NAME EV::Redis - Asynchronous redis client using hiredis and EV SYNOPSIS use EV::Redis; my $redis = EV::Redis->new; $redis->connect('127.0.0.1'); # or my $redis = EV::Redis->new( host => '127.0.0.1' ); # command $redis->set('foo' => 'bar', sub { my ($res, $err) = @_; print $res; # OK $redis->get('foo', sub { my ($res, $err) = @_; print $res; # bar $redis->disconnect; }); }); # start main loop EV::run; DESCRIPTION EV::Redis is a fork of EV::Hiredis by Daisuke Murase (typester), extended with reconnection, flow control, TLS, and RESP3 support. It is a drop-in replacement for EV::Hiredis, except that commands which would pair replies with the wrong callbacks croak (see Reply order note). This is an asynchronous client for Redis using hiredis and EV as backend. It connects to EV with C-level interface so that it runs faster. ANYEVENT INTEGRATION AnyEvent has a support for EV as its one of backends, so EV::Redis can be used in your AnyEvent applications seamlessly. NO UTF-8 SUPPORT Unlike other redis modules, this module doesn't support utf-8 string. This module handle all variables as bytes: command arguments are sent as their byte representation regardless of Perl's internal string encoding (upgraded strings are downgraded in place, unless read-only), and passing a string with characters above 0xFF croaks with a "Wide character" error. You should encode your utf-8 string before passing commands like following: use Encode; # set $val $redis->set(foo => encode_utf8 $val, sub { ... }); # get $val $redis->get('foo', sub { my $val = decode_utf8 $_[0]; }); SIGPIPE Writing to a connection the server has already closed raises "SIGPIPE", which terminates the process by default. Set "$SIG{PIPE} = 'IGNORE'" so the write fails instead and the connection error reaches "on_error" as usual. FORK A child process inherits the connection's socket, which the parent goes on using. The child never uses it: when the child's loop next has an event for that connection, the connection fails there with "connection inherited from the parent process" -- its pending commands get that error, "on_error" runs, and "on_disconnect" too unless it was still connecting, and with "reconnect" the child connects on its own. A child forked inside a callback still gets the replies the parent had already read. A child that exits, or destroys the object, without running the loop changes nothing. The parent is not affected. A loop other than the default needs "$loop->loop_fork" in the child, as EV requires. Objects cannot be copied: "Storable" croaks on them, and "Clone" must not be used on them: the copy shares the original's connection, and destroying either one frees it for both. RESP3 REPLIES After "HELLO 3", maps and sets arrive as array references (a map as a flat key, value, key, value list), doubles as numbers, booleans as 1 or 0, and big numbers and verbatim strings as plain strings, without the verbatim format tag. Attribute replies are not supported: one ahead of a reply, as "DEBUG PROTOCOL attrib" sends, is dropped and the reply arrives as usual; one inside an array or map, which some modules send, fails the connection with a protocol error. METHODS new(%options); Create new EV::Redis instance. Available %options are: * host => 'Str' * port => 'Int' Hostname and port number of redis-server to connect. Mutually exclusive with "path". * path => 'Str' UNIX socket path to connect. Mutually exclusive with "host". * on_error => $cb->($errstr) Error callback will be called when a connection level error occurs. If not provided (or "undef"), a default handler that calls "die" is installed. Note that exceptions thrown inside any handler (including this default) or command callback are caught and reported as warnings -- they do not propagate out of the event loop. In practice the default handler therefore turns connection errors into warnings; install your own "on_error" to handle them programmatically. To have no error handler, call "$obj->on_error(undef)" after construction. A $SIG{__WARN__} that dies while such a warning is reported cannot stop the loop either: both messages are printed to STDERR instead. Perl runs a %SIG handler when Perl code next runs, which inside "EV::run" is a callback; a handler that dies there aborts that callback and is reported as a warning, so an "alarm" that dies does not end "EV::run". Use "command_timeout" or an "EV::timer" instead. This callback can be set by "$obj->on_error($cb)" method any time. * on_connect => $cb->() Connection callback will be called when connection successful and completed to redis server. With "tls", it is called once the TCP connection