import asyncio import logging import threading from typing import Any, Callable, List, Optional from redis.asyncio.multidb.healthcheck import HealthCheck, HealthCheckPolicy from redis.background import BackgroundScheduler from redis.backoff import NoBackoff from redis.client import PubSubWorkerThread from redis.commands import CoreCommands, RedisModuleCommands from redis.maint_notifications import MaintNotificationsConfig from redis.multidb.circuit import CircuitBreaker from redis.multidb.circuit import State as CBState from redis.multidb.command_executor import DefaultCommandExecutor from redis.multidb.config import ( DEFAULT_GRACE_PERIOD, DatabaseConfig, InitialHealthCheck, MultiDbConfig, ) from redis.multidb.database import Database, Databases, SyncDatabase from redis.multidb.exception import ( InitialHealthCheckFailedError, NoValidDatabaseException, UnhealthyDatabaseException, ) from redis.multidb.failure_detector import FailureDetector from redis.observability.attributes import GeoFailoverReason from redis.retry import Retry from redis.utils import experimental logger = logging.getLogger(__name__) @experimental class MultiDBClient(RedisModuleCommands, CoreCommands): """ Client that operates on multiple logical Redis databases. Should be used in Client-side geographic failover database setups. """ def __init__(self, config: MultiDbConfig): self._databases = config.databases() self._health_checks = ( config.default_health_checks() if not config.health_checks else config.health_checks ) self._health_check_interval = config.health_check_interval self._health_check_policy: HealthCheckPolicy = ( config.health_check_policy.value() ) self._failure_detectors = ( config.default_failure_detectors() if not config.failure_detectors else config.failure_detectors ) self._failover_strategy = ( config.default_failover_strategy() if config.failover_strategy is None else config.failover_strategy ) self._failover_strategy.set_databases(self._databases) self._auto_fallback_interval = config.auto_fallback_interval self._event_dispatcher = config.event_dispatcher self._command_retry = config.command_retry self._command_retry.update_supported_errors((ConnectionRefusedError,)) self.command_executor = DefaultCommandExecutor( failure_detectors=self._failure_detectors, databases=self._databases, command_retry=self._command_retry, failover_strategy=self._failover_strategy, failover_attempts=config.failover_attempts, failover_delay=config.failover_delay, event_dispatcher=self._event_dispatcher, auto_fallback_interval=self._auto_fallback_interval, ) self.initialized = False self._bg_scheduler = BackgroundScheduler() self._hc_lock = threading.Lock() self._config = config def __del__(self): try: self.close() except Exception: # Suppress exceptions during garbage collection. # close() may fail if called during interpreter shutdown # or while an event loop is already running. pass def initialize(self): """ Perform initialization of databases to define their initial state. """ # Initial databases check to define initial state. # Uses run_coro_sync to run in the shared background loop - this ensures # connection pools created during initial health check remain valid for # subsequent recurring health checks (they use the same event loop). self._bg_scheduler.run_coro_sync(self._perform_initial_health_check) # Starts recurring health checks on the background. # Uses run_recurring_coro which shares the same event loop as run_coro_sync self._bg_scheduler.run_recurring_coro( self._health_check_interval, self._check_databases_health, ) is_active_db_found = False for database, weight in self._databases: # Set on state changed callback for each circuit. database.circuit.on_state_changed(self._on_circuit_state_change_callback) # Set states according to a weights and circuit state if database.circuit.state == CBState.CLOSED and not is_active_db_found: # Directly set the active database during initialization # without recording a geo failover metric self.command_executor._active_database = database is_active_db_found = True if not is_active_db_found: raise NoValidDatabaseException( "Initial connection failed - no active database found" ) self.initialized = True def get_databases(self) -> Databases: """ Returns a sorted (by weight) list of all databases. """ return self._databases def set_active_database(self, database: SyncDatabase) -> None: """ Promote one of the existing databases to become an active. """ exists = None for existing_db, _ in self._databases: if existing_db == database: exists = True break if not exists: raise ValueError("Given database is not a member of database list") self._bg_scheduler.run_coro_sync(self._check_db_health, database) if database.circuit.state == CBState.CLOSED: highest_weighted_db, _ = self._databases.get_top_n(1)[0] self.command_executor.active_database = ( database, GeoFailoverReason.MANUAL, ) return raise NoValidDatabaseException( "Cannot set active database, database is unhealthy" ) def add_database( self, config: DatabaseConfig, skip_initial_health_check: bool = True ): """ Adds a new database to the database list. Args: config: DatabaseConfig object that contains the database configuration. skip_initial_health_check: If True, adds the database even if it is unhealthy. """ # The retry object is not used in the lower level clients, so we can safely remove it. # We rely on command_retry in terms of global retries. config.client_kwargs["retry"] = Retry(retries=0, backoff=NoBackoff()) # Maintenance notifications are disabled by default in underlying clients, # but user can override this by providing their own config. if "maint_notifications_config" not in config.client_kwargs: config.client_kwargs["maint_notifications_config"] = ( MaintNotificationsConfig(enabled=False) ) if config.from_url: client = self._config.client_class.from_url( config.from_url, **config.client_kwargs ) elif config.from_pool: config.from_pool.set_retry(Retry(retries=0, backoff=NoBackoff())) client = self._config.client_class.from_pool( connection_pool=config.from_pool ) else: client = self._config.client_class(**config.client_kwargs) circuit = ( config.default_circuit_breaker() if config.circuit is None else config.circuit ) database = Database( client=client, circuit=circuit, weight=config.weight, health_check_url=config.health_check_url, ) try: self._bg_scheduler.run_coro_sync(self._check_db_health, database) except UnhealthyDatabaseException: if not skip_initial_health_check: raise highest_weighted_db, highest_weight = self._databases.get_top_n(1)[0] self._databases.add(database, database.weight) self._change_active_database(database, highest_weighted_db) def _change_active_database( self, new_database: SyncDatabase, highest_weight_database: SyncDatabase ): if ( new_database.weight > highest_weight_database.weight and new_database.circuit.state == CBState.CLOSED ): self.command_executor.active_database = ( new_database, GeoFailoverReason.AUTOMATIC, ) def remove_database(self, database: Database): """ Removes a database from the database list. """ weight = self._databases.remove(database) highest_weighted_db, highest_weight = self._databases.get_top_n(1)[0] if ( highest_weight <= weight and highest_weighted_db.circuit.state == CBState.CLOSED ): self.command_executor.active_database = ( highest_weighted_db, GeoFailoverReason.MANUAL, ) def update_database_weight(self, database: SyncDatabase, weight: float): """ Updates a database from the database list. """ exists = None for existing_db, _ in self._databases: if existing_db == database: exists = True break if not exists: raise ValueError("Given database is not a member of database list") highest_weighted_db, highest_weight = self._databases.get_top_n(1)[0] self._databases.update_weight(database, weight) database.weight = weight self._change_active_database(database, highest_weighted_db) def add_failure_detector(self, failure_detector: FailureDetector): """ Adds a new failure detector to the database. """ self._failure_detectors.append(failure_detector) def add_health_check(self, healthcheck: HealthCheck): """ Adds a new health check to the database. """ with self._hc_lock: self._health_checks.append(healthcheck) def execute_command(self, *args, **options): """ Executes a single command and return its result. """ if not self.initialized: self.initialize() return self.command_executor.execute_command(*args, **options) def pipeline(self): """ Enters into pipeline mode of the client. """ return Pipeline(self) def transaction(self, func: Callable[["Pipeline"], None], *watches, **options): """ Executes callable as transaction. """ if not self.initialized: self.initialize() return self.command_executor.execute_transaction(func, *watches, *options) def pubsub(self, **kwargs): """ Return a Publish/Subscribe object. With this object, you can subscribe to channels and listen for messages that get published to them. """ if not self.initialized: self.initialize() return PubSub(self, **kwargs) async def _check_db_health(self, database: SyncDatabase) -> bool: """ Runs health checks on the given database