forked-synapse/synapse/storage/background_updates.py

# Copyright 2014-2016 OpenMarket Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import abc
import logging
from enum import Enum, IntEnum
from types import TracebackType
from typing import (
    TYPE_CHECKING,
    Any,
    AsyncContextManager,
    Awaitable,
    Callable,
    Dict,
    Iterable,
    List,
    Optional,
    Sequence,
    Tuple,
    Type,
)

import attr
from pydantic import BaseModel

from synapse.metrics.background_process_metrics import run_as_background_process
from synapse.storage.engines import PostgresEngine
from synapse.storage.types import Connection, Cursor
from synapse.types import JsonDict
from synapse.util import Clock, json_encoder

from . import engines

if TYPE_CHECKING:
    from synapse.server import HomeServer
    from synapse.storage.database import DatabasePool, LoggingTransaction

logger = logging.getLogger(__name__)


ON_UPDATE_CALLBACK = Callable[[str, str, bool], AsyncContextManager[int]]
DEFAULT_BATCH_SIZE_CALLBACK = Callable[[str, str], Awaitable[int]]
MIN_BATCH_SIZE_CALLBACK = Callable[[str, str], Awaitable[int]]


class Constraint(metaclass=abc.ABCMeta):
    """Base class representing different constraints.

    Used by `register_background_validate_constraint_and_delete_rows`.
    """

    @abc.abstractmethod
    def make_check_clause(self, table: str) -> str:
        """Returns an SQL expression that checks the row passes the constraint."""
        pass

    @abc.abstractmethod
    def make_constraint_clause_postgres(self) -> str:
        """Returns an SQL clause for creating the constraint.

        Only used on Postgres DBs
        """
        pass


@attr.s(auto_attribs=True)
class ForeignKeyConstraint(Constraint):
    """A foreign key constraint.

    Attributes:
        referenced_table: The "parent" table name.
        columns: The list of mappings of columns from table to referenced table
        deferred: Whether to defer checking of the constraint to the end of the
            transaction. This is useful for e.g. backwards compatibility where
            an older version inserted data in the wrong order.
    """

    referenced_table: str
    columns: Sequence[Tuple[str, str]]
    deferred: bool

    def make_check_clause(self, table: str) -> str:
        join_clause = " AND ".join(
            f"{col1} = {table}.{col2}" for col1, col2 in self.columns
        )
        return f"EXISTS (SELECT 1 FROM {self.referenced_table} WHERE {join_clause})"

    def make_constraint_clause_postgres(self) -> str:
        column1_list = ", ".join(col1 for col1, col2 in self.columns)
        column2_list = ", ".join(col2 for col1, col2 in self.columns)
        defer_clause = " DEFERRABLE INITIALLY DEFERRED" if self.deferred else ""
        return f"FOREIGN KEY ({column1_list}) REFERENCES {self.referenced_table} ({column2_list}) {defer_clause}"


@attr.s(auto_attribs=True)
class NotNullConstraint(Constraint):
    """A NOT NULL column constraint"""

    column: str

    def make_check_clause(self, table: str) -> str:
        return f"{self.column} IS NOT NULL"

    def make_constraint_clause_postgres(self) -> str:
        return f"CHECK ({self.column} IS NOT NULL)"


class ValidateConstraintProgress(BaseModel):
    """The format of the progress JSON for validate constraint background
    updates.

    Used by `register_background_validate_constraint_and_delete_rows`.
    """

    class State(str, Enum):
        check = "check"
        validate = "validate"

    state: State = State.validate
    lower_bound: Sequence[Any] = ()


@attr.s(slots=True, frozen=True, auto_attribs=True)
class _BackgroundUpdateHandler:
    """A handler for a given background update.