is up; a failed TLS handshake then arrives as "on_error" and "on_disconnect". This callback can be set by "$obj->on_connect($cb)" method any time. * on_disconnect => $cb->() Disconnect callback will be called when disconnection occurs (both normal and error cases). This callback can be set by "$obj->on_disconnect($cb)" method any time. * on_push => $cb->($reply) RESP3 push callback for server-initiated out-of-band messages (Redis 6.0+). Called with the decoded push message (an array reference). This enables client-side caching invalidation and other server-push features. This callback can be set by "$obj->on_push($cb)" method any time. * connect_timeout => $num_of_milliseconds Connection timeout. * command_timeout => $num_of_milliseconds Command timeout. * max_pending => $num Maximum number of commands sent to Redis concurrently. When this limit is reached, additional commands are queued locally and sent as responses arrive. 0 means unlimited (default). Use "waiting_count" to check the local queue size. * waiting_timeout => $num_of_milliseconds Maximum time a command can wait in the local queue before being cancelled with "waiting timeout" error. 0 means unlimited (default). * resume_waiting_on_reconnect => $bool Controls behavior of waiting queue on disconnect. If false (default), waiting commands are cancelled with error on disconnect and on each failed connect attempt. If true, waiting commands are preserved and resumed after successful reconnection; an explicit disconnect(), a lost connection or failed connect without "reconnect", or giving up on reconnecting still cancels them. The settings as the connection is lost decide; a handler changing them affects the next one, except that reconnect(0) or disconnect() there cancels the kept commands at once. Commands cancelled on a lost connection are off the queue before "on_error" and "on_disconnect" run. * reconnect => $bool Enable automatic reconnection on connection failure or unexpected disconnection. Default is disabled (0). * reconnect_delay => $num_of_milliseconds Delay between reconnection attempts. Default is 1000 (1 second). Used only with "reconnect", as is "max_reconnect_attempts". * max_reconnect_attempts => $num Maximum number of reconnection attempts. 0 means unlimited. Default is 0. Negative values are treated as 0 (unlimited). Only connects that fail count: an established connection starts the count again, also one that is closed at once or fails its TLS handshake, so retries against such a server or proxy go on. * priority => $num Priority for the underlying libev IO watchers. Higher priority watchers are invoked before lower priority ones. Valid range is -2 (lowest) to +2 (highest), with 0 being the default. See EV documentation for details on priorities. * keepalive => $seconds Enable TCP keepalive with the specified interval in seconds (at most 32767). When enabled, the OS will periodically send probes on idle connections to detect dead peers. 0 means disabled (default). Recommended for long-lived connections behind NAT gateways or firewalls. The interval applies on Linux with glibc and on macOS; elsewhere only keepalive itself is enabled, with the system's timing. Ignored for unix sockets. * prefer_ipv4 => $bool Resolve host names to IPv4 addresses, falling back to IPv6 only when there is none; this is already the default. Mutually exclusive with "prefer_ipv6". * prefer_ipv6 => $bool Resolve host names to IPv6 addresses, falling back to IPv4 only when there is none. Mutually exclusive with "prefer_ipv4". * source_addr => 'Str' Local address to bind the outbound connection to. Useful on multi-homed servers to select a specific network interface. * tcp_user_timeout => $num_of_milliseconds Set the TCP_USER_TIMEOUT socket option (Linux-specific). Controls how long transmitted data may remain unacknowledged before the connection is dropped. Helps detect dead connections faster on lossy networks. Ignored for unix sockets; where the system lacks the option, TCP connects fail with an error. * cloexec => $bool Set SOCK_CLOEXEC on the Redis connection socket. Prevents the file descriptor from leaking to child processes after fork/exec. Default is enabled. * reuseaddr => $bool Set SO_REUSEADDR on the Redis connection socket. Allows rebinding to an address that is still in TIME_WAIT state. Default is disabled. * tls => $bool Enable TLS/SSL encryption for the connection. Requires that the module was built with TLS support (auto-detected at build time, or forced with "EV_REDIS_SSL=1"). Only valid with "host" connections, not "path". * tls_ca => 'Str' Path to CA certificate file for