monitor.py

import datetime
import logging
import traceback

from sqlalchemy.orm import defer
from sqlalchemy.sql.expression import and_, or_

from . import log  # This sets the appropriate log format and level.
from .config import Configuration
from .metadata_layer import TimestampData
from .model import (
    CachedFeed,
    CirculationEvent,
    Collection,
    CollectionMissing,
    CoverageRecord,
    Credential,
    CustomListEntry,
    Edition,
    Identifier,
    LicensePool,
    Measurement,
    Patron,
    PresentationCalculationPolicy,
    Subject,
    Timestamp,
    Work,
    get_one,
    get_one_or_create,
)
from .model.configuration import ConfigurationSetting
from .util.datetime_helpers import utc_now


class CollectionMonitorLogger(logging.LoggerAdapter):
    """Prefix log messages with a collection, if one is present."""

    def __init__(self, logger, extra):
        self.logger = logger
        self.extra = extra
        collection = self.extra.get("collection", None)
        self.log_prefix = "[{}] ".format(collection.name) if collection else ""
        self.warn = self.warning

    def process(self, msg, kwargs):
        return "{}{}".format(self.log_prefix, msg), kwargs


class Monitor(object):
    """A Monitor is responsible for running some piece of code on a
    regular basis. A Monitor has an associated Timestamp that tracks
    the last time it successfully ran; it may use this information on
    its next run to cover the intervening span of time.

    A Monitor will run to completion and then stop. To repeatedly run
    a Monitor, you'll need to repeatedly invoke it from some external
    source such as a cron job.

    This class is designed to be subclassed rather than instantiated
    directly. Subclasses must define SERVICE_NAME. Subclasses may
    define replacement values for DEFAULT_START_TIME and
    DEFAULT_COUNTER.

    Although any Monitor may be associated with a Collection, it's
    most useful to subclass CollectionMonitor if you're writing code
    that needs to be run on every Collection of a certain type.

    """

    # In your subclass, set this to the name of the service,
    # e.g. "Overdrive Circulation Monitor". All instances of your
    # subclass will give this as their service name and track their
    # Timestamps under this name.
    SERVICE_NAME = None

    # Some useful relative constants for DEFAULT_START_TIME (below).
    ONE_MINUTE_AGO = datetime.timedelta(seconds=60)
    ONE_YEAR_AGO = datetime.timedelta(seconds=60 * 60 * 24 * 365)
    NEVER = object()

    # If there is no Timestamp for this Monitor, this time will be
    # passed into `run_once()` as the `start_time` parameter.
    DEFAULT_START_TIME = ONE_MINUTE_AGO

    # When the Timestamp for this Monitor is created, this value will
    # be set for `Timestamp.counter`.
    #
    # This is only used by the SweepMonitor subclass.
    DEFAULT_COUNTER = None

    def __init__(self, _db, collection=None):
        self._db = _db
        cls = self.__class__
        if not self.SERVICE_NAME and not cls.SERVICE_NAME:
            raise ValueError("%s must define SERVICE_NAME." % cls.__name__)
        self.service_name = self.SERVICE_NAME
        default_start_time = cls.DEFAULT_START_TIME
        if isinstance(default_start_time, datetime.timedelta):
            default_start_time = utc_now() - default_start_time
        self.default_start_time = default_start_time
        self.default_counter = cls.DEFAULT_COUNTER

        # We store the collection ID rather than the Collection to
        # avoid breakage in case an app server with a scoped session
        # ever uses a Monitor.
        self.collection_id = None
        if collection:
            self.collection_id = collection.id

    @property
    def log(self):
        if not hasattr(self, "_log"):
            self._log = CollectionMonitorLogger(
                logging.getLogger(self.service_name),
                {"collection": self.collection},
            )
        return self._log

    @property
    def collection(self):
        """Retrieve the Collection object associated with this
        Monitor.
        """
        if not self.collection_id:
            return None
        return get_one(self._db, Collection, id=self.collection_id)

    @property
    def initial_start_time(self):
        """The time that should be used as the 'start time' the first
        time this Monitor is run.
        """
        if self.default_start_time is self.NEVER:
            return None
        if self.default_start_time:
            return self.default_start_time
        return utc_now()

    def timestamp(self):
        """Find or create a Timestamp for this Monitor.