until first failure. """ with self._hc_lock: health_checks = list(self._health_checks) # Health check will setup circuit state is_healthy = await self._health_check_policy.execute(health_checks, database) if not is_healthy: if database.circuit.state != CBState.OPEN: database.circuit.state = CBState.OPEN return is_healthy elif is_healthy and database.circuit.state != CBState.CLOSED: database.circuit.state = CBState.CLOSED return is_healthy async def _check_databases_health(self) -> dict[Database, bool]: """ Runs health checks as a recurring task. Runs health checks against all databases. """ task_to_db: dict[asyncio.Task, Database] = {} self._hc_tasks = [] for database, _ in self._databases: task = asyncio.create_task(self._check_db_health(database)) task_to_db[task] = database self._hc_tasks.append(task) results = await asyncio.gather(*self._hc_tasks, return_exceptions=True) # Map end results to databases db_results = { task_to_db[task]: result for task, result in zip(self._hc_tasks, results) } for database, result in db_results.items(): if isinstance(result, UnhealthyDatabaseException): unhealthy_db = result.database unhealthy_db.circuit.state = CBState.OPEN logger.debug( "Health check failed, due to exception", exc_info=result.original_exception, ) db_results[unhealthy_db] = False return db_results async def _perform_initial_health_check(self): """ Runs initial health check and evaluate healthiness based on initial_health_check_policy. """ results = await self._check_databases_health() is_healthy = True if self._config.initial_health_check_policy == InitialHealthCheck.ALL_AVAILABLE: is_healthy = False not in results.values() elif ( self._config.initial_health_check_policy == InitialHealthCheck.MAJORITY_AVAILABLE ): is_healthy = sum(results.values()) > len(results) / 2 elif ( self._config.initial_health_check_policy == InitialHealthCheck.ONE_AVAILABLE ): is_healthy = True in results.values() if not is_healthy: raise InitialHealthCheckFailedError( f"Initial health check failed. Initial health check policy: {self._config.initial_health_check_policy}" ) def _on_circuit_state_change_callback( self, circuit: CircuitBreaker, old_state: CBState, new_state: CBState ): if new_state == CBState.HALF_OPEN: self._bg_scheduler.run_coro_fire_and_forget( self._check_db_health, circuit.database ) return if old_state == CBState.CLOSED and new_state == CBState.OPEN: logger.warning( f"Database {circuit.database} is unreachable. Failover has been initiated." ) self._bg_scheduler.run_once( DEFAULT_GRACE_PERIOD, _half_open_circuit, circuit ) if old_state != CBState.CLOSED and new_state == CBState.CLOSED: logger.info(f"Database {circuit.database} is reachable again.") def close(self): """ Closes the client and all its resources. """ # Close health check policy BEFORE stopping the scheduler. # The policy's connection pools were created on the shared health check # event loop, so they must be disconnected on that same loop to avoid # leaking sockets/file descriptors. if self._bg_scheduler: try: self._bg_scheduler.run_coro_sync(self._health_check_policy.close) except Exception: pass self._bg_scheduler.stop() if self.command_executor.active_database: self.command_executor.active_database.client.close() def _half_open_circuit(circuit: CircuitBreaker): circuit.state = CBState.HALF_OPEN class Pipeline(RedisModuleCommands, CoreCommands): """ Pipeline implementation for multiple logical Redis databases. """ def __init__(self, client: MultiDBClient): self._command_stack = [] self._client = client def __enter__(self) -> "Pipeline": return self def __exit__(self, exc_type, exc_value, traceback): self.reset() def __del__(self): try: self.reset() except Exception: pass def __len__(self) -> int: return len(self._command_stack) def __bool__(self) -> bool: """Pipeline instances should always evaluate to True""" return True def reset(self) -> None: self._command_stack = [] def close(self) -> None: """Close the pipeline""" self.reset() def pipeline_execute_command(self, *args, **options) -> "Pipeline": """ Stage a command to be executed when execute() is next called Returns the current Pipeline object back so commands can be chained together, such as: pipe = pipe.set('foo', 'bar').incr('baz').decr('bang') At some other point, you can then run: pipe.execute(), which will execute all commands queued in the pipe. """ self._command_stack.append((args, options)) return self def execute_command(self, *args, **kwargs): """Adds a command to the stack""" return self.pipeline_execute_command(*args, **kwargs) def