server verification. If not specified, uses the system default CA store. * tls_capath => 'Str' Path to a directory containing CA certificate files in OpenSSL-compatible format (hashed filenames). Alternative to "tls_ca" for multiple CA certs. * tls_cert => 'Str' Path to client certificate file for mutual TLS authentication. Must be specified together with "tls_key". * tls_key => 'Str' Path to client private key file. Must be specified together with "tls_cert". * tls_server_name => 'Str' Server name for SNI (Server Name Indication), sent on every connection the object makes; without it no SNI is sent, which endpoints that route by SNI reject. It is not checked against the certificate. * tls_verify => $bool Enable or disable TLS peer verification. Default is true (verify). Set to false to accept self-signed certificates (not recommended for production). Verification checks the certificate chain against the CAs only: the bundled hiredis does not check that the certificate names the host, so any certificate those CAs issued is accepted. Use "tls_ca" with a CA dedicated to your Redis servers rather than the system store. * loop => 'EV::Loop', EV loop for running this instance. Default is "EV::default_loop". All parameters are optional. Unknown ones warn, unless "new" is called on a subclass. If parameters about connection (host&port or path) is not passed, you should call "connect" or "connect_unix" method by hand to connect to redis-server. connect($hostname [, $port]) connect_unix($path) Connect to a redis-server for "$hostname:$port" or $path. $port defaults to 6379. Croaks if a connection is already active, $port is outside 1..65535, or $path is too long for a unix socket (over 107 bytes on Linux, 103 on BSD and macOS). A failure detected inside the call (a missing unix socket, a host name that does not resolve) reaches "on_error" before the method returns, also when called from "new"; with reconnect enabled a retry is then scheduled. The host name is resolved inside the call, and again by each reconnect attempt: the event loop waits for the system resolver, "connect_timeout" does not cover it, and only the first address found is tried. Pass an IP address where DNS can be slow. command($commands..., [$cb->($result, $error)]) Do a redis command and return its result by callback. Returns "REDIS_OK" (0) on success or "REDIS_ERR" (-1) if the command could not be enqueued (the error is also delivered via callback, so the return value is rarely needed). $redis->command('get', 'foo', sub { my ($result, $error) = @_; print $result; # value for key 'foo' print $error; # redis error string, undef if no error }); If any error is occurred, $error presents the error message and $result is undef. If no error, $error is undef and $result presents response from redis. An error inside an array reply, such as "EXEC"'s result for a queued command that failed, arrives as its error text. Arrays nested more than 512 levels deep in a reply arrive empty; a reply nested more than 1024 levels deep is a protocol error that closes the connection. The callback is optional: only a code reference in the last position is taken as one, so an "undef" there is sent as an argument. Without a callback, the command runs in fire-and-forget mode: the reply from Redis is silently discarded and errors are not reported to Perl code (connection-level errors still trigger "on_error"). This is useful for high-volume writes where individual acknowledgement is not needed: $redis->set('counter', 42); # fire-and-forget, no callback The callback is the code reference passed, so reusing a variable for it is safe. Perl releases many closures in creation order slowly: with a million commands outstanding, each with a closure of its own, their replies or a cancel take about two minutes, against under a second with one shared code reference. NOTE: Alternatively all commands can be called via AUTOLOAD interface, including fire-and-forget: $redis->command('get', 'foo', sub { ... }); is equivalent to: $redis->get('foo', sub { ... }); The Redis "COMMAND" command itself is reached as "$redis->command('command', ...)", since "command" is this method. Note: Calling command() while not connected will croak with "connection required before calling command", unless automatic reconnection is active (reconnect timer running). Commands issued then wait locally and are sent in order once a connection is established; with "resume_waiting_on_reconnect", so are commands issued while an attempt is still connecting. Unless "resume_waiting_on_reconnect" is set, a failed attempt cancels waiting commands with its error. Queued