        This does not use TimestampData because it relies on checking
        whether a Timestamp already exists in the database.

        A new timestamp will have .finish set to None, since the first
        run is presumably in progress.
        """
        initial_timestamp = self.initial_start_time
        timestamp, new = get_one_or_create(
            self._db,
            Timestamp,
            service=self.service_name,
            service_type=Timestamp.MONITOR_TYPE,
            collection=self.collection,
            create_method_kwargs=dict(
                start=initial_timestamp,
                finish=None,
                counter=self.default_counter,
            ),
        )
        return timestamp

    def run(self):
        """Do all the work that has piled up since the
        last time the Monitor ran to completion.
        """
        # Use the existing Timestamp to determine the progress made up
        # to this point. It's the job of the subclass to decide where
        # to go from here.
        timestamp_obj = self.timestamp()
        progress = timestamp_obj.to_data()

        this_run_start = utc_now()
        exception = None

        ignorable = (None, TimestampData.CLEAR_VALUE)
        try:
            new_timestamp = self.run_once(progress)
            this_run_finish = utc_now()
            if new_timestamp is None:
                # Assume this Monitor has no special needs surrounding
                # its timestamp.
                new_timestamp = TimestampData()
            if new_timestamp.achievements not in ignorable:
                # This eliminates the need to create similar-looking
                # strings for TimestampData.achievements and for the log.
                self.log.info(new_timestamp.achievements)
            if new_timestamp.exception in ignorable:
                # run_once() completed with no exceptions being raised.
                # We can run the cleanup code and finalize the timestamp.
                self.cleanup()
                new_timestamp.finalize(
                    service=self.service_name,
                    service_type=Timestamp.MONITOR_TYPE,
                    collection=self.collection,
                    start=this_run_start,
                    finish=this_run_finish,
                    exception=None,
                )
                new_timestamp.apply(self._db)
            else:
                # This will be treated the same as an unhandled
                # exception, below.
                exception = new_timestamp.exception
        except Exception:
            this_run_finish = utc_now()
            self.log.exception(
                "Error running %s monitor. Timestamp will not be updated.",
                self.service_name,
            )
            exception = traceback.format_exc()
        if exception is not None:
            # We will update Timestamp.exception but not go through
            # the whole TimestampData.apply() process, which might
            # erase the information the Monitor needs to recover from
            # this failure.
            timestamp_obj.exception = exception

        self._db.commit()

        duration = this_run_finish - this_run_start
        self.log.info(
            "Ran %s monitor in %.2f sec.",
            self.service_name,
            duration.total_seconds(),
        )

    def run_once(self, progress):
        """Do the actual work of the Monitor.

        :param progress: A TimestampData representing the
           work done by the Monitor up to this point.

        :return: A TimestampData representing how you want the
            Monitor's entry in the `timestamps` table to look like
            from this point on.  NOTE: Modifying the incoming
            `progress` and returning it is generally a bad idea,
            because the incoming `progress` is full of old
            data. Instead, return a new TimestampData containing data
            for only the fields you want to set.
        """
        raise NotImplementedError()

    def cleanup(self):
        """Do any work that needs to be done at the end, once the main work
        has completed successfully.
        """
        pass


class TimelineMonitor(Monitor):
    """A monitor that needs to process everything that happened between
    two specific times.

    This Monitor uses `Timestamp.start` and `Timestamp.finish` to describe
    the span of time covered in the most recent run, not the time it
    actually took to run.
    """

    OVERLAP = datetime.timedelta(minutes=5)

    def run_once(self, progress):
        if progress.finish is None:
            # This monitor has never run before. Use the default
            # start time for this monitor.
            start = self.initial_start_time
        else:
            start = progress.finish - self.OVERLAP
        cutoff = utc_now()
        self.catch_up_from(start, cutoff, progress)

        if progress.is_failure:
            # Something has gone wrong. Stop immediately.
            #
            # TODO: Ideally we would undo any other changes made to
            # the TimestampData, but most Monitors don't set
            # .exception directly so it's not a big deal.
            return progress

        # We set `finish` to the time at which we _started_
        # running this process, to reduce the risk that we miss
        # events that happened while the process was running.
        progress.start = start
        progress.finish = cutoff
        return progress

    def catch_up_from(self, start, cutoff, progress):
        """Make sure all events between `start` and `cutoff` are covered.

