X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/e704e66bc35a0f0990620313ae9e2a630fa6821b..f2f8340b18430738a9527f05e707dd8f03508cc0:/sdk/python/arvados/keep.py?ds=sidebyside diff --git a/sdk/python/arvados/keep.py b/sdk/python/arvados/keep.py index ec9f6f6422..5b4770c4d0 100644 --- a/sdk/python/arvados/keep.py +++ b/sdk/python/arvados/keep.py @@ -1,4 +1,5 @@ import cStringIO +import collections import datetime import hashlib import logging @@ -9,8 +10,10 @@ import Queue import re import socket import ssl +import sys import threading import timer +import urlparse import arvados import arvados.config as config @@ -22,6 +25,17 @@ _logger = logging.getLogger('arvados.keep') global_client_object = None +# Monkey patch TCP constants when not available (apple). Values sourced from: +# http://www.opensource.apple.com/source/xnu/xnu-2422.115.4/bsd/netinet/tcp.h +if sys.platform == 'darwin': + if not hasattr(socket, 'TCP_KEEPALIVE'): + socket.TCP_KEEPALIVE = 0x010 + if not hasattr(socket, 'TCP_KEEPINTVL'): + socket.TCP_KEEPINTVL = 0x101 + if not hasattr(socket, 'TCP_KEEPCNT'): + socket.TCP_KEEPCNT = 0x102 + + class KeepLocator(object): EPOCH_DATETIME = datetime.datetime.utcfromtimestamp(0) HINT_RE = re.compile(r'^[A-Z][A-Za-z0-9@_-]+$') @@ -148,6 +162,8 @@ class KeepBlockCache(object): self._cache_lock = threading.Lock() class CacheSlot(object): + __slots__ = ("locator", "ready", "content") + def __init__(self, locator): self.locator = locator self.ready = threading.Event() @@ -210,85 +226,30 @@ class KeepBlockCache(object): self._cache.insert(0, n) return n, True -class KeepClient(object): - - # Default Keep server connection timeout: 2 seconds - # Default Keep server read timeout: 300 seconds - # Default Keep proxy connection timeout: 20 seconds - # Default Keep proxy read timeout: 300 seconds - DEFAULT_TIMEOUT = (2, 300) - DEFAULT_PROXY_TIMEOUT = (20, 300) +class Counter(object): + def __init__(self, v=0): + self._lk = threading.Lock() + self._val = v - class ThreadLimiter(object): - """Limit the number of threads writing to Keep at once. + def add(self, v): + with self._lk: + self._val += v - 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. + def get(self): + with self._lk: + return self._val - Should be used in a "with" block. - """ - def __init__(self, want_copies, max_service_replicas): - self._started = 0 - self._want_copies = want_copies - self._done = 0 - self._response = None - 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 write to Keep. - Return false otherwise. - """ - with self._done_lock: - 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, and the number of replicas it stored. - """ - with self._done_lock: - self._done += replicas_stored - self._response = response_body - def response(self): - """Return the body from the response to a PUT request.""" - with self._done_lock: - return self._response +class KeepClient(object): - def done(self): - """Return the total number of replicas successfully stored.""" - with self._done_lock: - return self._done + # Default Keep server connection timeout: 2 seconds + # Default Keep server read timeout: 256 seconds + # Default Keep server bandwidth minimum: 32768 bytes per second + # Default Keep proxy connection timeout: 20 seconds + # Default Keep proxy read timeout: 256 seconds + # Default Keep proxy bandwidth minimum: 32768 bytes per second + DEFAULT_TIMEOUT = (2, 256, 32768) + DEFAULT_PROXY_TIMEOUT = (20, 256, 32768) class KeepService(object): @@ -309,7 +270,9 @@ class KeepClient(object): arvados.errors.HttpError, ) - def __init__(self, root, user_agent_pool=Queue.LifoQueue(), **headers): + def __init__(self, root, user_agent_pool=Queue.LifoQueue(), + upload_counter=None, + download_counter=None, **headers): self.root = root self._user_agent_pool = user_agent_pool self._result = {'error': None} @@ -318,6 +281,8 @@ class KeepClient(object): 