quay/data/model/storage.py

import logging

from peewee import JOIN_LEFT_OUTER, fn, SQL

from data.model import config, db_transaction, InvalidImageException
from data.database import (ImageStorage, Image, DerivedStorageForImage, ImageStoragePlacement,
                           ImageStorageLocation, ImageStorageTransformation, ImageStorageSignature,
                           ImageStorageSignatureKind, Repository, Namespace)


logger = logging.getLogger(__name__)


def add_storage_placement(storage, location_name):
  """ Adds a storage placement for the given storage at the given location. """
  location = ImageStorageLocation.get(name=location_name)
  ImageStoragePlacement.create(location=location, storage=storage)


def garbage_collect_storage(storage_id_whitelist):
  if len(storage_id_whitelist) == 0:
    return

  def placements_query_to_paths_set(placements_query):
    return {(placement.location.name, get_layer_path(placement.storage))
            for placement in placements_query}

  def orphaned_storage_query(select_base_query, candidates, group_by):
    return (select_base_query
            .switch(ImageStorage)
            .join(Image, JOIN_LEFT_OUTER)
            .where(ImageStorage.id << list(candidates))
            .group_by(*group_by)
            .having(fn.Count(Image.id) == 0))

  # Note: Both of these deletes must occur in the same transaction (unfortunately) because a
  # storage without any placement is invalid, and a placement cannot exist without a storage.
  # TODO(jake): We might want to allow for null storages on placements, which would allow us to
  # delete the storages, then delete the placements in a non-transaction.
  logger.debug('Garbage collecting storages from candidates: %s', storage_id_whitelist)
  with db_transaction():
    # Track all of the data that should be removed from blob storage
    placements_to_remove = list(orphaned_storage_query(ImageStoragePlacement
                                                       .select(ImageStoragePlacement,
                                                               ImageStorage,
                                                               ImageStorageLocation)
                                                       .join(ImageStorageLocation)
                                                       .switch(ImageStoragePlacement)
                                                       .join(ImageStorage),
                                                       storage_id_whitelist,
                                                       (ImageStorage, ImageStoragePlacement,
                                                        ImageStorageLocation)))

    paths_to_remove = placements_query_to_paths_set(placements_to_remove)

    # Remove the placements for orphaned storages
    if len(placements_to_remove) > 0:
      placement_ids_to_remove = [placement.id for placement in placements_to_remove]
      placements_removed = (ImageStoragePlacement
                            .delete()
                            .where(ImageStoragePlacement.id << placement_ids_to_remove)
                            .execute())
      logger.debug('Removed %s image storage placements', placements_removed)

    # Remove all orphaned storages
    # The comma after ImageStorage.id is VERY important, it makes it a tuple, which is a sequence
    orphaned_storages = list(orphaned_storage_query(ImageStorage.select(ImageStorage.id),
                                                    storage_id_whitelist,
                                                    (ImageStorage.id,)).alias('osq'))
    if len(orphaned_storages) > 0:
      storages_removed = (ImageStorage
                          .delete()
                          .where(ImageStorage.id << orphaned_storages)
                          .execute())
      logger.debug('Removed %s image storage records', storages_removed)

  # We are going to make the conscious decision to not delete image storage blobs inside
  # transactions.
  # This may end up producing garbage in s3, trading off for higher availability in the database.
  for location_name, image_path in paths_to_remove:
    logger.debug('Removing %s from %s', image_path, location_name)
    config.store.remove({location_name}, image_path)


def create_v1_storage(location_name):
  storage = ImageStorage.create(cas_path=False)
  location = ImageStorageLocation.get(name=location_name)
  ImageStoragePlacement.create(location=location, storage=storage)
  storage.locations = {location_name}
  return storage


def find_or_create_storage_signature(storage, signature_kind):
  found = lookup_storage_signature(storage, signature_kind)
  if found is None:
    kind = ImageStorageSignatureKind.get(name=signature_kind)
    found = ImageStorageSignature.create(storage=storage, kind=kind)

  return found


def lookup_storage_signature(storage, signature_kind):
  kind = ImageStorageSignatureKind.get(name=signature_kind)
  try:
    return (ImageStorageSignature
            .select()
            .where(ImageStorageSignature.storage == storage, ImageStorageSignature.kind == kind)
            .get())
  except ImageStorageSignature.DoesNotExist:
    return None


def _get_storage(query_modifier):
  query = (ImageStoragePlacement
           .select(ImageStoragePlacement, ImageStorage, ImageStorageLocation)
           .join(ImageStorageLocation)
           .switch(ImageStoragePlacement)
           .join(ImageStorage))

  placements = list(query_modifier(query))

  if not placements:
    raise InvalidImageException()

  found = placements[0].storage
  found.locations = {placement.location.name for placement in placements}

  return found


def get_storage_by_subquery(subquery):
  """ Returns the storage (and its locations) for the storage id returned by the subquery. The
      subquery must return at most 1 result, which is a storage ID. """
  def filter_by_subquery(query):
    return query.where(ImageStorage.id == subquery)

  return _get_storage(filter_by_subquery)


def get_storage_by_uuid(storage_uuid):
  def filter_to_uuid(query):
    return query.where(ImageStorage.uuid == storage_uuid)

  try:
    return _get_storage(filter_to_uuid)
  except InvalidImageException:
    raise InvalidImageException('No storage found with uuid: %s', storage_uuid)


def get_layer_path(storage_record):
  """ Returns the path in the storage engine to the layer data referenced by the storage row. """
  store = config.store
  if not storage_record.cas_path:
    logger.debug('Serving layer from legacy v1 path')
    return store.v1_image_layer_path(storage_record.uuid)

  return store.blob_path(storage_record.content_checksum)


def lookup_repo_storages_by_content_checksum(repo, checksums):
  """ Looks up repository storages (without placements) matching the given repository
      and checksum. """
  # There may be many duplicates of the checksums, so for performance reasons we are going
  # to use a union to select just one storage with each checksum
  queries = []

  for counter, checksum in enumerate(set(checksums)):
    query_alias = 'q{0}'.format(counter)
    candidate_subq = (ImageStorage
                      .select(ImageStorage.id, ImageStorage.content_checksum, ImageStorage.image_size)
                      .join(Image)
                      .where(Image.repository == repo, ImageStorage.content_checksum == checksum)
                      .limit(1)
                      .alias(query_alias))
    queries.append(ImageStorage
                   .select(SQL('*'))
                   .from_(candidate_subq))

  return _reduce_as_tree(queries)


def _reduce_as_tree(queries_to_reduce):
  """ This method will split a list of queries into halves recursively until we reach individual
      queries, at which point it will start unioning the queries, or the already unioned subqueries.
      This works around a bug in peewee SQL generation where reducing linearly generates a chain
      of queries that will exceed the recursion depth limit when it has around 80 queries.
      """
  mid = len(queries_to_reduce)/2
  left = queries_to_reduce[:mid]
  right = queries_to_reduce[mid:]

  to_reduce_right = right[0]
  if len(right) > 1:
    to_reduce_right = _reduce_as_tree(right)

  if len(left) > 1:
    to_reduce_left = _reduce_as_tree(left)
  elif len(left) == 1:
    to_reduce_left = left[0]
  else:
    return to_reduce_right

  return to_reduce_left.union_all(to_reduce_right)


def get_storage_locations(uuid):
  query = (ImageStoragePlacement
           .select()
           .join(ImageStorageLocation)
           .switch(ImageStoragePlacement)
           .join(ImageStorage)
           .where(ImageStorage.uuid == uuid))

  return [location.location.name for location in query]