    Attributes:
        callback: The function to call to make progress on the background
            update.
        oneshot: Wether the update is likely to happen all in one go, ignoring
            the supplied target duration, e.g. index creation. This is used by
            the update controller to help correctly schedule the update.
    """

    callback: Callable[[JsonDict, int], Awaitable[int]]
    oneshot: bool = False


class _BackgroundUpdateContextManager:
    def __init__(
        self, sleep: bool, clock: Clock, sleep_duration_ms: int, update_duration: int
    ):
        self._sleep = sleep
        self._clock = clock
        self._sleep_duration_ms = sleep_duration_ms
        self._update_duration_ms = update_duration

    async def __aenter__(self) -> int:
        if self._sleep:
            await self._clock.sleep(self._sleep_duration_ms / 1000)

        return self._update_duration_ms

    async def __aexit__(
        self,
        exc_type: Optional[Type[BaseException]],
        exc: Optional[BaseException],
        tb: Optional[TracebackType],
    ) -> None:
        pass


class BackgroundUpdatePerformance:
    """Tracks the how long a background update is taking to update its items"""

    def __init__(self, name: str):
        self.name = name
        self.total_item_count = 0
        self.total_duration_ms = 0.0
        self.avg_item_count = 0.0
        self.avg_duration_ms = 0.0

    def update(self, item_count: int, duration_ms: float) -> None:
        """Update the stats after doing an update"""
        self.total_item_count += item_count
        self.total_duration_ms += duration_ms

        # Exponential moving averages for the number of items updated and
        # the duration.
        self.avg_item_count += 0.1 * (item_count - self.avg_item_count)
        self.avg_duration_ms += 0.1 * (duration_ms - self.avg_duration_ms)

    def average_items_per_ms(self) -> Optional[float]:
        """An estimate of how long it takes to do a single update.
        Returns:
            A duration in ms as a float
        """
        # We want to return None if this is the first background update item
        if self.total_item_count == 0:
            return None
        # Avoid dividing by zero
        elif self.avg_duration_ms == 0:
            return 0
        else:
            # Use the exponential moving average so that we can adapt to
            # changes in how long the update process takes.
            return float(self.avg_item_count) / float(self.avg_duration_ms)

    def total_items_per_ms(self) -> Optional[float]:
        """An estimate of how long it takes to do a single update.
        Returns:
            A duration in ms as a float
        """
        if self.total_duration_ms == 0:
            return 0
        elif self.total_item_count == 0:
            return None
        else:
            return float(self.total_item_count) / float(self.total_duration_ms)


class UpdaterStatus(IntEnum):
    # Use negative values for error conditions.
    ABORTED = -1
    DISABLED = 0
    NOT_STARTED = 1
    RUNNING_UPDATE = 2
    COMPLETE = 3


class BackgroundUpdater:
    """Background updates are updates to the database that run in the
    background. Each update processes a batch of data at once. We attempt to
    limit the impact of each update by monitoring how long each batch takes to
    process and autotuning the batch size.
    """

    def __init__(self, hs: "HomeServer", database: "DatabasePool"):
        self._clock = hs.get_clock()
        self.db_pool = database

        self._database_name = database.name()

        # if a background update is currently running, its name.
        self._current_background_update: Optional[str] = None

        self._on_update_callback: Optional[ON_UPDATE_CALLBACK] = None
        self._default_batch_size_callback: Optional[DEFAULT_BATCH_SIZE_CALLBACK] = None
        self._min_batch_size_callback: Optional[MIN_BATCH_SIZE_CALLBACK] = None

        self._background_update_performance: Dict[str, BackgroundUpdatePerformance] = {}
        self._background_update_handlers: Dict[str, _BackgroundUpdateHandler] = {}
        # TODO: all these bool flags make me feel icky---can we combine into a status
        # enum?
        self._all_done = False

        # Whether we're currently running updates
        self._running = False

        # Marker to be set if we abort and halt all background updates.
        self._aborted = False

        # Whether background updates are enabled. This allows us to
        # enable/disable background updates via the admin API.
        self.enabled = True

        self.minimum_background_batch_size = hs.config.background_updates.min_batch_size
        self.default_background_batch_size = (
            hs.config.background_updates.default_batch_size
        )
        self.update_duration_ms = hs.config.background_updates.update_duration_ms
        self.sleep_duration_ms = hs.config.background_updates.sleep_duration_ms
        self.sleep_enabled = hs.config.background_updates.sleep_enabled

