X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/af901a9a0434c9b5ed8b9349835a1749f45f95d3..c3f49b76d173818386f5c65db46b353e1d334d1e:/sdk/python/arvados/keep.py diff --git a/sdk/python/arvados/keep.py b/sdk/python/arvados/keep.py index 1b076b61ae..ec9f6f6422 100644 --- a/sdk/python/arvados/keep.py +++ b/sdk/python/arvados/keep.py @@ -1,28 +1,16 @@ -import bz2 +import cStringIO import datetime -import fcntl -import functools -import gflags import hashlib -import json import logging +import math import os -import pprint import pycurl import Queue import re import socket import ssl -import string -import cStringIO -import subprocess -import sys import threading -import time import timer -import types -import UserDict -import zlib import arvados import arvados.config as config @@ -232,58 +220,73 @@ class KeepClient(object): DEFAULT_PROXY_TIMEOUT = (20, 300) class ThreadLimiter(object): - """ - Limit the number of threads running at a given time to - {desired successes} minus {successes reported}. When successes - reported == desired, wake up the remaining threads and tell - them to quit. + """Limit the number of threads writing to Keep at once. + + This ensures that only a number of writer threads that could + potentially achieve the desired replication level run at once. + Once the desired replication level is achieved, queued threads + are instructed not to run. Should be used in a "with" block. """ - def __init__(self, todo): - self._todo = todo + def __init__(self, want_copies, max_service_replicas): + self._started = 0 + self._want_copies = want_copies self._done = 0 self._response = None - self._todo_lock = threading.Semaphore(todo) + self._start_lock = threading.Condition() + if (not max_service_replicas) or (max_service_replicas >= want_copies): + max_threads = 1 + else: + max_threads = math.ceil(float(want_copies) / max_service_replicas) + _logger.debug("Limiter max threads is %d", max_threads) + self._todo_lock = threading.Semaphore(max_threads) self._done_lock = threading.Lock() + self._local = threading.local() def __enter__(self): + self._start_lock.acquire() + if getattr(self._local, 'sequence', None) is not None: + # If the calling thread has used set_sequence(N), then + # we wait here until N other threads have started. + while self._started < self._local.sequence: + self._start_lock.wait() self._todo_lock.acquire() + self._started += 1 + self._start_lock.notifyAll() + self._start_lock.release() return self def __exit__(self, type, value, traceback): self._todo_lock.release() + def set_sequence(self, sequence): + self._local.sequence = sequence + def shall_i_proceed(self): """ - Return true if the current thread should do stuff. Return - false if the current thread should just stop. + Return true if the current thread should write to Keep. + Return false otherwise. """ with self._done_lock: - return (self._done < self._todo) + return (self._done < self._want_copies) def save_response(self, response_body, replicas_stored): """ Records a response body (a locator, possibly signed) returned by - the Keep server. It is not necessary to save more than - one response, since we presume that any locator returned - in response to a successful request is valid. + the Keep server, and the number of replicas it stored. """ with self._done_lock: self._done += replicas_stored self._response = response_body def response(self): - """ - Returns the body from the response to a PUT request. - """ + """Return the body from the response to a PUT request.""" with self._done_lock: return self._response def done(self): - """ - Return how many successes were reported. - """ + """Return the total number of replicas successfully stored.""" with self._done_lock: return self._done @@ -517,7 +520,11 @@ class KeepClient(object): return self._success def run(self): - with self.args['thread_limiter'] as limiter: + limiter = self.args['thread_limiter'] + sequence = self.args['thread_sequence'] + if sequence is not None: + limiter.set_sequence(sequence) + with limiter: if not limiter.shall_i_proceed(): # My turn arrived, but the job has been done without # me. @@ -639,6 +646,7 @@ class KeepClient(object): self.put = self.local_store_put else: self.num_retries = num_retries + self.max_replicas_per_service = None if proxy: if not proxy.endswith('/'): proxy += '/' @@ -646,12 +654,12 @@ class KeepClient(object): self._gateway_services = {} self._keep_services = [{ 'uuid': 'proxy', + 'service_type': 'proxy', '_service_root': proxy, }] self._writable_services = self._keep_services self.using_proxy = True self._static_services_list = True - self.max_replicas_per_service = None else: # It's important to avoid instantiating an API client # unless we actually need one, for testing's sake. @@ -664,7 +672,6 @@ class KeepClient(object): self._writable_services = None self.using_proxy = None self._static_services_list = False - self.max_replicas_per_service = None def current_timeout(self, attempt_number): """Return the appropriate timeout to