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 def usable(self): """Is it worth attempting a request?""" @@ -332,31 +297,46 @@ class KeepClient(object): def _get_user_agent(self): try: - return self._user_agent_pool.get(False) + return self._user_agent_pool.get(block=False) except Queue.Empty: return pycurl.Curl() def _put_user_agent(self, ua): try: ua.reset() - self._user_agent_pool.put(ua, False) + self._user_agent_pool.put(ua, block=False) 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) - s.setsockopt(socket.IPPROTO_TCP, socket.TCP_KEEPIDLE, 75) + # 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) return s - def get(self, locator, timeout=None): + def get(self, locator, method="GET", timeout=None): # locator is a KeepLocator object. url = self.root + str(locator) - _logger.debug("Request: GET %s", url) + _logger.debug("Request: %s %s", method, url) curl = self._get_user_agent() + ok = None try: with timer.Timer() as t: self._headers = {} @@ -368,7 +348,10 @@ class KeepClient(object): '{}: {}'.format(k,v) for k,v in self.get_headers.iteritems()]) curl.setopt(pycurl.WRITEFUNCTION, response_body.write) curl.setopt(pycurl.HEADERFUNCTION, self._headerfunction) + if method == "HEAD": + curl.setopt(pycurl.NOBODY, True) self._setcurltimeouts(curl, timeout) + try: curl.perform() except Exception as e: @@ -379,6 +362,7 @@ class KeepClient(object): 'headers': self._headers, 'error': False, } + ok = retry.check_http_response_success(self._result['status_code']) if not ok: self._result['error'] = arvados.errors.HttpError( @@ -388,7 +372,6 @@ class KeepClient(object): self._result = { 'error': e, } - ok = False self._usable = ok != False if self._result.get('status_code', None): # The client worked well enough to get an HTTP status @@ -403,11 +386,20 @@ class KeepClient(object): _logger.debug("Request fail: GET %s => %s: %s", url, type(self._result['error']), str(self._result['error'])) return None - _logger.info("%s response: %s bytes in %s msec (%.3f MiB/sec)", + if method == "HEAD": + _logger.info("HEAD %s: %s bytes", + self._result['status_code'], + self._result.get('content-length')) + return True + + _logger.info("GET %s: %s bytes in %s msec (%.3f MiB/sec)", self._result['status_code'], len(self._result['body']), t.msecs, (len(self._result['body'])/(1024.0*1024))/t.secs if t.secs > 0 else 0) + + if self.download_counter: + self.download_counter.add(len(self._result['body'])) resp_md5 = hashlib.md5(self._result['body']).hexdigest() if resp_md5 != locator.md5sum: _logger.warning("Checksum fail: md5(%s) = %s", @@ -421,37 +413,39 @@ class KeepClient(object): url = self.root + hash_s _logger.debug("Request: PUT %s", url) curl = self._get_user_agent() + ok = None try: - self._headers = {} - body_reader = cStringIO.StringIO(body) - response_body = cStringIO.StringIO() - curl.setopt(pycurl.NOSIGNAL, 1) - curl.setopt(pycurl.OPENSOCKETFUNCTION, self._socket_open) - 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" - # response) instead of sending the request body immediately. - # This allows the server to reject the request if the request - # is invalid or the server is read-only, without waiting for - # the client to send the entire block. - curl.setopt(pycurl.UPLOAD, True) - 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.iteritems()]) - curl.setopt(pycurl.WRITEFUNCTION, response_body.write) - curl.setopt(pycurl.HEADERFUNCTION, self._headerfunction) - self._setcurltimeouts(curl, timeout) - try: - curl.perform() - except Exception as e: - raise arvados.errors.HttpError(0, str(e)) - self._result = { - 'status_code': curl.getinfo(pycurl.RESPONSE_CODE), - 'body': response_body.getvalue(), - 'headers': self._headers, - 'error': False, - } + with timer.Timer() as t: + self._headers = {} + body_reader = cStringIO.StringIO(body) + response_body = cStringIO.StringIO() + curl.setopt(pycurl.NOSIGNAL, 