    def get_status(self) -> UpdaterStatus:
        """An integer summarising the updater status. Used as a metric."""
        if self._aborted:
            return UpdaterStatus.ABORTED
        # TODO: a status for "have seen at least one failure, but haven't aborted yet".
        if not self.enabled:
            return UpdaterStatus.DISABLED

        if self._all_done:
            return UpdaterStatus.COMPLETE
        if self._running:
            return UpdaterStatus.RUNNING_UPDATE
        return UpdaterStatus.NOT_STARTED

    def register_update_controller_callbacks(
        self,
        on_update: ON_UPDATE_CALLBACK,
        default_batch_size: Optional[DEFAULT_BATCH_SIZE_CALLBACK] = None,
        min_batch_size: Optional[DEFAULT_BATCH_SIZE_CALLBACK] = None,
    ) -> None:
        """Register callbacks from a module for each hook."""
        if self._on_update_callback is not None:
            logger.warning(
                "More than one module tried to register callbacks for controlling"
                " background updates. Only the callbacks registered by the first module"
                " (in order of appearance in Synapse's configuration file) that tried to"
                " do so will be called."
            )

            return

        self._on_update_callback = on_update

        if default_batch_size is not None:
            self._default_batch_size_callback = default_batch_size

        if min_batch_size is not None:
            self._min_batch_size_callback = min_batch_size

    def _get_context_manager_for_update(
        self,
        sleep: bool,
        update_name: str,
        database_name: str,
        oneshot: bool,
    ) -> AsyncContextManager[int]:
        """Get a context manager to run a background update with.

        If a module has registered a `update_handler` callback, use the context manager
        it returns.

        Otherwise, returns a context manager that will return a default value, optionally
        sleeping if needed.

        Args:
            sleep: Whether we can sleep between updates.
            update_name: The name of the update.
            database_name: The name of the database the update is being run on.
            oneshot: Whether the update will complete all in one go, e.g. index creation.
                In such cases the returned target duration is ignored.

        Returns:
            The target duration in milliseconds that the background update should run for.

            Note: this is a *target*, and an iteration may take substantially longer or
            shorter.
        """
        if self._on_update_callback is not None:
            return self._on_update_callback(update_name, database_name, oneshot)

        return _BackgroundUpdateContextManager(
            sleep, self._clock, self.sleep_duration_ms, self.update_duration_ms
        )

    async def _default_batch_size(self, update_name: str, database_name: str) -> int:
        """The batch size to use for the first iteration of a new background
        update.
        """
        if self._default_batch_size_callback is not None:
            return await self._default_batch_size_callback(update_name, database_name)

        return self.default_background_batch_size

    async def _min_batch_size(self, update_name: str, database_name: str) -> int:
        """A lower bound on the batch size of a new background update.

        Used to ensure that progress is always made. Must be greater than 0.
        """
        if self._min_batch_size_callback is not None:
            return await self._min_batch_size_callback(update_name, database_name)

        return self.minimum_background_batch_size

    def get_current_update(self) -> Optional[BackgroundUpdatePerformance]:
        """Returns the current background update, if any."""

        update_name = self._current_background_update
        if not update_name:
            return None

        perf = self._background_update_performance.get(update_name)
        if not perf:
            perf = BackgroundUpdatePerformance(update_name)

        return perf

    def start_doing_background_updates(self) -> None:
        if self.enabled:
            # if we start a new background update, not all updates are done.
            self._all_done = False
            sleep = self.sleep_enabled
            run_as_background_process(
                "background_updates", self.run_background_updates, sleep
            )

    async def run_background_updates(self, sleep: bool) -> None:
        if self._running or not self.enabled:
            return

        self._running = True

        back_to_back_failures = 0

        try:
            logger.info(
                "Starting background schema updates for database %s",
                self._database_name,
            )
            while self.enabled:
                try:
                    result = await self.do_next_background_update(sleep)
                    back_to_back_failures = 0
                except Exception:
                    back_to_back_failures += 1
                    if back_to_back_failures >= 5:
                        self._aborted = True
                        raise RuntimeError(
                            "5 back-to-back background update failures; aborting."
                        )
                    logger.exception("Error doing update")
                else:
                    if result:
                        logger.info(
                            "No more background updates to do."
                            " Unscheduling background update task."
                        )
                        self._all_done = True
                        return None
        finally:
            self._running = False

    async def has_completed_background_updates(self) -> bool:
        """Check if all the background updates have completed