        :param start: Start looking for events that happened at this
            time.
        :param cutoff: You're not responsible for events that happened
            after this time.
        :param progress: A TimestampData representing the progress so
            far. Unlike with run_once(), you are encouraged to can
            modify this in place, for instance to set .achievements.
            However, you cannot change .start and .finish -- any
            changes will be overwritten by run_once().
        """
        raise NotImplementedError()

    @classmethod
    def slice_timespan(cls, start, cutoff, increment):
        """Slice a span of time into segments no large than [increment].

        This lets you divide up a task like "gather the entire
        circulation history for a collection" into chunks of one day.

        :param start: A datetime.
        :param cutoff: A datetime.
        :param increment: A timedelta.
        """
        slice_start = start
        while slice_start < cutoff:
            full_slice = True
            slice_cutoff = slice_start + increment
            if slice_cutoff > cutoff:
                slice_cutoff = cutoff
                full_slice = False
            yield slice_start, slice_cutoff, full_slice
            slice_start = slice_start + increment


class CollectionMonitor(Monitor):
    """A Monitor that does something for all Collections that come
    from a certain provider.

    This class is designed to be subclassed rather than instantiated
    directly. Subclasses must define SERVICE_NAME and
    PROTOCOL. Subclasses may define replacement values for
    DEFAULT_START_TIME and DEFAULT_COUNTER.
    """

    # Set this to the name of the license provider managed by this
    # Monitor. If this value is set, the CollectionMonitor can only be
    # instantiated with Collections that get their licenses from this
    # provider. If this is unset, the CollectionMonitor can be
    # instantiated with any Collection, or with no Collection at all.
    PROTOCOL = None

    def __init__(self, _db, collection):
        cls = self.__class__
        self.protocol = cls.PROTOCOL
        if self.protocol:
            if collection is None:
                raise CollectionMissing()
        cls._validate_collection(collection, protocol=self.protocol)
        super(CollectionMonitor, self).__init__(_db, collection)

    @classmethod
    def _validate_collection(cls, collection, protocol=None):
        protocol = protocol or cls.PROTOCOL
        if protocol and collection.protocol != protocol:
            raise ValueError(
                "Collection protocol (%s) does not match Monitor protocol (%s)"
                % (collection.protocol, protocol)
            )

    @classmethod
    def all(cls, _db, collections=None, **constructor_kwargs):
        """Yield a sequence of CollectionMonitor objects: one for every
        Collection associated with cls.PROTOCOL.

        If `collections` is specified, then there must be a Monitor
        for each one and Monitors will be yielded in the same order
        that the collections are specified. Otherwise, Monitors will
        be yielded as follows...

        Monitors that have no Timestamp will be yielded first. After
        that, Monitors with older values for Timestamp.start will be
        yielded before Monitors with newer values.

        :param _db: Database session object.
        :param collections: An optional list of collections. If None,
            we'll process all collections.
        :type collections: List[Collection]
        :param constructor_kwargs: These keyword arguments will be passed
            into the CollectionMonitor constructor.

        """
        service_match = or_(
            Timestamp.service == cls.SERVICE_NAME, Timestamp.service == None
        )
        collections_for_protocol = Collection.by_protocol(_db, cls.PROTOCOL).outerjoin(
            Timestamp,
            and_(
                Timestamp.collection_id == Collection.id,
                service_match,
            ),
        )

        if collections:
            # verify that each specified collection exists in this context
            for coll in collections:  # type: str
                try:
                    cls._validate_collection(coll, cls.PROTOCOL)
                except ValueError as e:
                    additional_info = (
                        "Only the following collections are available: {!r}".format(
                            [c.name for c in collections_for_protocol]
                        )
                    )
                    e.args += (additional_info,)
                    raise ValueError(str(e) + "\n" + additional_info)
        else:
            collections = collections_for_protocol.order_by(
                Timestamp.start.asc().nullsfirst()
            )
        for collection in collections:
            yield cls(_db=_db, collection=collection, **constructor_kwargs)


class SweepMonitor(CollectionMonitor):
    """A monitor that does some work for every item in a database table,
    then stops.

