# Copyright (C) 2022 Anaconda, Inc # Copyright (C) 2023 conda # SPDX-License-Identifier: BSD-3-Clause """ Sharded repodata subsets. Traverse dependencies of installed and to-be-installed packages to generate a useful subset for the solver. The algorithm developed here is a direct result of the following CEP: - https://conda.org/learn/ceps/cep-0016 (Sharded Repodata) In this algorithm we treat a (channel, package name) as a node, its dependencies as edges. We then traverse all edges to discover all reachable (channel, package name) tuples. The solver should be able to find a solution with only this subset. This subset is overgenerous since the user is unlikely to want to install very old packages and their dependencies. If this is too slow, we could deploy heuristics that automatically ignore older package versions. We could also allow the user to configure minimum versions of common packages and ignore older versions and their dependencies, falling back to a full solve if unsatisfiable. We treat both sharded and monolithic repodata as if they were made up of per-package shards, computing a subset of both. This is because it is possible for the monolithic repodata to mention packages that exist in the true sharded repodata but would not be found by only traversing the shards. We treat all repodata as sharded, even if no actual sharded repodata has been found. ## Example usage The following constructs several repodata (`noarch` and `linux-64`) from a single channel name and a list of root packages: ``` from conda.models.channel import Channel from conda_libmamba_solver.shards_subset import build_repodata_subset channel = Channel("conda-forge-sharded/linux-64") channel_data = build_repodata_subset(["python", "pandas"], [channel.url()]) repodata = {} for url in channel_data: repodata[url] = channel_data.build_repodata() # ... this is what's fed to the solver ``` """ from __future__ import annotations import functools import logging import queue import sys import threading from collections import deque from concurrent.futures import Future, ThreadPoolExecutor from contextlib import suppress from dataclasses import dataclass from pathlib import Path from queue import SimpleQueue from typing import TYPE_CHECKING import conda.gateways.repodata import msgpack import zstandard from conda_libmamba_solver import shards_cache from conda_libmamba_solver.shards_cache import AnnotatedRawShard from .shards import ( ZSTD_MAX_SHARD_SIZE, Shards, _shards_connections, batch_retrieve_from_cache, batch_retrieve_from_network, fetch_channels, shard_mentioned_packages, ) log = logging.getLogger(__name__) if TYPE_CHECKING: from collections.abc import Iterable, Iterator, Sequence from queue import SimpleQueue as Queue from typing import Literal, TypeVar from conda.models.channel import Channel from conda_libmamba_solver.shards_cache import ShardCache from conda_libmamba_solver.shards_typing import ShardDict from .shards import ( ShardBase, ) # Waiting for worker threads to shutdown cleanly, or raise error. THREAD_WAIT_TIMEOUT = 5 # seconds REACHABLE_PIPELINED_MAX_TIMEOUTS = 10 # number of times we can timeout waiting for shards @dataclass(order=True) class Node: distance: int = sys.maxsize package: str = "" channel: str = "" visited: bool = False shard_url: str = "" def to_id(self) -> NodeId: return NodeId(self.package, self.channel, self.shard_url) @dataclass(order=True, eq=True, frozen=True) class NodeId: package: str channel: str shard_url: str = "" def __hash__(self): return hash((self.package, self.channel, self.shard_url)) def _nodes_from_packages( root_packages: list[str], shardlikes: Iterable[ShardBase] ) -> Iterator[tuple[NodeId, Node]]: """ Yield (NodeId, Node) for all root packages found in shardlikes. """ for package in root_packages: for shardlike in shardlikes: if package in shardlike: node = Node(0, package, shardlike.url, shard_url=shardlike.shard_url(package)) node_id = node.to_id() yield node_id, node @dataclass class RepodataSubset: nodes: dict[NodeId, Node] shardlikes: Sequence[ShardBase] DEFAULT_STRATEGY = "pipelined" def __init__(self, shardlikes: Iterable[ShardBase]): self.nodes = {} self.shardlikes = list(shardlikes) @classmethod def has_strategy(cls, strategy: str) -> bool: """ Return True if this class provides the named shard traversal strategy. """ return hasattr(cls, f"reachable_{strategy}") def neighbors(self, node: Node) -> Iterator[Node]: """ Retrieve