        Returns:
            True if all background updates have completed
        """
        # if we've previously determined that there is nothing left to do, that
        # is easy
        if self._all_done:
            return True

        # obviously, if we are currently processing an update, we're not done.
        if self._current_background_update:
            return False

        # otherwise, check if there are updates to be run. This is important,
        # as we may be running on a worker which doesn't perform the bg updates
        # itself, but still wants to wait for them to happen.
        updates = await self.db_pool.simple_select_onecol(
            "background_updates",
            keyvalues=None,
            retcol="1",
            desc="has_completed_background_updates",
        )
        if not updates:
            self._all_done = True
            return True

        return False

    async def has_completed_background_update(self, update_name: str) -> bool:
        """Check if the given background update has finished running."""
        if self._all_done:
            return True

        if update_name == self._current_background_update:
            return False

        update_exists = await self.db_pool.simple_select_one_onecol(
            "background_updates",
            keyvalues={"update_name": update_name},
            retcol="1",
            desc="has_completed_background_update",
            allow_none=True,
        )

        return not update_exists

    async def do_next_background_update(self, sleep: bool = True) -> bool:
        """Does some amount of work on the next queued background update

        Returns once some amount of work is done.

        Args:
            sleep: Whether to limit how quickly we run background updates or
                not.

        Returns:
            True if we have finished running all the background updates, otherwise False
        """

        def get_background_updates_txn(txn: Cursor) -> List[Dict[str, Any]]:
            txn.execute(
                """
                SELECT update_name, depends_on FROM background_updates
                ORDER BY ordering, update_name
                """
            )
            return self.db_pool.cursor_to_dict(txn)

        if not self._current_background_update:
            all_pending_updates = await self.db_pool.runInteraction(
                "background_updates",
                get_background_updates_txn,
            )
            if not all_pending_updates:
                # no work left to do
                return True

            # find the first update which isn't dependent on another one in the queue.
            pending = {update["update_name"] for update in all_pending_updates}
            for upd in all_pending_updates:
                depends_on = upd["depends_on"]
                if not depends_on or depends_on not in pending:
                    break
                logger.info(
                    "Not starting on bg update %s until %s is done",
                    upd["update_name"],
                    depends_on,
                )
            else:
                # if we get to the end of that for loop, there is a problem
                raise Exception(
                    "Unable to find a background update which doesn't depend on "
                    "another: dependency cycle?"
                )

            self._current_background_update = upd["update_name"]

        # We have a background update to run, otherwise we would have returned
        # early.
        assert self._current_background_update is not None
        update_info = self._background_update_handlers[self._current_background_update]

        async with self._get_context_manager_for_update(
            sleep=sleep,
            update_name=self._current_background_update,
            database_name=self._database_name,
            oneshot=update_info.oneshot,
        ) as desired_duration_ms:
            await self._do_background_update(desired_duration_ms)

        return False

    async def _do_background_update(self, desired_duration_ms: float) -> int:
        assert self._current_background_update is not None
        update_name = self._current_background_update
        logger.info("Starting update batch on background update '%s'", update_name)

        update_handler = self._background_update_handlers[update_name].callback

        performance = self._background_update_performance.get(update_name)

        if performance is None:
            performance = BackgroundUpdatePerformance(update_name)
            self._background_update_performance[update_name] = performance

        items_per_ms = performance.average_items_per_ms()

        if items_per_ms is not None:
            batch_size = int(desired_duration_ms * items_per_ms)
            # Clamp the batch size so that we always make progress
            batch_size = max(
                batch_size,
                await self._min_batch_size(update_name, self._database_name),
            )
        else:
            batch_size = await self._default_batch_size(
                update_name, self._database_name
            )

        progress_json = await self.db_pool.simple_select_one_onecol(
            "background_updates",
            keyvalues={"update_name": update_name},
            retcol="progress_json",
        )