    Progress through the table is stored in the Timestamp, so that if
    the Monitor crashes, the next time the Monitor is run, it starts
    at the item that caused the crash, rather than starting from the
    beginning of the table.
    """

    # The completion of each individual item should be logged at
    # this log level.
    COMPLETION_LOG_LEVEL = logging.INFO

    # Items will be processed in batches of this size.
    DEFAULT_BATCH_SIZE = 100

    DEFAULT_COUNTER = 0

    # The model class corresponding to the database table that this
    # Monitor sweeps over. This class must keep its primary key in the
    # `id` field.
    MODEL_CLASS = None

    def __init__(self, _db, collection=None, batch_size=None):
        cls = self.__class__
        if not batch_size or batch_size < 0:
            batch_size = cls.DEFAULT_BATCH_SIZE
        self.batch_size = batch_size
        if not cls.MODEL_CLASS:
            raise ValueError("%s must define MODEL_CLASS" % cls.__name__)
        self.model_class = cls.MODEL_CLASS
        super(SweepMonitor, self).__init__(_db, collection=collection)

    def run_once(self, *ignore):
        timestamp = self.timestamp()
        offset = timestamp.counter
        new_offset = offset
        exception = None

        # The timestamp for a SweepMonitor is purely informative --
        # we're not trying to capture all the events that happened
        # since a certain time -- so we're going to make sure the
        # timestamp is set from the start of the run to the end of the
        # last _successful_ batch.
        run_started_at = utc_now()
        timestamp.start = run_started_at

        total_processed = 0
        while True:
            old_offset = offset
            batch_started_at = utc_now()
            new_offset, batch_size = self.process_batch(offset)
            total_processed += batch_size
            batch_ended_at = utc_now()

            self.log.debug(
                "%s monitor went from offset %s to %s in %.2f sec",
                self.service_name,
                offset,
                new_offset,
                (batch_ended_at - batch_started_at).total_seconds(),
            )
            achievements = "Records processed: %d." % total_processed

            offset = new_offset
            if offset == 0:
                # We completed a sweep. We're done.
                break

            # We need to do another batch. If it should raise an exception,
            # we don't want to lose the progress we've already made.
            timestamp.update(
                counter=new_offset, finish=batch_ended_at, achievements=achievements
            )
            self._db.commit()

        # We're done with this run. The run() method will do the final
        # update.
        return TimestampData(counter=offset, achievements=achievements)

    def process_batch(self, offset):
        """Process one batch of work."""
        offset = offset or 0
        items = self.fetch_batch(offset).all()
        if items:
            self.process_items(items)
            # We've completed a batch. Return the ID of the last item
            # in the batch so we don't do this work again.
            return items[-1].id, len(items)
        else:
            # There are no more items in this database table, so we
            # are done with the sweep. Reset the counter.
            return 0, 0

    def process_items(self, items):
        """Process a list of items."""
        for item in items:
            self.process_item(item)
            self.log.log(self.COMPLETION_LOG_LEVEL, "Completed %r", item)

    def fetch_batch(self, offset):
        """Retrieve one batch of work from the database."""
        q = (
            self.item_query()
            .filter(self.model_class.id > offset)
            .order_by(self.model_class.id)
            .limit(self.batch_size)
        )
        return q

    def item_query(self):
        """Find the items that need to be processed in the sweep.

        :return: A query object.
        """
        # Start by getting everything in the table.
        qu = self._db.query(self.model_class)
        if self.collection:
            # Restrict to only those items associated with self.collection
            # somehow.
            qu = self.scope_to_collection(qu, self.collection)
        qu = qu.order_by(self.model_class.id)
        return qu

    def scope_to_collection(self, qu, collection):
        """Restrict the given query so that it only finds items
        associated with the given collection.

