X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/72d7d41944006d1f48f570784dafe56b9812b0c8..650265966e83fca4ce8e9a416e9b5e358e82be98:/sdk/python/arvados/keep.py diff --git a/sdk/python/arvados/keep.py b/sdk/python/arvados/keep.py index 9dfe0436de..ce4c6f81f6 100644 --- a/sdk/python/arvados/keep.py +++ b/sdk/python/arvados/keep.py @@ -15,6 +15,7 @@ from builtins import object import collections import datetime import hashlib +import errno import io import logging import math @@ -26,8 +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 @@ -39,6 +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 @@ -159,7 +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() @@ -175,11 +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") @@ -203,6 +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: @@ -210,12 +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 @@ -227,6 +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): @@ -242,10 +308,49 @@ 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 + + try: + if tryagain: + # There was an error. Evict some slots and try again. + self.cap_cache() + 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() + + self.cap_cache() + class Counter(object): def __init__(self, v=0): self._lk = threading.Lock() @@ -376,6 +481,7 @@ class KeepClient(object): 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": @@ -478,6 +584,7 @@ class KeepClient(object): 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) @@ -720,11 +827,11 @@ class KeepClient(object): 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']) + result.get('status_code'), + result.get('body')) raise self.TaskFailed() _logger.debug("KeepWriterThread %s succeeded %s+%i %s", @@ -1034,9 +1141,10 @@ 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: @@ -1055,7 +1163,7 @@ class KeepClient(object): 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, request_id=None, headers=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 @@ -1080,6 +1188,13 @@ class KeepClient(object): self.get_counter.add(1) + 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 + slot = None blob = None try: @@ -1087,6 +1202,13 @@ class KeepClient(object): 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: @@ -1096,12 +1218,6 @@ class KeepClient(object): self.misses_counter.add(1) - if headers is None: - headers = {} - headers['X-Request-Id'] = (request_id or - (hasattr(self, 'api_client') and self.api_client.request_id) or - arvados.util.new_request_id()) - # 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 @@ -1160,8 +1276,7 @@ class KeepClient(object): return blob finally: if slot is not None: - slot.set(blob) - self.block_cache.cap_cache() + 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? @@ -1171,14 +1286,14 @@ 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 {} after {}".format(loc_s, loop.attempts_str()), 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, request_id=None, classes=None): @@ -1215,10 +1330,11 @@ class KeepClient(object): return loc_s locator = KeepLocator(loc_s) + 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 or - (hasattr(self, 'api_client') and self.api_client.request_id) or - arvados.util.new_request_id()), + 'X-Request-Id': request_id, 'X-Keep-Desired-Replicas': str(copies), } roots_map = {} @@ -1275,15 +1391,15 @@ class KeepClient(object): 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 {} after {} (wanted {} copies but wrote {})".format( - data_hash, loop.attempts_str(), (copies, classes), 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, classes=[]): """A stub for put(). @@ -1328,6 +1444,3 @@ class KeepClient(object): return True if os.path.exists(os.path.join(self.local_store, locator.md5sum)): return True - - def is_cached(self, locator): - return self.block_cache.reserve_cache(expect_hash)