Source code for faust.tables.manager
"""Tables (changelog stream)."""
import asyncio
from typing import Any, MutableMapping, Optional, Set, Tuple
from mode import Service
from mode.utils.queues import ThrowableQueue
from faust.types import AppT, ChannelT, StoreT, TP
from faust.types.tables import CollectionT, TableManagerT
from faust.utils.tracing import traced_from_parent_span
from .recovery import Recovery
__all__ = [
'TableManager',
]
[docs]class TableManager(Service, TableManagerT):
"""Manage tables used by Faust worker."""
_channels: MutableMapping[CollectionT, ChannelT]
_changelogs: MutableMapping[str, CollectionT]
_recovery_started: asyncio.Event
_changelog_queue: ThrowableQueue
_pending_persisted_offsets: MutableMapping[TP, Tuple[StoreT, int]]
_recovery: Optional[Recovery] = None
def __init__(self, app: AppT, **kwargs: Any) -> None:
super().__init__(**kwargs)
self.app = app
self.data: MutableMapping = {}
self._changelog_queue = None
self._channels = {}
self._changelogs = {}
self._recovery_started = asyncio.Event(loop=self.loop)
self.actives_ready = False
self.standbys_ready = False
self._pending_persisted_offsets = {}
[docs] def persist_offset_on_commit(self,
store: StoreT,
tp: TP,
offset: int) -> None:
"""Mark the persisted offset for a TP to be saved on commit.
This is used for "exactly_once" processing guarantee.
Instead of writing the persisted offset to RocksDB when the message
is sent, we write it to disk when the offset is committed.
"""
existing_entry = self._pending_persisted_offsets.get(tp)
if existing_entry is not None:
_, existing_offset = existing_entry
if offset < existing_offset:
return
self._pending_persisted_offsets[tp] = (store, offset)
[docs] def on_commit(self, offsets: MutableMapping[TP, int]) -> None:
"""Call when committing source topic partitions."""
# flush any pending persisted offsets added by
# persist_offset_on_commit
for tp in offsets:
self.on_commit_tp(tp)
[docs] def on_commit_tp(self, tp: TP) -> None:
"""Call when committing source topic partition used by this table."""
entry = self._pending_persisted_offsets.get(tp)
if entry is not None:
store, offset = entry
store.set_persisted_offset(tp, offset)
[docs] def on_rebalance_start(self) -> None:
"""Call when a new rebalancing operation starts."""
self.actives_ready = False
self.standbys_ready = False
[docs] def on_actives_ready(self) -> None:
"""Call when actives are fully up-to-date."""
self.actives_ready = True
[docs] def on_standbys_ready(self) -> None:
"""Call when standbys are fully up-to-date and ready for failover."""
self.standbys_ready = True
def __hash__(self) -> int:
return object.__hash__(self)
@property
def changelog_topics(self) -> Set[str]:
"""Return the set of known changelog topics."""
return set(self._changelogs.keys())
@property
def changelog_queue(self) -> ThrowableQueue:
"""Queue used to buffer changelog events."""
if self._changelog_queue is None:
self._changelog_queue = self.app.FlowControlQueue(
maxsize=self.app.conf.stream_buffer_maxsize,
loop=self.loop,
clear_on_resume=True,
)
return self._changelog_queue
@property
def recovery(self) -> Recovery:
"""Recovery service used by this table manager."""
if self._recovery is None:
self._recovery = Recovery(
self.app, self, beacon=self.beacon, loop=self.loop)
return self._recovery
[docs] def add(self, table: CollectionT) -> CollectionT:
"""Add table to be managed by this table manager."""
if self._recovery_started.is_set():
raise RuntimeError('Too late to add tables at this point')
assert table.name is not None
if table.name in self:
raise ValueError(f'Table with name {table.name!r} already exists')
self[table.name] = table
self._changelogs[table.changelog_topic.get_topic_name()] = table
return table
[docs] async def on_start(self) -> None:
"""Call when table manager is starting."""
await self.sleep(1.0)
if not self.should_stop:
await self._update_channels()
await self.recovery.start()
async def _update_channels(self) -> None:
for table in self.values():
if table not in self._channels:
chan = table.changelog_topic.clone_using_queue(
self.changelog_queue)
self.app.topics.add(chan)
self._channels[table] = chan
await table.maybe_start()
self.app.consumer.pause_partitions({
tp for tp in self.app.consumer.assignment()
if tp.topic in self._changelogs
})
[docs] async def on_stop(self) -> None:
"""Call when table manager is stopping."""
await self.app._fetcher.stop()
if self._recovery:
await self._recovery.stop()
for table in self.values():
await table.stop()
[docs] def on_partitions_revoked(self, revoked: Set[TP]) -> None:
"""Call when cluster is rebalancing and partitions revoked."""
T = traced_from_parent_span()
T(self.recovery.on_partitions_revoked)(revoked)
[docs] async def on_rebalance(self,
assigned: Set[TP],
revoked: Set[TP],
newly_assigned: Set[TP]) -> None:
"""Call when the cluster is rebalancing."""
self._recovery_started.set() # cannot add more tables.
T = traced_from_parent_span()
for table in self.values():
await T(table.on_rebalance)(assigned, revoked, newly_assigned)
await T(self._update_channels)()
await T(self.recovery.on_rebalance)(assigned, revoked, newly_assigned)