# Copyright (C) 2022 Anaconda, Inc # Copyright (C) 2023 conda # SPDX-License-Identifier: BSD-3-Clause """ Models for sharded repodata, and to make monolithic repodata look like sharded repodata. """ from __future__ import annotations import abc import concurrent.futures import functools import json import logging from collections import defaultdict from pathlib import Path from typing import TYPE_CHECKING from urllib.parse import urljoin import conda.gateways.repodata import msgpack import zstandard from conda.base.context import context from conda.core.subdir_data import SubdirData from conda.gateways.connection.session import get_session from conda.gateways.repodata import ( _add_http_value_to_dict, conda_http_errors, ) from conda.models.channel import Channel from libmambapy.bindings import specs from requests import HTTPError from . import shards_cache log = logging.getLogger(__name__) if TYPE_CHECKING: from collections.abc import Iterable, KeysView from conda.gateways.repodata import RepodataCache from requests import Response from conda_libmamba_solver.shards_typing import RepodataDict, ShardsIndexDict from .shards_typing import PackageRecordDict, ShardDict SHARDS_CONNECTIONS_DEFAULT = 10 ZSTD_MAX_SHARD_SIZE = 2**20 * 16 # maximum size necessary when compressed data has no size header # For reference, the largest shard "conda-forge/linux-64/vim" is 2608283 bytes # or < 2**19*5 decompressed (486155 bytes compressed); the index is 575219 bytes # decompressed (514039 bytes compressed) and is mostly uncompressible hash data. def _shards_connections() -> int: """ If context.repodata_threads is not set, find the size of the connection pool in a typical https:// session. This should significantly reduce dropped connections. We match requests' default 10. Is this shared between all sessions? Or do we get a different pool for a different get_session(url)? Other adapters (file://, s3://) used in conda would have different concurrency behavior; we are not prepared to have separate threadpools per connection type. """ if context.repodata_threads is not None: return context.repodata_threads return SHARDS_CONNECTIONS_DEFAULT def ensure_hex_hash(record: PackageRecordDict): """ Convert bytes checksums to hex; leave unchanged if already str. """ for hash_type in "sha256", "md5": if hash_value := record.get(hash_type): if not isinstance(hash_value, str): record[hash_type] = bytes(hash_value).hex() return record @functools.cache def spec_to_package_name(spec: str) -> str: """ Given a dependency spec, return the package name. """ # Note: hope for no MatchSpec-without-name in repodata, although it is # possible in the MatchSpec grammar. parsed_spec = specs.MatchSpec.parse(spec) name = str(parsed_spec.name) return name def shard_mentioned_packages(shard: ShardDict) -> Iterable[str]: """ Return all dependency names mentioned in a shard, not including the shard's own package name. """ unique_specs = set() for package in (*shard["packages"].values(), *shard["packages.conda"].values()): ensure_hex_hash(package) # otherwise we could do this at serialization for spec in (*package.get("depends", ()),): # , *package.get("constrains", ())): if spec in unique_specs: continue unique_specs.add(spec) name = spec_to_package_name(spec) yield name # not much improvement from only yielding unique names class ShardBase(abc.ABC): """ Abstract base class for shard-like objects. Defines the common interface for both sharded repodata (Shards) and traditional repodata presented as shards (ShardLike). """ url: str repodata_no_packages: RepodataDict visited: dict[str, ShardDict | None] _base_url: str @property @abc.abstractmethod def package_names(self) -> KeysView[str]: """Return the names of all packages available in this shard collection.""" ... @property def base_url(self) -> str: """ Return self.url joined with base_url from repodata, or self.url if no base_url was present. Packages are found here. Note base_url can be a relative or an absolute url. The double urljoin ensures proper URL normalization: first join with _base_url, then with "." to add trailing slash if needed. """ return urljoin(urljoin(self.url, self._base_url), ".") def __contains__(self, package: str) -> bool: """Check if a package is available in this shard collection.""" return package in self.package_names @abc.abstractmethod def shard_url(self, package: str) -> str: """ Return shard URL for a given package. For monolithic repodata, should not be fetched but is a unique identifier. Raise KeyError if package is not in the index. """ ... @abc.abstractmethod def shard_loaded(self, package: str) -> bool: """ Return True if the given package's shard is in memory. """ ... def visit_package(self, package: str) -> ShardDict: """ Return a shard that is already loaded in memory and mark as visited. """ ... def visit_shard(self, package: str, shard: ShardDict): """ Store new shard data in the visited dict. """ self.visited[package] = shard @abc.abstractmethod def fetch_shard(self, package: str) -> ShardDict: """ Fetch an individual shard for the given package. """ ... @abc.abstractmethod def fetch_shards(self, packages: Iterable[str]) -> dict[str, ShardDict]: """ Fetch multiple shards in one go. """ ... def build_repodata(self) -> RepodataDict: """ Return monolithic repodata including all visited shards. """ repodata = self.repodata_no_packages.copy() repodata.update({"packages": {}, "packages.conda": {}}) for _, shard in self.visited.items(): if shard is None: continue # recorded visited but not available shards for package_group in ("packages", "packages.conda"): repodata[package_group].update(shard[package_group]) return repodata class ShardLike(ShardBase): """ Present a "classic" repodata.json as per-package shards. """ def __init__(self, repodata: RepodataDict, url: str = ""): """ url: must be unique for all ShardLike used together. """ self.repodata_no_packages: RepodataDict = { **repodata, "packages": {}, "packages.conda": {}, } all_packages = { "packages": repodata.get("packages", {}), "packages.conda": repodata.get("packages.conda", {}), } self.url = url shards = defaultdict(lambda: {"packages": {}, "packages.conda": {}}) for group_name, group in all_packages.items(): for package, record in group.items(): name = record["name"] shards[name][group_name][package] = record # defaultdict behavior no longer wanted self.shards: dict[str, ShardDict] = dict(shards) # type: ignore # used to write out repodata subset self.visited: dict[str, ShardDict | None] = {} # alternate location for packages, if not self.url try: base_url = self.repodata_no_packages["info"]["base_url"] if not isinstance(base_url, str): log.warning(f'repodata["info"]["base_url"] was not a str, got {type(base_url)}') raise TypeError() self._base_url = base_url except KeyError: self._base_url = "" def __repr__(self): left, right = super().__repr__().split(maxsplit=1) return f"{left} {self.url} {right}" @property def package_names(self) -> KeysView[str]: return self.shards.keys() def shard_url(self, package: str) -> str: """ Return shard URL for a given package. Raise KeyError if package is not in the index. """ self.shards[package] return f"{self.url}#{package}" def shard_loaded(self, package: str) -> bool: """ Return True if the given package's shard is in memory. """ return package in self.shards def visit_package(self, package: str) -> ShardDict: """ Return a shard that is already in memory and mark as visited. """ shard = self.fetch_shard(package) assert shard is not None return shard def fetch_shard(self, package: str) -> ShardDict: """ "Fetch" an individual shard. Update self.visited with all not-None packages. Raise KeyError if package is not in the index. """ shard = self.shards[package] self.visited[package] = shard return shard def fetch_shards(self, packages: Iterable[str]) -> dict[str, ShardDict]: """ Fetch multiple shards in one go. Update self.visited with all not-None packages. """ return {package: self.fetch_shard(package) for package in packages} def _shards_base_url(url, shards_base_url) -> str: """ Return shards_base_url joined with base_url and url. Note shards_base_url can be a relative or an absolute url. """ if shards_base_url and not shards_base_url.endswith("/"): shards_base_url += "/" return urljoin(urljoin(url, shards_base_url), ".") class Shards(ShardBase): """ Handle repodata_shards.msgpack.zst and individual per-package shards. """ # cache for shards_base_url() _shards_base_url = "" _shards_base_url_key = (None, None) def __init__(self, shards_index: ShardsIndexDict, url: str, cache: shards_cache.ShardCache): """ Args: shards_index: raw parsed msgpack dict url: URL of repodata_shards.msgpack.zst """ self.shards_index = shards_index self.url = url self.shards_cache = cache # Use the channel's base URL to share session amongst subdir locations channel_base_url = Channel(self.shards_base_url).base_url self.session = get_session(channel_base_url) self.repodata_no_packages = { "info": shards_index["info"], "packages": {}, "packages.conda": {}, "repodata_version": 2, } # used to write out repodata subset # not used in traversal algorithm self.visited: dict[str, ShardDict | None] = {} # https://github.com/conda/conda-index/pull/209 ensures that sharded # repodata will always include base_url, even if it is an empty string; # rattler/pixi require these keys. self._base_url = shards_index["info"]["base_url"] @property def package_names(self): return self.packages_index.keys() @property def packages_index(self): return self.shards_index["shards"] @property def shards_base_url(self) -> str: """ Return self.url joined with shards_base_url. Note shards_base_url can be a relative or an absolute url. """ # could be simplified by restricting self.shards_index assignment shards_base_url_ = self.shards_index["info"].get("shards_base_url", "") cache_key = (self.url, shards_base_url_) if self._shards_base_url_key != cache_key: self._shards_base_url_key = cache_key self._shards_base_url = _shards_base_url(self.url, shards_base_url_) return self._shards_base_url def shard_url(self, package: str) -> str: """ Return shard URL for a given package. Raise KeyError if package is not in