        # Avoid a circular import.
        from synapse.storage._base import db_to_json

        progress = db_to_json(progress_json)

        time_start = self._clock.time_msec()
        items_updated = await update_handler(progress, batch_size)
        time_stop = self._clock.time_msec()

        duration_ms = time_stop - time_start

        performance.update(items_updated, duration_ms)

        logger.info(
            "Running background update %r. Processed %r items in %rms."
            " (total_rate=%r/ms, current_rate=%r/ms, total_updated=%r, batch_size=%r)",
            update_name,
            items_updated,
            duration_ms,
            performance.total_items_per_ms(),
            performance.average_items_per_ms(),
            performance.total_item_count,
            batch_size,
        )

        return len(self._background_update_performance)

    def register_background_update_handler(
        self,
        update_name: str,
        update_handler: Callable[[JsonDict, int], Awaitable[int]],
    ) -> None:
        """Register a handler for doing a background update.

        The handler should take two arguments:

        * A dict of the current progress
        * An integer count of the number of items to update in this batch.

        The handler should return a deferred or coroutine which returns an integer count
        of items updated.

        The handler is responsible for updating the progress of the update.

        Args:
            update_name: The name of the update that this code handles.
            update_handler: The function that does the update.
        """
        self._background_update_handlers[update_name] = _BackgroundUpdateHandler(
            update_handler
        )

    def register_background_index_update(
        self,
        update_name: str,
        index_name: str,
        table: str,
        columns: Iterable[str],
        where_clause: Optional[str] = None,
        unique: bool = False,
        psql_only: bool = False,
        replaces_index: Optional[str] = None,
    ) -> None:
        """Helper for store classes to do a background index addition

        To use:

        1. use a schema delta file to add a background update. Example:
            INSERT INTO background_updates (update_name, progress_json) VALUES
                ('my_new_index', '{}');

        2. In the Store constructor, call this method

        Args:
            update_name: update_name to register for
            index_name: name of index to add
            table: table to add index to
            columns: columns/expressions to include in index
            where_clause: A WHERE clause to specify a partial unique index.
            unique: true to make a UNIQUE index
            psql_only: true to only create this index on psql databases (useful
                for virtual sqlite tables)
            replaces_index: The name of an index that this index replaces.
                The named index will be dropped upon completion of the new index.
        """

        async def updater(progress: JsonDict, batch_size: int) -> int:
            await self.create_index_in_background(
                index_name=index_name,
                table=table,
                columns=columns,
                where_clause=where_clause,
                unique=unique,
                psql_only=psql_only,
                replaces_index=replaces_index,
            )
            await self._end_background_update(update_name)
            return 1

        self._background_update_handlers[update_name] = _BackgroundUpdateHandler(
            updater, oneshot=True
        )

    def register_background_validate_constraint(
        self, update_name: str, constraint_name: str, table: str
    ) -> None:
        """Helper for store classes to do a background validate constraint.

        This only applies on PostgreSQL.

        To use:

        1. use a schema delta file to add a background update. Example:
            INSERT INTO background_updates (update_name, progress_json) VALUES
                ('validate_my_constraint', '{}');

        2. In the Store constructor, call this method

        Args:
            update_name: update_name to register for
            constraint_name: name of constraint to validate
            table: table the constraint is applied to
        """

        def runner(conn: Connection) -> None:
            c = conn.cursor()

            sql = f"""
            ALTER TABLE {table} VALIDATE CONSTRAINT {constraint_name};
            """
            logger.debug("[SQL] %s", sql)
            c.execute(sql)

        async def updater(progress: JsonDict, batch_size: int) -> int:
            assert isinstance(
                self.db_pool.engine, engines.PostgresEngine
            ), "validate constraint background update registered for non-Postres database"

            logger.info("Validating constraint %s to %s", constraint_name, table)
            await self.db_pool.runWithConnection(runner)
            await self._end_background_update(update_name)
            return 1

        self._background_update_handlers[update_name] = _BackgroundUpdateHandler(
            updater, oneshot=True
        )

    async def create_index_in_background(
        self,
        index_name: str,
        table: str,
        columns: Iterable[str],
        where_clause: Optional[str] = None,
        unique: bool = False,
        psql_only: bool = False,
        replaces_index: Optional[str] = None,
    ) -> None:
        """Add an index in the background.