all unvisited neighbors of a node Neighbors in the context are dependencies of a package """ discovered = set() for shardlike in self.shardlikes: if node.package not in shardlike: continue # check that we don't fetch the same shard twice... shard = shardlike.fetch_shard( node.package ) # XXX this is the only place that in-memory (repodata.json) shards are found for the first time for package in shard_mentioned_packages(shard): node_id = NodeId(package, shardlike.url) if node_id not in self.nodes: self.nodes[node_id] = Node(node.distance + 1, package, shardlike.url) yield self.nodes[node_id] if package not in discovered: # now this is per package name, not per (name, channel) tuple discovered.add(package) def outgoing(self, node: Node): """ All nodes that can be reached by this node, plus cost. """ # If we set a greater cost for sharded repodata than the repodata that # is already in memory and tracked nodes as (channel, package) tuples, # we might be able to find more shards-to-fetch-in-parallel more # quickly. On the other hand our goal is that the big channels will all # be sharded. for n in self.neighbors(node): yield n, 1 def reachable_bfs(self, root_packages): """ Fetch all packages reachable from `root_packages`' by following dependencies using the "breadth-first search" algorithm. Update associated `self.shardlikes` to contain enough data to build a repodata subset. """ self.nodes = dict(_nodes_from_packages(root_packages, self.shardlikes)) node_queue = deque(self.nodes.values()) sharded = [s for s in self.shardlikes if isinstance(s, Shards)] while node_queue: # Batch fetch all nodes at current level to_retrieve = {node.package for node in node_queue if not node.visited} if to_retrieve: not_in_cache = batch_retrieve_from_cache(sharded, sorted(to_retrieve)) batch_retrieve_from_network(not_in_cache) # Process one level level_size = len(node_queue) for _ in range(level_size): node = node_queue.popleft() if node.visited: # pragma: no cover continue # we should never add visited nodes to node_queue node.visited = True for next_node, _ in self.outgoing(node): if not next_node.visited: node_queue.append(next_node) def reachable_pipelined(self, root_packages): """ Fetch all packages reachable from `root_packages`' by following dependencies. Build repodata subset using concurrent threads to follow dependencies, fetch from cache, and fetch from network. """ # Ignore cache on shards object, use our own. Necessary if there are no # sharded channels. cache = shards_cache.ShardCache(Path(conda.gateways.repodata.create_cache_dir())) cache_in_queue: SimpleQueue[list[NodeId] | None] = SimpleQueue() shard_out_queue: SimpleQueue[list[tuple[NodeId, ShardDict]] | Exception] = SimpleQueue() cache_miss_queue: SimpleQueue[list[NodeId] | None] = SimpleQueue() cache_thread = threading.Thread( target=cache_fetch_thread, args=(cache_in_queue, shard_out_queue, cache_miss_queue, cache), daemon=True, # may have to set to False if we ever want to run in a subinterpreter ) network_thread = threading.Thread( target=network_fetch_thread, args=(cache_miss_queue, shard_out_queue, cache, self.shardlikes), daemon=True, ) try: cache_thread.start() network_thread.start() self.pipelined_main_thread( root_packages, cache_in_queue, shard_out_queue, cache_thread, network_thread ) finally: cache_in_queue.put(None) # These should finish almost immediately, but if not, raise an error: cache_thread.join(THREAD_WAIT_TIMEOUT) network_thread.join(THREAD_WAIT_TIMEOUT) def pipelined_main_thread( self, root_packages, cache_in_queue, shard_out_queue, cache_thread, network_thread ): """ Run reachibility algorithm given queues to submit and receive shards. """ shardlikes_by_url = {s.url: s for s in self.shardlikes} pending: set[NodeId] = set() in_flight: set[NodeId] = set() timeouts = 0 shutdown_initiated = False self.nodes = {} # create start condition parent_node = Node(0) self.visit_node(pending, parent_node, root_packages) def pump(): """ Find shards we already have and those we need. Submit those need to cache_in_queue, those we have to shard_out_queue. """ have, need = self.drain_pending(pending, shardlikes_by_url) if need: in_flight.update(need) cache_in_queue.put(need) if have: in_flight.update(node_id for node_id, _ in have) shard_out_queue.put(have) return len(have) + len(need) running = True while running: pump() try: new_shards = shard_out_queue.get(timeout=1) if new_shards is None: running = False continue # or break if isinstance(new_shards, Exception): # error propagated from worker thread raise new_shards except queue.Empty: pump_count = pump() log.debug("Shard timeout %s, pump_count=%d", timeouts, pump_count) log.debug("pending: %s...", sorted(str(node_id) for node_id in pending)[:10]) log.debug("in_flight: %s...", sorted(str(node_id) for node_id in in_flight)[:10]) log.debug("nodes: %d", len(self.nodes)) log.debug("cache_thread.is_alive(): %s", cache_thread.is_alive()) log.debug("network_thread.is_alive(): %s", network_thread.is_alive()) log.debug("shard_out_queue.qsize(): %s", shard_out_queue.qsize()) if not pending and not in_flight: log.debug("All shards have finished processing") break timeouts += 1 if timeouts > REACHABLE_PIPELINED_MAX_TIMEOUTS: raise TimeoutError( f"Timeout waiting for shard_out_queue after {timeouts} attempts. " f"pending={len(pending)}, in_flight={len(in_flight)}, " f"cache_thread_alive={cache_thread.is_alive()}, " f"network_thread_alive={network_thread.is_alive()}" ) continue # immediately calls pump() at top of loop for node_id, shard in new_shards: in_flight.remove(node_id) # add shard to appropriate ShardLike parent_node = self.nodes[node_id] shardlike = shardlikes_by_url[node_id.channel] shardlike.visit_shard(node_id.package, shard) self.visit_node(pending, parent_node, shard_mentioned_packages(shard)) if not pending and not in_flight and not shutdown_initiated: log.debug("Initiating shutdown: sending None to cache_in_queue") cache_in_queue.put(None) shutdown_initiated = True def visit_node( self, pending: set[NodeId], parent_node: Node, mentioned_packages: Iterable[str] ): """Broadcast mentioned packages across channels to pending.""" # NOTE we have visit for Nodes which is used in the graph traversal # algorithm, and a separate visit for ShardBase which means "include # this package in the output repodata". for package in mentioned_packages: for shardlike in self.shardlikes: if package in shardlike: new_node_id = NodeId(package, shardlike.url, shardlike.shard_url(package)) if new_node_id not in self.nodes: new_node = Node( distance=parent_node.distance + 1, package=new_node_id.package, channel=new_node_id.channel, shard_url=new_node_id.shard_url, ) self.nodes[new_node_id] = new_node pending.add(new_node_id) parent_node.visited = True def drain_pending( self, pending: set[NodeId], shardlikes_by_url: dict[str, ShardBase] ) -> tuple[list[tuple[NodeId, ShardDict]], list[NodeId]]: """ Check pending for in-memory shards. Clear pending. Return a list of shards we have and shards we need to fetch. """ shards_need = [] shards_have = [] for node_id in pending: # we should already have these nodes. shardlike = shardlikes_by_url[node_id.channel] if shardlike.shard_loaded(node_id.package): # for monolithic repodata shards_have.append((node_id, shardlike.visit_package(node_id.package))) else: if self.nodes[node_id].visited: # pragma: no cover log.debug("Skip visited, should not be reached") continue shards_need.append(node_id) pending.clear() return shards_have, shards_need def build_repodata_subset( root_packages: Iterable[str], channels: dict[str, Channel], algorithm: Literal["bfs", "pipelined"] = RepodataSubset.DEFAULT_STRATEGY, ) -> dict[str, ShardBase]: """ Retrieve all necessary information to build a repodata subset. Params: root_packages: iterable of installed and requested package names channels: iterable of Channel objects algorithm: desired traversal algorithm """ if isinstance(channels, dict): # True when called by LibMambaIndexHelper channels_: list[Channel] = list(channels.values()) else: channels_ = channels channel_data = fetch_channels(channels_) subset = RepodataSubset((*channel_data.values(),)) getattr(subset, f"reachable_{algorithm}")(root_packages) log.debug("%d (channel, package) nodes discovered", len(subset.nodes)) return channel_data # region workers if TYPE_CHECKING: _T = TypeVar("_T") def combine_batches_until_none( in_queue: Queue[Sequence[_T] | None], ) -> Iterator[Sequence[_T]]: """ Combine lists from in_queue until we see None. Yield combined lists. """ running = True while running: try: # Add timeout to prevent indefinite blocking if producer thread fails batch = in_queue.get(timeout=5) if batch is None: break except queue.Empty: # If we timeout, continue waiting - producer might still send data continue node_ids = list(batch) with suppress(queue.Empty): while True: # loop exits with break