        :param qu: A query object.
        :param collection: A Collection object, presumed to not be None.
        """
        raise NotImplementedError()

    def process_item(self, item):
        """Do the work that needs to be done for a given item."""
        raise NotImplementedError()


class IdentifierSweepMonitor(SweepMonitor):
    """A Monitor that does some work for every Identifier."""

    MODEL_CLASS = Identifier

    def scope_to_collection(self, qu, collection):
        """Only find Identifiers licensed through the given Collection."""
        return qu.join(Identifier.licensed_through).filter(
            LicensePool.collection == collection
        )


class SubjectSweepMonitor(SweepMonitor):
    """A Monitor that does some work for every Subject."""

    MODEL_CLASS = Subject

    # It's usually easy to process a Subject, so make the batch size
    # large.
    DEFAULT_BATCH_SIZE = 500

    def __init__(self, _db, subject_type=None, filter_string=None):
        """Constructor.
        :param subject_type: Only process Subjects of this type.
        :param filter_string: Only process Subjects whose .identifier
           or .name contain this string.
        """
        super(SubjectSweepMonitor, self).__init__(_db, None)
        self.subject_type = subject_type
        self.filter_string = filter_string

    def item_query(self):
        """Find only Subjects that match the given filters."""
        qu = self._db.query(Subject)
        if self.subject_type:
            qu = qu.filter(Subject.type == self.subject_type)
        if self.filter_string:
            filter_string = "%" + self.filter_string + "%"
            or_clause = or_(
                Subject.identifier.ilike(filter_string),
                Subject.name.ilike(filter_string),
            )
            qu = qu.filter(or_clause)
        return qu

    def scope_to_collection(self, qu, collection):
        """Refuse to scope this query to a Collection."""
        return qu


class CustomListEntrySweepMonitor(SweepMonitor):
    """A Monitor that does something to every CustomListEntry."""

    MODEL_CLASS = CustomListEntry

    def scope_to_collection(self, qu, collection):
        """Restrict the query to only find CustomListEntries whose
        Work is in the given Collection.
        """
        return (
            qu.join(CustomListEntry.work)
            .join(Work.license_pools)
            .filter(LicensePool.collection == collection)
        )


class EditionSweepMonitor(SweepMonitor):
    """A Monitor that does something to every Edition."""

    MODEL_CLASS = Edition

    def scope_to_collection(self, qu, collection):
        """Restrict the query to only find Editions whose
        primary Identifier is licensed to the given Collection.
        """
        return (
            qu.join(Edition.primary_identifier)
            .join(Identifier.licensed_through)
            .filter(LicensePool.collection == collection)
        )


class WorkSweepMonitor(SweepMonitor):
    """A Monitor that does something to every Work."""

    MODEL_CLASS = Work

    def scope_to_collection(self, qu, collection):
        """Restrict the query to only find Works found in the given
        Collection.
        """
        return qu.join(Work.license_pools).filter(LicensePool.collection == collection)


class PresentationReadyWorkSweepMonitor(WorkSweepMonitor):
    """A Monitor that does something to every presentation-ready Work."""

    def item_query(self):
        return (
            super(PresentationReadyWorkSweepMonitor, self)
            .item_query()
            .filter(Work.presentation_ready == True)
        )


class NotPresentationReadyWorkSweepMonitor(WorkSweepMonitor):
    """A Monitor that does something to every Work that is not
    presentation-ready.
    """

    def item_query(self):
        not_presentation_ready = or_(
            Work.presentation_ready == False, Work.presentation_ready == None
        )
        return (
            super(NotPresentationReadyWorkSweepMonitor, self)
            .item_query()
            .filter(not_presentation_ready)
        )


# SweepMonitors that do something specific.


class OPDSEntryCacheMonitor(PresentationReadyWorkSweepMonitor):
    """A Monitor that recalculates the OPDS entries for every
    presentation-ready Work.