1) + curl.setopt(pycurl.OPENSOCKETFUNCTION, self._socket_open) + 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" + # response) instead of sending the request body immediately. + # This allows the server to reject the request if the request + # is invalid or the server is read-only, without waiting for + # the client to send the entire block. + curl.setopt(pycurl.UPLOAD, True) + 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.iteritems()]) + curl.setopt(pycurl.WRITEFUNCTION, response_body.write) + curl.setopt(pycurl.HEADERFUNCTION, self._headerfunction) + self._setcurltimeouts(curl, timeout) + try: + curl.perform() + except Exception as e: + raise arvados.errors.HttpError(0, str(e)) + self._result = { + 'status_code': curl.getinfo(pycurl.RESPONSE_CODE), + 'body': response_body.getvalue(), + 'headers': self._headers, + 'error': False, + } ok = retry.check_http_response_success(self._result['status_code']) if not ok: self._result['error'] = arvados.errors.HttpError( @@ -461,7 +455,6 @@ class KeepClient(object): self._result = { 'error': e, } - ok = False self._usable = ok != False # still usable if ok is True or None if self._result.get('status_code', None): # Client is functional. See comment in get(). @@ -472,17 +465,30 @@ class KeepClient(object): _logger.debug("Request fail: PUT %s => %s: %s", url, type(self._result['error']), str(self._result['error'])) return False + _logger.info("PUT %s: %s bytes in %s msec (%.3f MiB/sec)", + self._result['status_code'], + len(body), + t.msecs, + (len(body)/(1024.0*1024))/t.secs if t.secs > 0 else 0) + if self.upload_counter: + self.upload_counter.add(len(body)) return True def _setcurltimeouts(self, curl, timeouts): if not timeouts: return elif isinstance(timeouts, tuple): - conn_t, xfer_t = timeouts + if len(timeouts) == 2: + conn_t, xfer_t = timeouts + bandwidth_bps = KeepClient.DEFAULT_TIMEOUT[2] + else: + conn_t, xfer_t, bandwidth_bps = timeouts else: conn_t, xfer_t = (timeouts, timeouts) + bandwidth_bps = KeepClient.DEFAULT_TIMEOUT[2] curl.setopt(pycurl.CONNECTTIMEOUT_MS, int(conn_t*1000)) - curl.setopt(pycurl.TIMEOUT_MS, int(xfer_t*1000)) + 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): header_line = header_line.decode('iso-8859-1') @@ -502,68 +508,150 @@ 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): + Queue.Queue.__init__(self) # Old-style superclass + self.wanted_copies = copies + self.successful_copies = 0 + self.response = None + self.successful_copies_lock = threading.Lock() + self.pending_tries = copies + self.pending_tries_notification = threading.Condition() + + def write_success(self, response, replicas_nr): + with self.successful_copies_lock: + self.successful_copies += replicas_nr + 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: + return self.wanted_copies - self.successful_copies + + def get_next_task(self): + with self.pending_tries_notification: + while True: + if self.pending_copies() < 1: + # This notify_all() is unnecessary -- + # write_success() already called notify_all() + # when pending<1 became true, so it's not + # possible for any other thread to be in + # wait() now -- but it's cheap insurance + # against deadlock so we do it anyway: + self.pending_tries_notification.notify_all() + # Drain the queue and then raise Queue.Empty + while True: + self.get_nowait() + self.task_done() + elif self.pending_tries > 0: + service, service_root = self.get_nowait() + if service.finished(): + self.task_done() + continue + self.pending_tries -= 1 + return service, service_root + elif self.empty(): + self.pending_tries_notification.notify_all() + raise Queue.Empty + else: + self.pending_tries_notification.wait() + + + class KeepWriterThreadPool(object): + def __init__(self, data, data_hash, copies, max_service_replicas, timeout=None): + 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(float(copies) / max_service_replicas)) + _logger.debug("Pool max threads is %d", num_threads) + self.workers = [] + self.queue = KeepClient.KeepWriterQueue(copies) + # 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 + + 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): - """ - Write a blob of data to the given Keep server. On success, call - save_response() of the given ThreadLimiter to save the returned - locator. - """ - def __init__(self, keep_service, **kwargs): - super(KeepClient.KeepWriterThread, self).