or queue.Empty exception batch = in_queue.get_nowait() if batch is None: # do the work but then quit running = False break else: node_ids.extend(batch) yield node_ids def exception_to_queue(func): """ Decorator to send unhandled exceptions to the second argument out_queue. """ @functools.wraps(func) def wrapper(in_queue, out_queue, *args, **kwargs): try: return func(in_queue, out_queue, *args, **kwargs) except Exception as e: in_queue.put(None) # signal termination out_queue.put(e) return wrapper @exception_to_queue def cache_fetch_thread( in_queue: Queue[Sequence[NodeId] | None], shard_out_queue: Queue[Sequence[tuple[NodeId, ShardDict] | Exception] | None], network_out_queue: Queue[Sequence[NodeId] | None], cache: ShardCache, ): """ Fetch batches of shards from cache until in_queue sees None. Enqueue found shards to shard_out_queue, and not found shards to network_out_queue. When we see None on in_queue, send None to both out queues and exit. Args: in_queue: NodeId (URLs) to fetch. shard_out_queue: fetched shards sent to queue. network_out_queue: cache misses forwarded to queue. Same queue is network_fetch_thread's in_queue. cache: used to retrieve shards. """ cache = cache.copy() for node_ids in combine_batches_until_none(in_queue): cached = cache.retrieve_multiple([node_id.shard_url for node_id in node_ids]) # should we add this into retrieve_multiple? found: list[tuple[NodeId, ShardDict]] = [] not_found: list[NodeId] = [] for node_id in node_ids: if shard := cached.get(node_id.shard_url): found.append((node_id, shard)) else: not_found.append(node_id) # Might wake up the network thread by calling it first: if not_found: network_out_queue.put(not_found) if found: shard_out_queue.put(found) network_out_queue.put(None) shard_out_queue.put(None) @exception_to_queue def network_fetch_thread( in_queue: Queue[Sequence[NodeId | Future] | None], shard_out_queue: Queue[list[tuple[NodeId, ShardDict] | Exception] | None], cache: ShardCache, shardlikes: list[ShardBase], ): """ Fetch shards from the network that are received on in_queue, until we see None. Unhandled exceptions also go to shard_out_queue, and exit this thread. Args: in_queue: NodeId (URLs) to fetch. shard_out_queue: fetched shards sent to queue. cache: once shards are decoded they are stored in cache. shardlikes: list of (network-only) shard index objects. """ cache = cache.copy() dctx = zstandard.ZstdDecompressor(max_window_size=ZSTD_MAX_SHARD_SIZE) shardlikes_by_url = {s.url: s for s in shardlikes} def fetch(s, url: str, node_id: NodeId): response = s.get(url) response.raise_for_status() data = response.content return url, node_id, data def submit(node_id: NodeId): # this worker should only receive network node_id's: shardlike = shardlikes_by_url[node_id.channel] if not isinstance(shardlike, Shards): raise TypeError("network_fetch_thread got non-network shardlike") session = shardlike.session url = shardlikes_by_url[node_id.channel].shard_url(node_id.package) return executor.submit(fetch, session, url, node_id) def handle_result(future: Future): url, node_id, data = future.result() log.debug("Fetch thread got %s (%s bytes)", url, len(data)) # Decompress and parse. If it decodes as # msgpack.zst, insert into cache. Then put "known # good" shard into out queue. shard: ShardDict = msgpack.loads( dctx.decompress(data, max_output_size=ZSTD_MAX_SHARD_SIZE) ) # type: ignore[assign] # We could send this back into the cache thread instead to # serialize access to sqlite3 if lock contention becomes an issue. cache.insert(AnnotatedRawShard(url, node_id.package, data)) shard_out_queue.put([(node_id, shard)]) def result_to_in_queue(future: Future): # Simplify waiting by putting responses back into in_queue. This # function is called in the ThreadPoolExecutor's thread, but we want to # serialize result processing in the network_fetch_thread. in_queue.put([future]) with ThreadPoolExecutor(max_workers=_shards_connections()) as executor: for node_ids_and_results in combine_batches_until_none(in_queue): for node_id_or_result in node_ids_and_results: if isinstance(node_id_or_result, Future): handle_result(node_id_or_result) else: future = submit(node_id_or_result) future.add_done_callback(result_to_in_queue) # TODO call executor.shutdown(cancel_futures=True) on error or otherwise # prevent new HTTP requests from being started e.g. "skip" flag in # fetch() function. Also possible to shutdown(wait=False). # endregion