        Args:
            update_name: update_name to register for
            index_name: name of index to add
            table: table to add index to
            columns: columns/expressions to include in index
            where_clause: A WHERE clause to specify a partial unique index.
            unique: true to make a UNIQUE index
            psql_only: true to only create this index on psql databases (useful
                for virtual sqlite tables)
            replaces_index: The name of an index that this index replaces.
                The named index will be dropped upon completion of the new index.
        """

        def create_index_psql(conn: Connection) -> None:
            conn.rollback()
            # postgres insists on autocommit for the index
            conn.set_session(autocommit=True)  # type: ignore

            try:
                c = conn.cursor()

                # If a previous attempt to create the index was interrupted,
                # we may already have a half-built index. Let's just drop it
                # before trying to create it again.

                sql = "DROP INDEX IF EXISTS %s" % (index_name,)
                logger.debug("[SQL] %s", sql)
                c.execute(sql)

                sql = (
                    "CREATE %(unique)s INDEX CONCURRENTLY %(name)s"
                    " ON %(table)s"
                    " (%(columns)s) %(where_clause)s"
                ) % {
                    "unique": "UNIQUE" if unique else "",
                    "name": index_name,
                    "table": table,
                    "columns": ", ".join(columns),
                    "where_clause": "WHERE " + where_clause if where_clause else "",
                }
                logger.debug("[SQL] %s", sql)
                c.execute(sql)

                if replaces_index is not None:
                    # We drop the old index as the new index has now been created.
                    sql = f"DROP INDEX IF EXISTS {replaces_index}"
                    logger.debug("[SQL] %s", sql)
                    c.execute(sql)
            finally:
                conn.set_session(autocommit=False)  # type: ignore

        def create_index_sqlite(conn: Connection) -> None:
            # Sqlite doesn't support concurrent creation of indexes.
            #
            # We assume that sqlite doesn't give us invalid indices; however
            # we may still end up with the index existing but the
            # background_updates not having been recorded if synapse got shut
            # down at the wrong moment - hance we use IF NOT EXISTS. (SQLite
            # has supported CREATE TABLE|INDEX IF NOT EXISTS since 3.3.0.)
            sql = (
                "CREATE %(unique)s INDEX IF NOT EXISTS %(name)s ON %(table)s"
                " (%(columns)s) %(where_clause)s"
            ) % {
                "unique": "UNIQUE" if unique else "",
                "name": index_name,
                "table": table,
                "columns": ", ".join(columns),
                "where_clause": "WHERE " + where_clause if where_clause else "",
            }

            c = conn.cursor()
            logger.debug("[SQL] %s", sql)
            c.execute(sql)

            if replaces_index is not None:
                # We drop the old index as the new index has now been created.
                sql = f"DROP INDEX IF EXISTS {replaces_index}"
                logger.debug("[SQL] %s", sql)
                c.execute(sql)

        if isinstance(self.db_pool.engine, engines.PostgresEngine):
            runner: Optional[Callable[[Connection], None]] = create_index_psql
        elif psql_only:
            runner = None
        else:
            runner = create_index_sqlite

        if runner is None:
            return

        logger.info("Adding index %s to %s", index_name, table)
        await self.db_pool.runWithConnection(runner)

    def register_background_validate_constraint_and_delete_rows(
        self,
        update_name: str,
        table: str,
        constraint_name: str,
        constraint: Constraint,
        unique_columns: Sequence[str],
    ) -> None:
        """Helper for store classes to do a background validate constraint, and
        delete rows that do not pass the constraint check.

        Note: This deletes rows that don't match the constraint. This may not be
        appropriate in all situations, and so the suitability of using this
        method should be considered on a case-by-case basis.

        This only applies on PostgreSQL.

        For SQLite the table gets recreated as part of the schema delta and the
        data is copied over synchronously (or whatever the correct way to
        describe it as).