__init__() - self.service = keep_service - self.args = kwargs - self._success = False + TaskFailed = RuntimeError() - def success(self): - return self._success + def __init__(self, queue, data, data_hash, timeout=None): + super(KeepClient.KeepWriterThread, self).__init__() + self.timeout = timeout + self.queue = queue + self.data = data + self.data_hash = data_hash + self.daemon = True def run(self): - 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. + while True: + try: + service, service_root = self.queue.get_next_task() + except Queue.Empty: return - self.run_with_limiter(limiter) - - def run_with_limiter(self, limiter): - if self.service.finished(): - return - _logger.debug("KeepWriterThread %s proceeding %s+%i %s", - str(threading.current_thread()), - self.args['data_hash'], - len(self.args['data']), - self.args['service_root']) - self._success = bool(self.service.put( - self.args['data_hash'], - self.args['data'], - timeout=self.args.get('timeout', None))) - result = self.service.last_result() - if self._success: - _logger.debug("KeepWriterThread %s succeeded %s+%i %s", - str(threading.current_thread()), - self.args['data_hash'], - len(self.args['data']), - self.args['service_root']) - # Tick the 'done' counter for the number of replica - # reported stored by the server, for the case that - # we're talking to a proxy or other backend that - # stores to multiple copies for us. try: - replicas_stored = int(result['headers']['x-keep-replicas-stored']) - except (KeyError, ValueError): - replicas_stored = 1 - limiter.save_response(result['body'].strip(), replicas_stored) - elif result.get('status_code', None): - _logger.debug("Request fail: PUT %s => %s %s", - self.args['data_hash'], - result['status_code'], - result['body']) + locator, copies = self.do_task(service, service_root) + except Exception as e: + if e is not self.TaskFailed: + _logger.exception("Exception in KeepWriterThread") + self.queue.write_fail(service) + else: + self.queue.write_success(locator, copies) + finally: + self.queue.task_done() + + def do_task(self, service, service_root): + success = bool(service.put(self.data_hash, + self.data, + timeout=self.timeout)) + result = service.last_result() + + if not success: + if result.get('status_code', None): + _logger.debug("Request fail: PUT %s => %s %s", + self.data_hash, + result['status_code'], + result['body']) + raise self.TaskFailed + + _logger.debug("KeepWriterThread %s succeeded %s+%i %s", + str(threading.current_thread()), + self.data_hash, + len(self.data), + service_root) + try: + replicas_stored = int(result['headers']['x-keep-replicas-stored']) + except (KeyError, ValueError): + replicas_stored = 1 + + return result['body'].strip(), replicas_stored def __init__(self, api_client=None, proxy=None, @@ -581,25 +669,28 @@ class KeepClient(object): :proxy: If specified, this KeepClient will send requests to this Keep proxy. Otherwise, KeepClient will fall back to the setting of the - ARVADOS_KEEP_PROXY configuration setting. If you want to ensure - KeepClient does not use a proxy, pass in an empty string. + ARVADOS_KEEP_SERVICES or ARVADOS_KEEP_PROXY configuration settings. + If you want to KeepClient does not use a proxy, pass in an empty + string. :timeout: The initial timeout (in seconds) for HTTP requests to Keep - non-proxy servers. A tuple of two floats is interpreted as - (connection_timeout, read_timeout): see - http://docs.python-requests.org/en/latest/user/advanced/#timeouts. - Because timeouts are often a result of transient server load, the - actual connection timeout will be increased by a factor of two on - each retry. - Default: (2, 300). + non-proxy servers. A tuple of three floats is interpreted as + (connection_timeout, read_timeout, minimum_bandwidth). A connection + will be aborted if the average traffic rate falls below + minimum_bandwidth bytes per second over an interval of read_timeout + seconds. Because timeouts are often a result of transient server + load, the actual connection timeout will be increased by a factor + of two on each retry. + Default: (2, 256, 32768). :proxy_timeout: The initial timeout (in seconds) for HTTP requests to - Keep proxies. A tuple of two floats is interpreted as - (connection_timeout, read_timeout). The behavior described - above for adjusting connection timeouts on retry also applies. - Default: (20, 300). + Keep proxies. A tuple of three floats is interpreted as + (connection_timeout, read_timeout, minimum_bandwidth). The behavior + described above for adjusting connection timeouts on retry also + applies. + Default: (20, 256, 32768). :api_token: If you're not using an API client, but only talking @@ -623,7 +714,10 @@ class KeepClient(object): """ self.lock = threading.Lock() if proxy is None: - proxy = config.get('ARVADOS_KEEP_PROXY') + if config.get('ARVADOS_KEEP_SERVICES'): + proxy = config.get('ARVADOS_KEEP_SERVICES') + else: + proxy = config.get('ARVADOS_KEEP_PROXY') if api_token is None: if api_client is None: api_token = config.get('ARVADOS_API_TOKEN') @@ -639,6 +733,12 @@ class KeepClient(object): self.timeout = timeout self.proxy_timeout = proxy_timeout self._user_agent_pool = Queue.LifoQueue() + self.upload_counter = Counter() + self.download_counter = Counter() + self.put_counter = Counter() + self.get_counter = Counter() + self.hits_counter = Counter() + self.misses_counter = Counter() if local_store: self.local_store = local_store @@ -648,15 +748,21 @@ class KeepClient(object): self.num_retries = num_retries self.max_replicas_per_service = None if proxy: - if not proxy.endswith('/'): - proxy += '/' + proxy_uris = proxy.split() + for i in range(len(proxy_uris)): + if not proxy_uris[i].endswith('/'): + proxy_uris[i] += '/' + # URL validation + url = urlparse.urlparse(proxy_uris[i]) + if not (url.scheme and url.netloc): + raise arvados.errors.ArgumentError("Invalid proxy URI: {}".format(proxy_uris[i])) self.api_token = api_token self._gateway_services = {} self._keep_services = [{ - 'uuid': 'proxy', + 'uuid': "00000-bi6l4-%015d" % idx, 'service_type': 'proxy', - '_service_root': proxy, - }] + '_service_root': uri, + } for idx, uri in enumerate(proxy_uris)] self._writable_services = self._keep_services self.using_proxy = True self._static_services_list = True @@ -686,8 +792,10 @@ class KeepClient(object): # TODO(twp): the timeout should be a property of a # KeepService, not a KeepClient. See #4488. t = self.proxy_timeout if self.using_proxy else self.timeout - return (t[0] * (1 << attempt_number), t[1]) - + if len(t) == 2: + return (t[0] * (1 << attempt_number), t[1]) + else: + return (t[0] * (1 << attempt_number), t[1], t[2]) def _any_nondisk_services(self, service_list): return any(ks.get('service_type', 'disk') != 'disk' for ks in service_list) @@ -790,7 +898,10 @@ class KeepClient(object): for root in local_roots: if root not in roots_map: roots_map[root] = self.KeepService( - root, self._user_agent_pool, **headers) + root, self._user_agent_pool, + upload_counter=self.upload_counter, + download_counter=self.download_counter, + **headers) return local_roots @staticmethod @@ -819,8 +930,15 @@ class KeepClient(object): else: return None + @retry.retry_method + def head(self, loc_s, num_retries=None): + return