Why Gemfury? Push, build, and install  RubyGems npm packages Python packages Maven artifacts PHP packages Go Modules Debian packages RPM packages NuGet packages

Repository URL to install this package:

Details    
faust-streaming / faust / stores / rocksdb.py
Size: Mime:
"""RocksDB storage."""
import asyncio
import gc
import math
import shutil
import time
import typing
from collections import defaultdict
from datetime import datetime, timedelta
from pathlib import Path
from typing import (
    Any,
    Callable,
    DefaultDict,
    Dict,
    Iterable,
    Iterator,
    Mapping,
    MutableMapping,
    NamedTuple,
    Optional,
    Set,
    Tuple,
    Union,
    cast,
)

from mode.utils.collections import LRUCache
from yarl import URL

from faust.exceptions import ImproperlyConfigured
from faust.streams import current_event
from faust.types import TP, AppT, CollectionT, EventT
from faust.utils import platforms
from . import base

_max_open_files = platforms.max_open_files()
if _max_open_files is not None:  # pragma: no cover
    _max_open_files = math.ceil(_max_open_files * 0.90)
DEFAULT_MAX_OPEN_FILES = _max_open_files
DEFAULT_WRITE_BUFFER_SIZE = 67108864
DEFAULT_MAX_WRITE_BUFFER_NUMBER = 3
DEFAULT_TARGET_FILE_SIZE_BASE = 67108864
DEFAULT_BLOCK_CACHE_SIZE = 2 * 1024 ** 3
DEFAULT_BLOCK_CACHE_COMPRESSED_SIZE = 500 * 1024 ** 2
DEFAULT_BLOOM_FILTER_SIZE = 3

try:  # pragma: no cover
    import rocksdb
    from rocksdb.errors import Corruption
except ImportError:  # pragma: no cover
    rocksdb = None  # noqa

if typing.TYPE_CHECKING:  # pragma: no cover
    from rocksdb import DB, Options
else:

    class DB:  # noqa
        """Dummy DB."""

    class Options:  # noqa
        """Dummy Options."""


class PartitionDB(NamedTuple):
    """Tuple of ``(partition, rocksdb.DB)``."""

    partition: int
    db: DB


class _DBValueTuple(NamedTuple):
    db: DB
    value: bytes


class RocksDBOptions:
    """Options required to open a RocksDB database."""

    max_open_files: Optional[int] = DEFAULT_MAX_OPEN_FILES
    write_buffer_size: int = DEFAULT_WRITE_BUFFER_SIZE
    max_write_buffer_number: int = DEFAULT_MAX_WRITE_BUFFER_NUMBER
    target_file_size_base: int = DEFAULT_TARGET_FILE_SIZE_BASE
    block_cache_size: int = DEFAULT_BLOCK_CACHE_SIZE
    block_cache_compressed_size: int = DEFAULT_BLOCK_CACHE_COMPRESSED_SIZE
    bloom_filter_size: int = DEFAULT_BLOOM_FILTER_SIZE
    extra_options: Mapping

    def __init__(
        self,
        max_open_files: int = None,
        write_buffer_size: int = None,
        max_write_buffer_number: int = None,
        target_file_size_base: int = None,
        block_cache_size: int = None,
        block_cache_compressed_size: int = None,
        bloom_filter_size: int = None,
        **kwargs: Any,
    ) -> None:
        if max_open_files is not None:
            self.max_open_files = max_open_files
        if write_buffer_size is not None:
            self.write_buffer_size = write_buffer_size
        if max_write_buffer_number is not None:
            self.max_write_buffer_number = max_write_buffer_number
        if target_file_size_base is not None:
            self.target_file_size_base = target_file_size_base
        if block_cache_size is not None:
            self.block_cache_size = block_cache_size
        if block_cache_compressed_size is not None:
            self.block_cache_compressed_size = block_cache_compressed_size
        if bloom_filter_size is not None:
            self.bloom_filter_size = bloom_filter_size
        self.extra_options = kwargs

    def open(self, path: Path, *, read_only: bool = False) -> DB:
        """Open RocksDB database using this configuration."""
        return rocksdb.DB(str(path), self.as_options(), read_only=read_only)