        Args:
            update_name: The name of the background update.
            table: The table with the invalid constraint.
            constraint_name: The name of the constraint
            constraint: A `Constraint` object matching the type of constraint.
            unique_columns: A sequence of columns that form a unique constraint
              on the table. Used to iterate over the table.
        """

        assert isinstance(
            self.db_pool.engine, engines.PostgresEngine
        ), "validate constraint background update registered for non-Postres database"

        async def updater(progress: JsonDict, batch_size: int) -> int:
            return await self.validate_constraint_and_delete_in_background(
                update_name=update_name,
                table=table,
                constraint_name=constraint_name,
                constraint=constraint,
                unique_columns=unique_columns,
                progress=progress,
                batch_size=batch_size,
            )

        self._background_update_handlers[update_name] = _BackgroundUpdateHandler(
            updater, oneshot=True
        )

    async def validate_constraint_and_delete_in_background(
        self,
        update_name: str,
        table: str,
        constraint_name: str,
        constraint: Constraint,
        unique_columns: Sequence[str],
        progress: JsonDict,
        batch_size: int,
    ) -> int:
        """Validates a table constraint that has been marked as `NOT VALID`,
        deleting rows that don't pass the constraint check.

        This will delete rows that do not meet the validation check.

        update_name: str,
        table: str,
        constraint_name: str,
        constraint: Constraint,
        unique_columns: Sequence[str],
        """

        # We validate the constraint by:
        #   1. Trying to validate the constraint as is. If this succeeds then
        #      we're done.
        #   2. Otherwise, we manually scan the table to remove rows that don't
        #      match the constraint.
        #   3. We try re-validating the constraint.

        parsed_progress = ValidateConstraintProgress.parse_obj(progress)

        if parsed_progress.state == ValidateConstraintProgress.State.check:
            return_columns = ", ".join(unique_columns)
            order_columns = ", ".join(unique_columns)

            where_clause = ""
            args: List[Any] = []
            if parsed_progress.lower_bound:
                where_clause = f"""WHERE ({order_columns}) > ({", ".join("?" for _ in unique_columns)})"""
                args.extend(parsed_progress.lower_bound)

            args.append(batch_size)

            sql = f"""
                SELECT
                    {return_columns},
                    {constraint.make_check_clause(table)} AS check
                FROM {table}
                {where_clause}
                ORDER BY {order_columns}
                LIMIT ?
            """

            def validate_constraint_in_background_check(
                txn: "LoggingTransaction",
            ) -> None:
                txn.execute(sql, args)
                rows = txn.fetchall()

                new_progress = parsed_progress.copy()

                if not rows:
                    new_progress.state = ValidateConstraintProgress.State.validate
                    self._background_update_progress_txn(
                        txn, update_name, new_progress.dict()
                    )
                    return

                new_progress.lower_bound = rows[-1][:-1]

                to_delete = [row[:-1] for row in rows if not row[-1]]

                if to_delete:
                    logger.warning(
                        "Deleting %d rows that do not pass new constraint",
                        len(to_delete),
                    )

                    self.db_pool.simple_delete_many_batch_txn(
                        txn, table=table, keys=unique_columns, values=to_delete
                    )

                self._background_update_progress_txn(
                    txn, update_name, new_progress.dict()
                )

            await self.db_pool.runInteraction(
                "validate_constraint_in_background_check",
                validate_constraint_in_background_check,
            )

            return batch_size

        elif parsed_progress.state == ValidateConstraintProgress.State.validate:
            sql = f"ALTER TABLE {table} VALIDATE CONSTRAINT {constraint_name}"

            def validate_constraint_in_background_validate(
                txn: "LoggingTransaction",
            ) -> None:
                txn.execute(sql)

            try:
                await self.db_pool.runInteraction(
                    "validate_constraint_in_background_validate",
                    validate_constraint_in_background_validate,
                )

                await self._end_background_update(update_name)
            except self.db_pool.engine.module.IntegrityError as e:
                # If we get an integrity error here, then we go back and recheck the table.
                logger.warning("Integrity error when validating constraint: %s", e)
                await self._background_update_progress(
                    update_name,
                    ValidateConstraintProgress(
                        state=ValidateConstraintProgress.State.check
                    ).dict(),
                )

            return batch_size
        else:
            raise Exception(
                f"Unrecognized state '{parsed_progress.state}' when trying to validate_constraint_and_delete_in_background"
            )

    async def _end_background_update(self, update_name: str) -> None:
        """Removes a completed background update task from the queue.

