild=(tries_left < num_retries), + need_writable=False, + headers=headers) + except Exception as error: + loop.save_result(error) + continue - # Query KeepService objects that haven't returned - # permanent failure, in our specified shuffle order. - services_to_try = [roots_map[root] - for root in sorted_roots - if roots_map[root].usable()] - for keep_service in services_to_try: - blob = keep_service.get(locator, method=method, timeout=self.current_timeout(num_retries-tries_left)) - if blob is not None: - break - loop.save_result((blob, len(services_to_try))) - - # Always cache the result, then return it if we succeeded. - if method == "GET": - slot.set(blob) - self.block_cache.cap_cache() - if loop.success(): - if method == "HEAD": - return True - else: + # Query KeepService objects that haven't returned + # permanent failure, in our specified shuffle order. + services_to_try = [roots_map[root] + for root in sorted_roots + if roots_map[root].usable()] + for keep_service in services_to_try: + blob = keep_service.get(locator, method=method, timeout=self.current_timeout(num_retries-tries_left)) + if blob is not None: + break + loop.save_result((blob, len(services_to_try))) + + # Always cache the result, then return it if we succeeded. + if loop.success(): return blob + finally: + if slot is not None: + self.block_cache.set(slot, blob) # Q: Including 403 is necessary for the Keep tests to continue # passing, but maybe they should expect KeepReadError instead? @@ -1043,17 +1292,17 @@ class KeepClient(object): for key in sorted_roots) if not roots_map: raise arvados.errors.KeepReadError( - "failed to read {}: no Keep services available ({})".format( - loc_s, loop.last_result())) + "[{}] failed to read {}: no Keep services available ({})".format( + request_id, loc_s, loop.last_result())) elif not_founds == len(sorted_roots): raise arvados.errors.NotFoundError( - "{} not found".format(loc_s), service_errors) + "[{}] {} not found".format(request_id, loc_s), service_errors) else: raise arvados.errors.KeepReadError( - "failed to read {}".format(loc_s), service_errors, label="service") + "[{}] failed to read {} after {}".format(request_id, loc_s, loop.attempts_str()), service_errors, label="service") @retry.retry_method - def put(self, data, copies=2, num_retries=None): + def put(self, data, copies=2, num_retries=None, request_id=None, classes=None): """Save data in Keep. This method will get a list of Keep services from the API server, and @@ -1070,8 +1319,12 @@ class KeepClient(object): *each* Keep server if it returns temporary failures, with exponential backoff. The default value is set when the KeepClient is initialized. + * classes: An optional list of storage class names where copies should + be written. """ + classes = classes or self._default_classes + if not isinstance(data, bytes): data = data.encode() @@ -1083,51 +1336,78 @@ class KeepClient(object): return loc_s locator = KeepLocator(loc_s) - headers = {} - # Tell the proxy how many copies we want it to store - headers['X-Keep-Desired-Replicas'] = str(copies) + request_id = (request_id or + (hasattr(self, 'api_client') and self.api_client.request_id) or + arvados.util.new_request_id()) + headers = { + 'X-Request-Id': request_id, + 'X-Keep-Desired-Replicas': str(copies), + } roots_map = {} loop = retry.RetryLoop(num_retries, self._check_loop_result, backoff_start=2) - done = 0 + done_copies = 0 + done_classes = [] for tries_left in loop: try: sorted_roots = self.map_new_services( roots_map, locator, - force_rebuild=(tries_left < num_retries), need_writable=True, **headers) + force_rebuild=(tries_left < num_retries), + need_writable=True, + headers=headers) except Exception as error: loop.save_result(error) continue - writer_pool = KeepClient.KeepWriterThreadPool(data=data, + pending_classes = [] + if done_classes is not None: + pending_classes = list(set(classes) - set(done_classes)) + writer_pool = KeepClient.KeepWriterThreadPool(data=data, data_hash=data_hash, - copies=copies - done, + copies=copies - done_copies, max_service_replicas=self.max_replicas_per_service, - timeout=self.current_timeout(num_retries - tries_left)) + timeout=self.current_timeout(num_retries - tries_left), + classes=pending_classes) for service_root, ks in [(root, roots_map[root]) for root in sorted_roots]: if ks.finished(): continue writer_pool.add_task(ks, service_root) writer_pool.join() - done += writer_pool.done() - loop.save_result((done >= copies, writer_pool.total_task_nr)) + pool_copies, pool_classes = writer_pool.done() + done_copies += pool_copies + if (done_classes is not None) and (pool_classes is not None): + done_classes += pool_classes + loop.save_result( + (done_copies >= copies and set(done_classes) == set(classes), + writer_pool.total_task_nr)) + else: + # Old keepstore contacted without storage classes support: + # success is determined only by successful copies. + # + # Disable storage classes tracking from this point forward. + if not self._storage_classes_unsupported_warning: + self._storage_classes_unsupported_warning = True + _logger.warning("X-Keep-Storage-Classes header not supported by the cluster") + done_classes = None + loop.save_result( + (done_copies >= copies, writer_pool.total_task_nr)) if loop.success(): return writer_pool.response() if not roots_map: raise arvados.errors.KeepWriteError( - "failed to write {}: no Keep services available ({})".format( - data_hash, loop.last_result())) + "[{}] failed to write {}: no Keep services available ({})".format( + request_id, data_hash, loop.last_result())) else: service_errors = ((key, roots_map[key].last_result()['error']) for key in sorted_roots if roots_map[key].last_result()['error']) raise arvados.errors.KeepWriteError( - "failed to write {} (wanted {} copies but wrote {})".format( - data_hash, copies, writer_pool.done()), service_errors, label="service") + "[{}] failed to write {} after {} (wanted {} copies but wrote {})".format( + request_id, data_hash, loop.attempts_str(), (copies, classes), writer_pool.done()), service_errors, label="service") - def local_store_put(self, data, copies=1, num_retries=None): + def local_store_put(self, data, copies=1, num_retries=None, classes=[]): """A stub for put(). This method is used in place of the real put() method when @@ -1159,5 +1439,14 @@ class KeepClient(object): with open(os.path.join(self.local_store, locator.md5sum), 'rb') as f: return f.read() - def is_cached(self, locator): - return self.block_cache.reserve_cache(expect_hash) + def local_store_head(self, loc_s, num_retries=None): + """Companion to local_store_put().""" + try: + locator = KeepLocator(loc_s) + except ValueError: + raise arvados.errors.NotFoundError( + "Invalid data locator: '%s'" % loc_s) + if locator.md5sum == config.EMPTY_BLOCK_LOCATOR.split('+')[0]: + return True + if os.path.exists(os.path.join(self.local_store, locator.md5sum)): + return True