    def as_options(self) -> Options:
        """Return :class:`rocksdb.Options` object using this configuration."""
        return rocksdb.Options(
            create_if_missing=True,
            max_open_files=self.max_open_files,
            write_buffer_size=self.write_buffer_size,
            max_write_buffer_number=self.max_write_buffer_number,
            target_file_size_base=self.target_file_size_base,
            table_factory=rocksdb.BlockBasedTableFactory(
                filter_policy=rocksdb.BloomFilterPolicy(self.bloom_filter_size),
                block_cache=rocksdb.LRUCache(self.block_cache_size),
                block_cache_compressed=rocksdb.LRUCache(
                    self.block_cache_compressed_size
                ),
            ),
            **self.extra_options,
        )


class Store(base.SerializedStore):
    """RocksDB table storage."""

    offset_key = b"__faust\0offset__"

    #: Decides the size of the K=>TopicPartition index (10_000).
    key_index_size: int

    #: Used to configure the RocksDB settings for table stores.
    rocksdb_options: RocksDBOptions

    _dbs: MutableMapping[int, DB]
    _key_index: LRUCache[bytes, int]
    rebalance_ack: bool
    db_lock: asyncio.Lock

    def __init__(
        self,
        url: Union[str, URL],
        app: AppT,
        table: CollectionT,
        *,
        key_index_size: int = None,
        options: Mapping[str, Any] = None,
        **kwargs: Any,
    ) -> None:
        if rocksdb is None:
            error = ImproperlyConfigured(
                "RocksDB bindings not installed? pip install python-rocksdb"
            )
            try:
                import rocksdb as _rocksdb  # noqa: F401
            except Exception as exc:  # pragma: no cover
                raise error from exc
            else:  # pragma: no cover
                raise error
        super().__init__(url, app, table, **kwargs)
        if not self.url.path:
            self.url /= self.table_name
        self.options = options or {}
        self.rocksdb_options = RocksDBOptions(**self.options)
        if key_index_size is None:
            key_index_size = app.conf.table_key_index_size
        self.key_index_size = key_index_size
        self._dbs = {}
        self._last_checkpoints = {}
        self.checkpoint_interval = timedelta(seconds=app.conf.checkpoint_interval)
        self._key_index = LRUCache(limit=self.key_index_size)
        self.db_lock = asyncio.Lock()
        self.rebalance_ack = False

    def persisted_offset(self, tp: TP) -> Optional[int]:
        """Return the last persisted offset.

        See :meth:`set_persisted_offset`.
        """
        offset = self._db_for_partition(tp.partition).get(self.offset_key)
        if offset is not None:
            return int(offset)
        return None

    def set_persisted_offset(self, tp: TP, offset: int) -> None:
        """Set the last persisted offset for this table.

        This will remember the last offset that we wrote to RocksDB,
        so that on rebalance/recovery we can seek past this point
        to only read the events that occurred recently while
        we were not an active replica.
        """
        db = self._db_for_partition(tp.partition)
        db.put(self.offset_key, str(offset).encode())

        last_checkpointed = self._last_checkpoints.get(tp)
        if not last_checkpointed or (datetime.utcnow() - last_checkpointed) > self.checkpoint_interval:
            start = time.time()
            backup_engine = rocksdb.BackupEngine(str(self.backup_path(tp.partition)))
            backup_engine.create_backup(db)
            backup_engine.purge_old_backups(2)
            self._last_checkpoints[tp] = datetime.utcnow()
            elapsed = time.time() - start
            self.logger.info(f"Checkpointed rocksdb state for {tp}, took: {elapsed}")

    async def need_active_standby_for(self, tp: TP) -> bool:
        """Decide if an active standby is needed for this topic partition.

        Since other workers may be running on the same local machine,
        we can decide to not actively read standby messages, since
        that database file is already being populated.

        Currently it is recommended that you use
        separate data directories for multiple workers on the same machine.