        Args:
            update_name:: The name of the completed task to remove

        Returns:
            None, completes once the task is removed.
        """
        if update_name != self._current_background_update:
            raise Exception(
                "Cannot end background update %s which isn't currently running"
                % update_name
            )
        self._current_background_update = None
        await self.db_pool.simple_delete_one(
            "background_updates", keyvalues={"update_name": update_name}
        )

    async def _background_update_progress(
        self, update_name: str, progress: dict
    ) -> None:
        """Update the progress of a background update

        Args:
            update_name: The name of the background update task
            progress: The progress of the update.
        """

        await self.db_pool.runInteraction(
            "background_update_progress",
            self._background_update_progress_txn,
            update_name,
            progress,
        )

    def _background_update_progress_txn(
        self, txn: "LoggingTransaction", update_name: str, progress: JsonDict
    ) -> None:
        """Update the progress of a background update

        Args:
            txn: The transaction.
            update_name: The name of the background update task
            progress: The progress of the update.
        """

        progress_json = json_encoder.encode(progress)

        self.db_pool.simple_update_one_txn(
            txn,
            "background_updates",
            keyvalues={"update_name": update_name},
            updatevalues={"progress_json": progress_json},
        )


def run_validate_constraint_and_delete_rows_schema_delta(
    txn: "LoggingTransaction",
    ordering: int,
    update_name: str,
    table: str,
    constraint_name: str,
    constraint: Constraint,
    sqlite_table_name: str,
    sqlite_table_schema: str,
) -> None:
    """Runs a schema delta to add a constraint to the table. This should be run
    in a schema delta file.

    For PostgreSQL the constraint is added and validated in the background.

    For SQLite the table is recreated and data copied across immediately. This
    is done by the caller passing in a script to create the new table. Note that
    table indexes and triggers are copied over automatically.

    There must be a corresponding call to
    `register_background_validate_constraint_and_delete_rows` to register the
    background update in one of the data store classes.

    Attributes:
        txn ordering, update_name: For adding a row to background_updates table.
        table: The table to add constraint to. constraint_name: The name of the
        new constraint constraint: A `Constraint` object describing the
        constraint sqlite_table_name: For SQLite the name of the empty copy of
        table sqlite_table_schema: A SQL script for creating the above table.
    """

    if isinstance(txn.database_engine, PostgresEngine):
        # For postgres we can just add the constraint and mark it as NOT VALID,
        # and then insert a background update to go and check the validity in
        # the background.
        txn.execute(
            f"""
            ALTER TABLE {table}
            ADD CONSTRAINT {constraint_name} {constraint.make_constraint_clause_postgres()}
            NOT VALID
            """
        )

        txn.execute(
            "INSERT INTO background_updates (ordering, update_name, progress_json) VALUES (?, ?, '{}')",
            (ordering, update_name),
        )
    else:
        # For SQLite, we:
        #   1. fetch all indexes/triggers/etc related to the table
        #   2. create an empty copy of the table
        #   3. copy across the rows (that satisfy the check)
        #   4. replace the old table with the new able.
        #   5. add back all the indexes/triggers/etc

        # Fetch the indexes/triggers/etc. Note that `sql` column being null is
        # due to indexes being auto created based on the class definition (e.g.
        # PRIMARY KEY), and so don't need to be recreated.
        txn.execute(
            """
            SELECT sql FROM sqlite_master
            WHERE tbl_name = ? AND type != 'table' AND sql IS NOT NULL
            """,
            (table,),
        )
        extras = [row[0] for row in txn]

        txn.execute(sqlite_table_schema)

        sql = f"""
            INSERT INTO {sqlite_table_name} SELECT * FROM {table}
            WHERE {constraint.make_check_clause(table)}
        """

        txn.execute(sql)

        txn.execute(f"DROP TABLE {table}")
        txn.execute(f"ALTER TABLE {sqlite_table_name} RENAME TO {table}")

        for extra in extras:
            txn.execute(extra)