    This is different from the OPDSEntryWorkCoverageProvider,
    which only processes works that are missing a WorkCoverageRecord
    with the 'generate-opds' operation.
    """

    SERVICE_NAME = "ODPS Entry Cache Monitor"

    def process_item(self, work):
        work.calculate_opds_entries()


class PermanentWorkIDRefreshMonitor(EditionSweepMonitor):
    """A monitor that calculates or recalculates the permanent work ID for
    every edition.
    """

    SERVICE_NAME = "Permanent work ID refresh"

    def process_item(self, edition):
        edition.calculate_permanent_work_id()


class MakePresentationReadyMonitor(NotPresentationReadyWorkSweepMonitor):
    """A monitor that makes works presentation ready.

    By default this works by passing the work's active edition into
    ensure_coverage() for each of a list of CoverageProviders. If all
    the ensure_coverage() calls succeed, presentation of the work is
    calculated and the work is marked presentation ready.
    """

    SERVICE_NAME = "Make Works Presentation Ready"

    def __init__(self, _db, coverage_providers, collection=None):
        super(MakePresentationReadyMonitor, self).__init__(_db, collection)
        self.coverage_providers = coverage_providers
        self.policy = PresentationCalculationPolicy(choose_edition=False)

    def run(self):
        """Before doing anything, consolidate works."""
        LicensePool.consolidate_works(self._db)
        return super(MakePresentationReadyMonitor, self).run()

    def process_item(self, work):
        """Do the work necessary to make one Work presentation-ready,
        and handle exceptions.
        """
        exception = None

        try:
            self.prepare(work)
        except CoverageProvidersFailed as e:
            exception = "Provider(s) failed: %s" % e
        except Exception as e:
            self.log.error("Exception processing work %r", work, exc_info=e)
            exception = str(e)

        if exception:
            # Unlike with most Monitors, an exception is not a good
            # reason to stop doing our job. Note it inside the Work
            # and keep going.
            work.presentation_ready_exception = exception
        else:
            # Success!
            work.calculate_presentation(self.policy)
            work.set_presentation_ready()

    def prepare(self, work):
        """Try to make a single Work presentation-ready.

        :raise CoverageProvidersFailed: If we can't make a Work
            presentation-ready because one or more CoverageProviders
            failed.
        """
        edition = work.presentation_edition
        if not edition:
            work = work.calculate_presentation()
        identifier = edition.primary_identifier
        overall_success = True
        failures = []
        for provider in self.coverage_providers:
            covered_types = provider.input_identifier_types
            if covered_types and identifier.type in covered_types:
                coverage_record = provider.ensure_coverage(identifier)
                if (
                    not isinstance(coverage_record, CoverageRecord)
                    or coverage_record.status != CoverageRecord.SUCCESS
                    or coverage_record.exception is not None
                ):
                    # This provider has failed.
                    failures.append(provider)
        if failures:
            raise CoverageProvidersFailed(failures)
        return failures


class CoverageProvidersFailed(Exception):
    """We tried to run CoverageProviders on a Work's identifier,
    but some of the providers failed.
    """

    def __init__(self, failed_providers):
        self.failed_providers = failed_providers
        super(CoverageProvidersFailed, self).__init__(
            ", ".join([x.service_name for x in failed_providers])
        )


class CustomListEntryWorkUpdateMonitor(CustomListEntrySweepMonitor):

    """Set or reset the Work associated with each custom list entry."""

    SERVICE_NAME = "Update Works for custom list entries"
    DEFAULT_BATCH_SIZE = 100

    def process_item(self, item):
        item.set_work()


class ReaperMonitor(Monitor):
    """A Monitor that deletes database rows that have expired but
    have no other process to delete them.

    A subclass of ReaperMonitor MUST define values for the following
    constants:
    * MODEL_CLASS - The model class this monitor is reaping, e.g. Credential.
    * TIMESTAMP_FIELD - Within the model class, the DateTime field to be
    used when deciding which rows to deleting,
    e.g. 'expires'. The reaper will be more efficient if there's
    an index on this field.
    * MAX_AGE - A datetime.timedelta or number of days representing
    the time that must pass before an item can be safely deleted.