        For example if you have a 4 CPU core machine, you can run
        four worker instances on that machine, but using separate
        data directories:

        .. sourcecode:: console

            $ myproj --datadir=/var/faust/w1 worker -l info --web-port=6066
            $ myproj --datadir=/var/faust/w2 worker -l info --web-port=6067
            $ myproj --datadir=/var/faust/w3 worker -l info --web-port=6068
            $ myproj --datadir=/var/faust/w4 worker -l info --web-port=6069
        """
        try:
            self._db_for_partition(tp.partition)
        except rocksdb.errors.RocksIOError as exc:
            if "lock" not in repr(exc):
                raise
            return False
        else:
            return True

    def apply_changelog_batch(
        self,
        batch: Iterable[EventT],
        to_key: Callable[[Any], Any],
        to_value: Callable[[Any], Any],
    ) -> None:
        """Write batch of changelog events to local RocksDB storage.

        Arguments:
            batch: Iterable of changelog events (:class:`faust.Event`)
            to_key: A callable you can use to deserialize the key
                of a changelog event.
            to_value: A callable you can use to deserialize the value
                of a changelog event.
        """
        batches: DefaultDict[int, rocksdb.WriteBatch]
        batches = defaultdict(rocksdb.WriteBatch)
        tp_offsets: Dict[TP, int] = {}
        for event in batch:
            tp, offset = event.message.tp, event.message.offset
            tp_offsets[tp] = (
                offset if tp not in tp_offsets else max(offset, tp_offsets[tp])
            )
            msg = event.message
            if msg.value is None:
                batches[msg.partition].delete(msg.key)
            else:
                batches[msg.partition].put(msg.key, msg.value)

        for partition, batch in batches.items():
            self._db_for_partition(partition).write(batch)

        for tp, offset in tp_offsets.items():
            self.set_persisted_offset(tp, offset)

    def _set(self, key: bytes, value: Optional[bytes]) -> None:
        event = current_event()
        assert event is not None
        partition = event.message.partition
        db = self._db_for_partition(partition)
        self._key_index[key] = partition
        db.put(key, value)

    def _db_for_partition(self, partition: int) -> DB:
        try:
            return self._dbs[partition]
        except KeyError:
            db = self._dbs[partition] = self._open_for_partition(partition)
            return db

    def _open_for_partition(self, partition: int) -> DB:
        db_path = self.partition_path(partition)
        try:
            return self.rocksdb_options.open(db_path)
        except Corruption:  # db file corrupted, clear and reopen
            self.log.exception("RocksDB files corrupted: {}, remove corrupted files and re-open", db_path)
            shutil.rmtree(db_path)
            return self.rocksdb_options.open(db_path)

    def _get(self, key: bytes) -> Optional[bytes]:
        event = current_event()
        partition_from_message = (
            event is not None
            and not self.table.is_global
            and not self.table.use_partitioner
        )
        if partition_from_message:
            partition = event.message.partition
            db = self._db_for_partition(partition)
            value = db.get(key)
            if value is not None:
                self._key_index[key] = partition
            return value
        else:
            dbvalue = self._get_bucket_for_key(key)
            if dbvalue is None:
                return None
            db, value = dbvalue

            if value is None:
                if db.key_may_exist(key)[0]:
                    return db.get(key)
            return value

    def _get_bucket_for_key(self, key: bytes) -> Optional[_DBValueTuple]:
        dbs: Iterable[PartitionDB]
        try:
            partition = self._key_index[key]
            dbs = [PartitionDB(partition, self._dbs[partition])]
        except KeyError:
            dbs = cast(Iterable[PartitionDB], self._dbs.items())

        for partition, db in dbs:
            if db.key_may_exist(key)[0]:
                value = db.get(key)
                if value is not None:
                    self._key_index[key] = partition
                    return _DBValueTuple(db, value)
        return None

    def _del(self, key: bytes) -> None:
        for db in self._dbs_for_key(key):
            db.delete(key)

    async def on_rebalance(
        self,
        assigned: Set[TP],
        revoked: Set[TP],
        newly_assigned: Set[TP],
        generation_id: int = 0,
    ) -> None:
        """Rebalance occurred.