the index. """ shard_name = f"{bytes(self.packages_index[package]).hex()}.msgpack.zst" # "Individual shards are stored under the URL .msgpack.zst" return f"{self.shards_base_url}{shard_name}" def shard_loaded(self, package: str) -> bool: """ Return True if the given package's shard is in memory. """ return package in self.visited def visit_package(self, package: str) -> ShardDict: """ Return a shard that is already in memory and mark as visited. """ shard = self.visited[package] return shard def fetch_shard(self, package: str) -> ShardDict: """ Fetch an individual shard for the given package. Default implementation calls fetch_shards() with a single package. Subclasses may override for more efficient single-fetch operations. Raise KeyError if package is not in the index. """ return self.fetch_shards([package])[package] def fetch_shards(self, packages: Iterable[str]) -> dict[str, ShardDict]: """ Return mapping of *package names* to Shard for given packages. If a shard is already in self.visited, it is not fetched again. """ results = {} def fetch(s, url, package_to_fetch): response = s.get(url) response.raise_for_status() data = response.content return shards_cache.AnnotatedRawShard( url=url, package=package_to_fetch, compressed_shard=data ) packages = sorted(list(packages)) urls_packages = {} # package shards to fetch for package in packages: if package in self.visited: results[package] = self.visited[package] else: urls_packages[self.shard_url(package)] = package with concurrent.futures.ThreadPoolExecutor(max_workers=_shards_connections()) as executor: futures = { executor.submit(fetch, self.session, url, package): (url, package) for url, package in urls_packages.items() if package not in results } for future in concurrent.futures.as_completed(futures): log.debug(". %s", futures[future]) url, package = futures[future] self._process_fetch_result(future, url, package, results) self.visited.update(results) return results def _process_fetch_result(self, future, url, package, results): """ Process a single fetched shard. """ with conda_http_errors(url, package): fetch_result = future.result() # Decompress and save record results[fetch_result.package] = msgpack.loads( zstandard.decompress( fetch_result.compressed_shard, max_output_size=ZSTD_MAX_SHARD_SIZE ) ) # Cache fetched shard self.shards_cache.insert(fetch_result) def repodata_shards(url, cache: RepodataCache) -> bytes: """ Fetch shards index with cache. Update cache state. Return shards data, either newly fetched or from cache. """ session = get_session(url) state = cache.state headers = {} etag = state.etag last_modified = state.mod if etag: headers["If-None-Match"] = str(etag) if last_modified: headers["If-Modified-Since"] = str(last_modified) filename = "repodata_shards.msgpack.zst" with conda_http_errors(url, filename): timeout = ( context.remote_connect_timeout_secs, context.remote_read_timeout_secs, ) response: Response = session.get( url, headers=headers, proxies=session.proxies, timeout=timeout ) response.raise_for_status() response_bytes = response.content if response.status_code == 304: # should we save cache-control to state here to put another n # seconds on the "make a remote request" clock and/or touch cache # mtime return cache.cache_path_shards.read_bytes() saved_fields = {conda.gateways.repodata.URL_KEY: url} for header, key in ( ("Etag", conda.gateways.repodata.ETAG_KEY), ( "Last-Modified", conda.gateways.repodata.LAST_MODIFIED_KEY, ), ("Cache-Control", conda.gateways.repodata.CACHE_CONTROL_KEY), ): _add_http_value_to_dict(response, header, saved_fields, key) state.update(saved_fields) # should we return the response and let caller save cache data to state? return response_bytes def fetch_shards_index( sd: SubdirData, cache: shards_cache.ShardCache | None = None ) -> Shards | None: """ Check a SubdirData's URL for shards. Return shards index bytes from cache or network. Return None if not found; caller should fetch normal repodata. TODO: If this function fails to retrieve the sharded repodata index file, it will mark it is as not supporting this feature in cache. This can problematic because sometimes server errors can happen which will lead it to wrongly assuming the channel doesn't support sharding. We need to rethink our logic for determining shard support. """ fetch = sd.repo_fetch repo_cache = fetch.repo_cache # cache.load_state() will clear the file on JSONDecodeError but cache.load() # will raise the exception. # repo_cache.load_state( # binary=True # ) # won't succeed when .msgpack.zst is missing as it wants to compare the timestamp (returns empty state) # Load state ourselves to avoid clearing when binary cached data is missing. # If we fall back to monolithic repodata.json, the standard fetch code will # load the state again in text mode. try: with repo_cache.lock("r+") as state_file: # cannot use pathlib.read_text / write_text on any locked file, as # it will release the lock early state = json.loads(state_file.read()) repo_cache.state.update(state) except (FileNotFoundError, json.JSONDecodeError): pass cache_state = repo_cache.state if cache is None: cache = shards_cache.ShardCache(Path(conda.gateways.repodata.create_cache_dir())) if cache_state.should_check_format("shards"): # look for shards index shards_data = None shards_index_url = f"{sd.url_w_subdir}/repodata_shards.msgpack.zst" if not repo_cache.cache_path_shards.exists(): # avoid 304 not modified if we don't have the file cache_state.etag = "" cache_state.mod = "" elif not repo_cache.stale(): # load from cache without network request shards_data = repo_cache.cache_path_shards.read_bytes() if shards_data is None: try: shards_data = repodata_shards(shards_index_url, repo_cache) cache_state.set_has_format("shards", True) # this will also set state["refresh_ns"] = time.time_ns(); we could # call cache.refresh() if we got a 304 instead: repo_cache.save(shards_data) except (HTTPError, conda.gateways.repodata.RepodataIsEmpty): # fetch repodata.json / repodata.json.zst instead cache_state.set_has_format("shards", False) repo_cache.refresh() if shards_data: # basic parse (move into caller?) shards_index: ShardsIndexDict = msgpack.loads( zstandard.decompress(shards_data, max_output_size=ZSTD_MAX_SHARD_SIZE) ) # type: ignore shards = Shards(shards_index, shards_index_url, cache) return shards return None def batch_retrieve_from_cache(sharded: list[Shards], packages: list[str]): """ Given a list of Shards objects and a list of package names, fetch all URLs from a shared local cache, and update Shards with those per-package shards. Return the remaining URLs that must be fetched from the network. """ sharded = [shardlike for shardlike in sharded if isinstance(shardlike, Shards)] wanted = [] # XXX update batch_retrieve_from_cache to work with (Shards, package name) # tuples instead of broadcasting across shards itself. for shard in sharded: for package_name in packages: if package_name in shard: # and not package_name in shard.visited wanted.append((shard, package_name, shard.shard_url(package_name))) log.debug("%d shards to fetch", len(wanted)) if not sharded: log.debug("No sharded channels found.") return wanted shared_shard_cache = sharded[0].shards_cache from_cache = shared_shard_cache.retrieve_multiple([shard_url for *_, shard_url in wanted]) # add fetched Shard objects to Shards objects visited dict for shard, package, shard_url in wanted: if from_cache_shard := from_cache.get(shard_url): shard.visit_shard(package, from_cache_shard) return wanted def batch_retrieve_from_network(wanted: list[tuple[Shards, str, str]]): """ Given a list of (Shards, package name, shard URL) tuples, group by Shards and call fetch_shards with a list of all URLs for that Shard. """ shard_packages: dict[Shards, list[str]] = defaultdict(list) for shard, package, _ in wanted: shard_packages[shard].append(package) # XXX it might be better to pull networking and Session() out of Shards(), # so that we can e.g. use the same session for a Channel(); typically a # noarch+arch pair of subdirs. # Could we share a ThreadPoolExecutor and see better session utilization? for shard, packages in shard_packages.items(): shard.fetch_shards(packages) def fetch_channels(channels: Iterable[Channel | str]) -> dict[str, ShardBase]: """ Return a dict mapping of a channel URL to a `Shard` or `ShardLike` object. Attempt to fetch the sharded index first and then fall back to retrieving a traditional `repodata.json` file. """ # metaclass returns same channel, or casts to channel. channels = [Channel(c) for c in channels] # type: ignore # Eliminate duplicates for example if this class is called with # channels=[Channel(f"{load_channel}/linux-64")], # subdirs=( # "noarch", # "linux-64", # ), url_to_channel = dict( (channel_url, Channel(channel_url)) for channel in channels for channel_url in channel.urls(True, context.subdirs) ) channel_data: dict[str, ShardBase] = {} # share single disk cache for all Shards() instances cache = shards_cache.ShardCache(Path(conda.gateways.repodata.create_cache_dir())) # The parallel version may reorder channels, does this matter? with concurrent.futures.ThreadPoolExecutor(max_workers=_shards_connections()) as executor: futures = { executor.submit(fetch_shards_index, SubdirData(channel), cache): channel_url for (channel_url, channel) in url_to_channel.items() } futures_non_sharded = {} for future in concurrent.futures.as_completed(futures): channel_url = futures[future] found = future.result() if found: channel_data[channel_url] = found else: futures_non_sharded[ executor.submit( SubdirData(Channel(channel_url)).repo_fetch.fetch_latest_parsed ) ] = channel_url # if all are None then don't do ShardLike... for future in concurrent.futures.as_completed(futures_non_sharded): channel_url = futures_non_sharded[future] repodata_json, _ = future.result() # the filename is not strictly repodata.json since we could have # fetched the same data from repodata.json.zst; but makes the # urljoin consistent with shards which end with # /repodata_shards.msgpack.zst url = f"{channel_url}/repodata.json" found = ShardLike(repodata_json, url) channel_data[channel_url] = found return channel_data