    A subclass of ReaperMonitor MAY define values for the following constants:
    * BATCH_SIZE - The number of rows to fetch for deletion in a single
    batch. The default is 1000.

    If your model class has fields that might contain a lot of data
    and aren't important to the reaping process, put their field names
    into a list called LARGE_FIELDS and the Reaper will avoid fetching
    that information, improving performance.
    """

    MODEL_CLASS = None
    TIMESTAMP_FIELD = None
    MAX_AGE = None
    BATCH_SIZE = 1000

    REGISTRY = []

    def __init__(self, *args, **kwargs):
        self.SERVICE_NAME = "Reaper for %s" % self.MODEL_CLASS.__name__
        if self.TIMESTAMP_FIELD is not None:
            self.SERVICE_NAME += ".%s" % self.TIMESTAMP_FIELD

        super(ReaperMonitor, self).__init__(*args, **kwargs)

    @property
    def cutoff(self):
        """Items with a timestamp earlier than this time will be reaped."""
        if isinstance(self.MAX_AGE, datetime.timedelta):
            max_age = self.MAX_AGE
        else:
            max_age = datetime.timedelta(days=self.MAX_AGE)
        return utc_now() - max_age

    @property
    def timestamp_field(self):
        return getattr(self.MODEL_CLASS, self.TIMESTAMP_FIELD)

    @property
    def where_clause(self):
        """A SQLAlchemy clause that identifies the database rows to be reaped."""
        return self.timestamp_field < self.cutoff

    def run_once(self, *args, **kwargs):
        rows_deleted = 0
        qu = self.query()
        to_defer = getattr(self.MODEL_CLASS, "LARGE_FIELDS", [])
        for x in to_defer:
            qu = qu.options(defer(x))
        count = qu.count()
        self.log.info("Deleting %d row(s)", count)
        while count > 0:
            for i in qu.limit(self.BATCH_SIZE):
                self.log.info("Deleting %r", i)
                self.delete(i)
                rows_deleted += 1
            self._db.commit()
            count = qu.count()
        return TimestampData(achievements="Items deleted: %d" % rows_deleted)

    def delete(self, row):
        """Delete a row from the database.

        CAUTION: If you override this method such that it doesn't
        actually delete the database row, then run_once() may enter an
        infinite loop.
        """
        self._db.delete(row)

    def query(self):
        return self._db.query(self.MODEL_CLASS).filter(self.where_clause)


# ReaperMonitors that do something specific.


class CachedFeedReaper(ReaperMonitor):
    """Removed cached feeds older than thirty days."""

    MODEL_CLASS = CachedFeed
    TIMESTAMP_FIELD = "timestamp"
    MAX_AGE = 30


ReaperMonitor.REGISTRY.append(CachedFeedReaper)


class CredentialReaper(ReaperMonitor):
    """Remove Credentials that expired more than a day ago."""

    MODEL_CLASS = Credential
    TIMESTAMP_FIELD = "expires"
    MAX_AGE = 1


ReaperMonitor.REGISTRY.append(CredentialReaper)


class PatronRecordReaper(ReaperMonitor):
    """Remove patron records that expired more than 60 days ago"""

    MODEL_CLASS = Patron
    TIMESTAMP_FIELD = "authorization_expires"
    MAX_AGE = 60


ReaperMonitor.REGISTRY.append(PatronRecordReaper)


class WorkReaper(ReaperMonitor):
    """Remove Works that have no associated LicensePools.

    Unlike other reapers, no timestamp is relevant. As soon as a Work
    loses its last LicensePool it can be removed.
    """

    MODEL_CLASS = Work

    def __init__(self, *args, **kwargs):
        from .external_search import ExternalSearchIndex

        search_index_client = kwargs.pop("search_index_client", None)
        super(WorkReaper, self).__init__(*args, **kwargs)
        self.search_index_client = search_index_client or ExternalSearchIndex(self._db)

    def query(self):
        return (
            self._db.query(Work)
            .outerjoin(Work.license_pools)
            .filter(LicensePool.id == None)
        )

    def delete(self, work):
        """Delete work from elasticsearch and database."""
        work.delete(self.search_index_client)


ReaperMonitor.REGISTRY.append(WorkReaper)


class CollectionReaper(ReaperMonitor):
    """Remove collections that have been marked for deletion."""