        Arguments:
            assigned: Set of all assigned topic partitions.
            revoked: Set of newly revoked topic partitions.
            newly_assigned: Set of newly assigned topic partitions,
                for which we were not assigned the last time.
            generation_id: the metadata generation identifier for the re-balance
        """
        self.rebalance_ack = False
        async with self.db_lock:
            self.revoke_partitions(self.table, revoked)
            await self.assign_partitions(self.table, newly_assigned, generation_id)

    async def stop(self) -> None:
        self.logger.info("Closing rocksdb on stop")
        # for db in self._dbs.values():
        #     db.close()
        self._dbs.clear()
        gc.collect()

    def revoke_partitions(self, table: CollectionT, tps: Set[TP]) -> None:
        """De-assign partitions used on this worker instance.

        Arguments:
            table: The table that we store data for.
            tps: Set of topic partitions that we should no longer
                be serving data for.
        """
        standby_tps = self.app.assignor.assigned_standbys()

        for tp in tps:
            if tp in standby_tps:
                self._recover_db(tp, True)

            if tp.topic in table.changelog_topic.topics:
                db = self._dbs.pop(tp.partition, None)
                self._last_checkpoints.pop(tp.partition, None)
                if db is not None:
                    self.logger.info(f"closing db {tp.topic} partition {tp.partition}")
                    # db.close()

        gc.collect()

    async def assign_partitions(
        self, table: CollectionT, tps: Set[TP], generation_id: int = 0
    ) -> None:
        """Assign partitions to this worker instance.

        Arguments:
            table: The table that we store data for.
            tps: Set of topic partitions we have been assigned.
        """
        self.rebalance_ack = True
        standby_tps = self.app.assignor.assigned_standbys()
        my_topics = table.changelog_topic.topics
        for tp in tps:
            if tp.topic in my_topics:
                is_standby = tp in standby_tps
                self._recover_db(tp, is_standby)

                if not is_standby and self.rebalance_ack:
                    await self._try_open_db_for_partition(
                        tp.partition, generation_id=generation_id
                    )
                await asyncio.sleep(0)

    def _recover_db(self, tp: TP, is_standby: bool) -> None:
        try:
            del self._dbs[tp.partition]  # db needs to be re-opened
        except KeyError:
            pass

        db_path = self.partition_path(tp.partition)
        backup_engine = rocksdb.BackupEngine(str(self.backup_path(tp.partition)))
        backups = backup_engine.get_backup_info()
        tp_mode = "standby" if is_standby else "active"

        if backups:
            start = time.time()
            backup_engine.restore_latest_backup(str(db_path), str(db_path))
            backup_info = backups[-1]
            backup_id = backup_info['backup_id']
            backup_time = datetime.fromtimestamp(backup_info['timestamp'])
            elapsed = time.time() - start
            self.logger.info(f"Recovered rocksdb state for {tp_mode} {tp} "
                             f"from backup {backup_id}, {backup_time}, took: {elapsed}")

        elif db_path.exists():
            shutil.rmtree(db_path)
            self.logger.info(f"Cleared rocksdb state for {tp} as no stable backup found")

    async def _try_open_db_for_partition(
        self,
        partition: int,
        max_retries: int = 30,
        retry_delay: float = 1.0,
        generation_id: int = 0,
    ) -> DB:
        for i in range(max_retries):
            try:
                # side effect: opens db and adds to self._dbs.
                self.logger.info(
                    f"opening partition {partition} for gen id "
                    f"{generation_id} app id {self.app.consumer_generation_id}"
                )
                return self._db_for_partition(partition)
            except rocksdb.errors.RocksIOError as exc:
                if i == max_retries - 1 or "lock" not in repr(exc):
                    # release all the locks and crash
                    self.log.warning(
                        "DB for partition %r retries timed out ", partition
                    )
                    await self.stop()
                    raise
                self.log.info(
                    "DB for partition %r is locked! Retry in 1s...", partition
                )
                if generation_id != self.app.consumer_generation_id:
                    self.log.info(
                        f"Rebalanced again giving up partition {partition} gen id"
                        f" {generation_id} app {self.app.consumer_generation_id}"
                    )
                    return
                await self.sleep(retry_delay)
        else:  # pragma: no cover
            ...