use for this client. @@ -681,6 +688,10 @@ class KeepClient(object): t = self.proxy_timeout if self.using_proxy else self.timeout return (t[0] * (1 << attempt_number), t[1]) + def _any_nondisk_services(self, service_list): + return any(ks.get('service_type', 'disk') != 'disk' + for ks in service_list) + def build_services_list(self, force_rebuild=False): if (self._static_services_list or (self._keep_services and not force_rebuild)): @@ -691,12 +702,16 @@ class KeepClient(object): except Exception: # API server predates Keep services. keep_services = self.api_client.keep_disks().list() - accessible = keep_services.execute().get('items') - if not accessible: + # Gateway services are only used when specified by UUID, + # so there's nothing to gain by filtering them by + # service_type. + self._gateway_services = {ks['uuid']: ks for ks in + keep_services.execute()['items']} + if not self._gateway_services: raise arvados.errors.NoKeepServersError() # Precompute the base URI for each service. - for r in accessible: + for r in self._gateway_services.itervalues(): host = r['service_host'] if not host.startswith('[') and host.find(':') >= 0: # IPv6 URIs must be formatted like http://[::1]:80/... @@ -706,28 +721,19 @@ class KeepClient(object): host, r['service_port']) - # Gateway services are only used when specified by UUID, - # so there's nothing to gain by filtering them by - # service_type. - self._gateway_services = {ks.get('uuid'): ks for ks in accessible} _logger.debug(str(self._gateway_services)) - self._keep_services = [ - ks for ks in accessible - if ks.get('service_type') in ['disk', 'proxy']] - self._writable_services = [ - ks for ks in accessible - if (ks.get('service_type') in ['disk', 'proxy']) and (True != ks.get('read_only'))] - _logger.debug(str(self._keep_services)) - - self.using_proxy = any(ks.get('service_type') == 'proxy' - for ks in self._keep_services) - # Set max_replicas_per_service to 1 for disk typed services. - # In case of, non-disk typed services, we will use None to mean unknown. - self.max_replicas_per_service = 1 - for ks in accessible: - if ('disk' != ks.get('service_type')) and not ks.get('read_only'): - self.max_replicas_per_service = None + ks for ks in self._gateway_services.itervalues() + if not ks.get('service_type', '').startswith('gateway:')] + self._writable_services = [ks for ks in self._keep_services + if not ks.get('read_only')] + + # For disk type services, max_replicas_per_service is 1 + # It is unknown (unlimited) for other service types. + if self._any_nondisk_services(self._writable_services): + self.max_replicas_per_service = None + else: + self.max_replicas_per_service = 1 def _service_weight(self, data_hash, service_uuid): """Compute the weight of a Keep service endpoint for a data @@ -752,8 +758,8 @@ class KeepClient(object): for hint in locator.hints: if hint.startswith('K@'): if len(hint) == 7: - sorted_roots.append( - "https://keep.{}.arvadosapi.com/".format(hint[2:])) + sorted_roots.append( + "https://keep.{}.arvadosapi.com/".format(hint[2:])) elif len(hint) == 29: svc = self._gateway_services.get(hint[2:]) if svc: @@ -764,7 +770,8 @@ class KeepClient(object): # in that order. use_services = self._keep_services if need_writable: - use_services = self._writable_services + use_services = self._writable_services + self.using_proxy = self._any_nondisk_services(use_services) sorted_roots.extend([ svc['_service_root'] for svc in sorted( use_services, @@ -948,20 +955,22 @@ class KeepClient(object): # Tell the proxy how many copies we want it to store headers['X-Keep-Desired-Replication'] = str(copies) roots_map = {} - thread_limiter = KeepClient.ThreadLimiter(1 if self.max_replicas_per_service is None else copies) loop = retry.RetryLoop(num_retries, self._check_loop_result, backoff_start=2) for tries_left in loop: try: - local_roots = self.map_new_services( + sorted_roots = self.map_new_services( roots_map, locator, force_rebuild=(tries_left < num_retries), need_writable=True, **headers) except Exception as error: loop.save_result(error) continue + thread_limiter = KeepClient.ThreadLimiter( + copies, self.max_replicas_per_service) threads = [] - for service_root, ks in roots_map.iteritems(): + for service_root, ks in [(root, roots_map[root]) + for root in sorted_roots]: if ks.finished(): continue t = KeepClient.KeepWriterThread( @@ -970,7 +979,8 @@ class KeepClient(object): data_hash=data_hash, service_root=service_root, thread_limiter=thread_limiter, - timeout=self.current_timeout(num_retries-tries_left)) + timeout=self.current_timeout(num_retries-tries_left), + thread_sequence=len(threads)) t.start() threads.append(t) for t in threads: @@ -985,7 +995,7 @@ class KeepClient(object): data_hash, loop.last_result())) else: service_errors = ((key, roots_map[key].last_result()['error']) - for key in local_roots + for key in sorted_roots if roots_map[key].last_result()['error']) raise arvados.errors.KeepWriteError( "failed to write {} (wanted {} copies but wrote {})".format(