    MODEL_CLASS = Collection

    @property
    def where_clause(self):
        """A SQLAlchemy clause that identifies the database rows to be reaped."""
        return Collection.marked_for_deletion == True

    def delete(self, collection):
        """Delete a Collection from the database.

        Database deletion of a Collection might take a really long
        time, so we call a special method that will do the deletion
        incrementally and can pick up where it left off if there's a
        failure.
        """
        collection.delete()


ReaperMonitor.REGISTRY.append(CollectionReaper)


class MeasurementReaper(ReaperMonitor):
    """Remove measurements that are not the most recent"""

    MODEL_CLASS = Measurement

    def run(self):
        enabled = ConfigurationSetting.sitewide(
            self._db, Configuration.MEASUREMENT_REAPER
        ).bool_value
        if enabled is not None and not enabled:
            self.log.info(
                "{} skipped because it is disabled in configuration.".format(
                    self.service_name
                )
            )
            return
        return super(ReaperMonitor, self).run()

    @property
    def where_clause(self):
        return Measurement.is_most_recent == False

    def run_once(self, *args, **kwargs):
        rows_deleted = self.query().delete()
        self._db.commit()
        return TimestampData(achievements="Items deleted: %d" % rows_deleted)


ReaperMonitor.REGISTRY.append(MeasurementReaper)


class ScrubberMonitor(ReaperMonitor):
    """Scrub information from the database.

    Unlike the other ReaperMonitors, this class doesn't delete rows
    from the database -- it only clears out specific data fields.

    In addition to the constants required for ReaperMonitor, a
    subclass of ScrubberMonitor MUST define a value for the following
    constant:

    * SCRUB_FIELD - The field whose value will be set to None when a row
      is scrubbed.
    """

    def __init__(self, *args, **kwargs):
        """Set the name of the Monitor based on which field is being
        scrubbed.
        """
        super(ScrubberMonitor, self).__init__(*args, **kwargs)
        self.SERVICE_NAME = "Scrubber for %s.%s" % (
            self.MODEL_CLASS.__name__,
            self.SCRUB_FIELD,
        )

    def run_once(self, *args, **kwargs):
        """Find all rows that need to be scrubbed, and scrub them."""
        rows_scrubbed = 0
        cls = self.MODEL_CLASS
        update = (
            cls.__table__.update()
            .where(self.where_clause)
            .values({self.SCRUB_FIELD: None})
            .returning(cls.id)
        )
        scrubbed = self._db.execute(update).fetchall()
        self._db.commit()
        return TimestampData(achievements="Items scrubbed: %d" % len(scrubbed))

    @property
    def where_clause(self):
        """Find rows that are older than MAX_AGE _and_ which have a non-null
        SCRUB_FIELD. If the field is already null, there's no need to
        scrub it.
        """
        return and_(super(ScrubberMonitor, self).where_clause, self.scrub_field != None)

    @property
    def scrub_field(self):
        """Find the SQLAlchemy representation of the model field to be
        scrubbed.
        """
        return getattr(self.MODEL_CLASS, self.SCRUB_FIELD)


class CirculationEventLocationScrubber(ScrubberMonitor):
    """Scrub location information from old CirculationEvents."""

    MODEL_CLASS = CirculationEvent
    TIMESTAMP_FIELD = "start"
    MAX_AGE = 365
    SCRUB_FIELD = "location"


ReaperMonitor.REGISTRY.append(CirculationEventLocationScrubber)


class PatronNeighborhoodScrubber(ScrubberMonitor):
    """Scrub cached neighborhood information from patrons who haven't been
    seen in a while.
    """

    MODEL_CLASS = Patron
    TIMESTAMP_FIELD = "last_external_sync"
    MAX_AGE = Patron.MAX_SYNC_TIME
    SCRUB_FIELD = "cached_neighborhood"


ReaperMonitor.REGISTRY.append(PatronNeighborhoodScrubber)