execute(self) -> List[Any]: """Execute all the commands in the current pipeline""" if not self._client.initialized: self._client.initialize() try: return self._client.command_executor.execute_pipeline( tuple(self._command_stack) ) finally: self.reset() class PubSub: """ PubSub object for multi database client. """ def __init__(self, client: MultiDBClient, **kwargs): """Initialize the PubSub object for a multi-database client. Args: client: MultiDBClient instance to use for pub/sub operations **kwargs: Additional keyword arguments to pass to the underlying pubsub implementation """ self._client = client self._client.command_executor.pubsub(**kwargs) def __enter__(self) -> "PubSub": return self def __del__(self) -> None: try: # if this object went out of scope prior to shutting down # subscriptions, close the connection manually before # returning it to the connection pool self.reset() except Exception: pass def reset(self) -> None: return self._client.command_executor.execute_pubsub_method("reset") def close(self) -> None: self.reset() @property def subscribed(self) -> bool: return self._client.command_executor.active_pubsub.subscribed def execute_command(self, *args): return self._client.command_executor.execute_pubsub_method( "execute_command", *args ) def psubscribe(self, *args, **kwargs): """ Subscribe to channel patterns. Patterns supplied as keyword arguments expect a pattern name as the key and a callable as the value. A pattern's callable will be invoked automatically when a message is received on that pattern rather than producing a message via ``listen()``. """ return self._client.command_executor.execute_pubsub_method( "psubscribe", *args, **kwargs ) def punsubscribe(self, *args): """ Unsubscribe from the supplied patterns. If empty, unsubscribe from all patterns. """ return self._client.command_executor.execute_pubsub_method( "punsubscribe", *args ) def subscribe(self, *args, **kwargs): """ Subscribe to channels. Channels supplied as keyword arguments expect a channel name as the key and a callable as the value. A channel's callable will be invoked automatically when a message is received on that channel rather than producing a message via ``listen()`` or ``get_message()``. """ return self._client.command_executor.execute_pubsub_method( "subscribe", *args, **kwargs ) def unsubscribe(self, *args): """ Unsubscribe from the supplied channels. If empty, unsubscribe from all channels """ return self._client.command_executor.execute_pubsub_method("unsubscribe", *args) def ssubscribe(self, *args, **kwargs): """ Subscribes the client to the specified shard channels. Channels supplied as keyword arguments expect a channel name as the key and a callable as the value. A channel's callable will be invoked automatically when a message is received on that channel rather than producing a message via ``listen()`` or ``get_sharded_message()``. """ return self._client.command_executor.execute_pubsub_method( "ssubscribe", *args, **kwargs ) def sunsubscribe(self, *args): """ Unsubscribe from the supplied shard_channels. If empty, unsubscribe from all shard_channels """ return self._client.command_executor.execute_pubsub_method( "sunsubscribe", *args ) def get_message( self, ignore_subscribe_messages: bool = False, timeout: float = 0.0 ): """ Get the next message if one is available, otherwise None. If timeout is specified, the system will wait for `timeout` seconds before returning. Timeout should be specified as a floating point number, or None, to wait indefinitely. """ return self._client.command_executor.execute_pubsub_method( "get_message", ignore_subscribe_messages=ignore_subscribe_messages, timeout=timeout, ) def get_sharded_message( self, ignore_subscribe_messages: bool = False, timeout: float = 0.0 ): """ Get the next message if one is available in a sharded channel, otherwise None. If timeout is specified, the system will wait for `timeout` seconds before returning. Timeout should be specified as a floating point number, or None, to wait indefinitely. """ return self._client.command_executor.execute_pubsub_method( "get_sharded_message", ignore_subscribe_messages=ignore_subscribe_messages, timeout=timeout, ) def run_in_thread( self, sleep_time: float = 0.0, daemon: bool = False, exception_handler: Optional[Callable] = None, sharded_pubsub: bool = False, ) -> "PubSubWorkerThread": return self._client.command_executor.execute_pubsub_run( sleep_time, daemon=daemon, exception_handler=exception_handler, pubsub=self, sharded_pubsub=sharded_pubsub, )