    def _contains(self, key: bytes) -> bool:
        event = current_event()
        partition_from_message = (
            event is not None
            and not self.table.is_global
            and not self.table.use_partitioner
        )
        if partition_from_message:
            partition = event.message.partition
            db = self._db_for_partition(partition)
            value = db.get(key)
            if value is not None:
                return True
            else:
                return False
        else:
            for db in self._dbs_for_key(key):
                # bloom filter: false positives possible, but not false negatives
                if db.key_may_exist(key)[0] and db.get(key) is not None:
                    return True
            return False

    def _dbs_for_key(self, key: bytes) -> Iterable[DB]:
        # Returns cached db if key is in index, otherwise all dbs
        # for linear search.
        try:
            return [self._dbs[self._key_index[key]]]
        except KeyError:
            return self._dbs.values()

    def _dbs_for_actives(self) -> Iterator[DB]:
        actives = self.app.assignor.assigned_actives()
        topic = self.table.changelog_topic_name
        for partition, db in self._dbs.items():
            tp = TP(topic=topic, partition=partition)
            # for global tables, keys from all
            # partitions are available.
            if tp in actives or self.table.is_global:
                yield db

    def _size(self) -> int:
        return sum(self._size1(db) for db in self._dbs_for_actives())

    def _visible_keys(self, db: DB) -> Iterator[bytes]:
        it = db.iterkeys()  # noqa: B301
        it.seek_to_first()
        for key in it:
            if key != self.offset_key:
                yield key

    def _visible_items(self, db: DB) -> Iterator[Tuple[bytes, bytes]]:
        it = db.iteritems()  # noqa: B301
        it.seek_to_first()
        for key, value in it:
            if key != self.offset_key:
                yield key, value

    def _visible_values(self, db: DB) -> Iterator[bytes]:
        for _, value in self._visible_items(db):
            yield value

    def _size1(self, db: DB) -> int:
        return sum(1 for _ in self._visible_keys(db))

    def _iterkeys(self) -> Iterator[bytes]:
        for db in self._dbs_for_actives():
            yield from self._visible_keys(db)

    def _itervalues(self) -> Iterator[bytes]:
        for db in self._dbs_for_actives():
            yield from self._visible_values(db)

    def _iteritems(self) -> Iterator[Tuple[bytes, bytes]]:
        for db in self._dbs_for_actives():
            yield from self._visible_items(db)

    def _clear(self) -> None:
        raise NotImplementedError("TODO")  # XXX cannot reset tables

    def reset_state(self) -> None:
        """Remove all data stored in this table.

        Notes:
            Only local data will be removed, table changelog partitions
            in Kafka will not be affected.
        """
        self._dbs.clear()
        self._last_checkpoints.clear()
        self._key_index.clear()
        # with suppress(FileNotFoundError):
        #     shutil.(self.path.absolute())

    def partition_path(self, partition: int) -> Path:
        """Return :class:`pathlib.Path` to db file of specific partition."""
        p = self.path / self.basename
        return self._path_with_suffix(p.with_name(f"{p.name}-{partition}"))

    def backup_path(self, partition: int) -> Path:
        """Return :class:`pathlib.Path` to db file of specific partition."""
        p = self.path / self.basename
        return self._path_with_suffix(p.with_name(f"{p.name}-{partition}"), suffix=".bak")

    def _path_with_suffix(self, path: Path, *, suffix: str = ".db") -> Path:
        # Path.with_suffix should not be used as this will
        # not work if the table name has dots in it (Issue #184).
        return path.with_name(f"{path.name}{suffix}")

    @property
    def path(self) -> Path:
        """Path to directory where tables are stored.

        See Also:
            :setting:`tabledir` (default value for this path).

        Returns:
            :class:`pathlib.Path`.
        """
        return self.app.conf.tabledir

    @property
    def basename(self) -> Path:
        """Return the name of this table, used as filename prefix."""
        return Path(self.url.path)