self._get_or_head(loc_s, method="HEAD", num_retries=num_retries) + @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_or_head(self, loc_s, method="GET", num_retries=None): """Get data from Keep. This method fetches one or more blocks of data from Keep. It @@ -842,11 +960,18 @@ class KeepClient(object): """ if ',' in loc_s: return ''.join(self.get(x) for x in loc_s.split(',')) + + self.get_counter.add(1) + locator = KeepLocator(loc_s) - slot, first = self.block_cache.reserve_cache(locator.md5sum) - if not first: - v = slot.get() - return v + 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, @@ -862,7 +987,9 @@ class KeepClient(object): )]) # Map root URLs to their KeepService objects. roots_map = { - root: self.KeepService(root, self._user_agent_pool) + root: self.KeepService(root, self._user_agent_pool, + upload_counter=self.upload_counter, + download_counter=self.download_counter) for root in hint_roots } @@ -892,16 +1019,20 @@ class KeepClient(object): for root in sorted_roots if roots_map[root].usable()] for keep_service in services_to_try: - blob = keep_service.get(locator, timeout=self.current_timeout(num_retries-tries_left)) + 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. - slot.set(blob) - self.block_cache.cap_cache() + if method == "GET": + slot.set(blob) + self.block_cache.cap_cache() if loop.success(): - return blob + if method == "HEAD": + return True + else: + return blob # Q: Including 403 is necessary for the Keep tests to continue # passing, but maybe they should expect KeepReadError instead? @@ -945,6 +1076,8 @@ class KeepClient(object): elif not isinstance(data, str): raise arvados.errors.ArgumentError("Argument 'data' to KeepClient.put is not type 'str'") + self.put_counter.add(1) + data_hash = hashlib.md5(data).hexdigest() loc_s = data_hash + '+' + str(len(data)) if copies < 1: @@ -953,10 +1086,11 @@ class KeepClient(object): headers = {} # Tell the proxy how many copies we want it to store - headers['X-Keep-Desired-Replication'] = str(copies) + headers['X-Keep-Desired-Replicas'] = str(copies) roots_map = {} loop = retry.RetryLoop(num_retries, self._check_loop_result, backoff_start=2) + done = 0 for tries_left in loop: try: sorted_roots = self.map_new_services( @@ -966,29 +1100,22 @@ class KeepClient(object): loop.save_result(error) continue - thread_limiter = KeepClient.ThreadLimiter( - copies, self.max_replicas_per_service) - threads = [] + writer_pool = KeepClient.KeepWriterThreadPool(data=data, + data_hash=data_hash, + copies=copies - done, + max_service_replicas=self.max_replicas_per_service, + timeout=self.current_timeout(num_retries - tries_left)) for service_root, ks in [(root, roots_map[root]) for root in sorted_roots]: if ks.finished(): continue - t = KeepClient.KeepWriterThread( - ks, - data=data, - data_hash=data_hash, - service_root=service_root, - thread_limiter=thread_limiter, - timeout=self.current_timeout(num_retries-tries_left), - thread_sequence=len(threads)) - t.start() - threads.append(t) - for t in threads: - t.join() - loop.save_result((thread_limiter.done() >= copies, len(threads))) + writer_pool.add_task(ks, service_root) + writer_pool.join() + done += writer_pool.done() + loop.save_result((done >= copies, writer_pool.total_task_nr)) if loop.success(): - return thread_limiter.response() + return writer_pool.response() if not roots_map: raise arvados.errors.KeepWriteError( "failed to write {}: no Keep services available ({})".format( @@ -999,7 +1126,7 @@ class KeepClient(object): if roots_map[key].last_result()['error']) raise arvados.errors.KeepWriteError( "failed to write {} (wanted {} copies but wrote {})".format( - data_hash, copies, thread_limiter.done()), service_errors, label="service") + data_hash, copies, writer_pool.done()), service_errors, label="service") def local_store_put(self, data, copies=1, num_retries=None): """A stub for put().