commands respect "waiting_timeout" if set. A command issued from a command's callback that reports the loss or timeout of an established connection waits too: with "reconnect" and "resume_waiting_on_reconnect" it is sent on the next connection, otherwise it fails with that error once "on_error" and "on_disconnect" have run. After a failed connect it croaks as above, unless a reconnect is scheduled: then it waits for that attempt. From "on_error" and "on_disconnect" themselves it croaks as above: the reconnect is scheduled only after they return. Pub/Sub note: For "subscribe" and "psubscribe", the callback is persistent and receives all messages; at least one channel/pattern argument is required. For "unsubscribe" and "punsubscribe", the confirmation is delivered through the original subscribe callback (this is hiredis behavior). Any callback passed to unsubscribe commands is silently discarded, unless hiredis refuses the command (as when nothing is subscribed): then it gets the error. Subscribing again to a channel or pattern that already has a callback moves it to the new callback, including confirmations still due; a callback left with no channels is never called again. An error reply that arrives while the connection is subscribed (RESP2 or RESP3) is taken by hiredis as a connection error and closes the connection, so keep pub/sub on a connection of its own. A subscribed connection waits for messages without a timeout: set "keepalive" to notice a server that is gone without closing it. When the connection is lost or disconnect() closes it, a subscribe callback gets one error for each channel or pattern it still holds; when the object is destroyed it gets one. Sharded pub/sub note: "ssubscribe" and "sunsubscribe" are not supported and croak: the bundled hiredis has no sharded pub/sub support, so it cannot deliver "smessage" messages to a callback. "spublish" works as a regular command. MONITOR note: "monitor" requires an idle connection (no pending, waiting, or subscribed commands), and once active no further commands may be sent on that connection -- command() croaks. Both restrictions exist because hiredis's monitor mode re-queues callback records in a way that is unsafe to mix with other traffic. Use a dedicated connection; the state clears on disconnect. Reply order note: hiredis hands each reply to the oldest callback still waiting for one, so commands that change how the server answers are refused: "CLIENT REPLY OFF", "CLIENT REPLY SKIP", "REPLCONF ACK" and "REPLCONF GETACK" croak, and so do "SYNC" and "PSYNC", whose replication stream is not a sequence of replies; "RESET" croaks while the connection is subscribed or a subscribe is waiting (leave with "unsubscribe" and "punsubscribe" and wait for their replies, or disconnect), and fails through its callback if a subscription was made while it waited; pub/sub commands, "monitor" and "HELLO" sent inside "MULTI" fail through their callback. Pub/sub and "HELLO" wait locally for outstanding transaction replies before they are checked and sent; each such wait costs a round trip, and commands issued meanwhile queue behind it. A "MONITOR" the server refuses leaves the connection usable. An "EXEC" cancelled before it is sent (by "waiting_timeout" or "skip_waiting") leaves the transaction open: later commands are answered "QUEUED" until a "DISCARD". Nested event loop note: while a callback runs for an event on this connection (a reply, message or push, the connect, or the connection's loss), its I/O watchers are paused, so a nested "EV::run" (or condvar-style wait) inside it cannot process replies for the same connection -- they are delivered after the outer callback returns. Callbacks of commands cancelled locally ("waiting_timeout", "skip_waiting", "skip_pending", and waiting commands that disconnect() cancels) do not pause them. Use a separate connection if you must wait for Redis inside a callback. disconnect Disconnect from redis-server. Safe to call when already disconnected. Stops any pending reconnect timer, so explicit disconnect prevents automatic reconnection. Triggers the "on_disconnect" callback when disconnecting from an established connection. Called while the connection is still being established, it skips "on_connect" for it; "on_disconnect" still runs if commands already sent keep it open until they are answered. Waiting commands are cancelled with a "disconnected" error before it returns, also while the connection is still being established or finishing its pending replies, and when already disconnected (e.g., commands kept by "resume_waiting_on_reconnect"). Commands already sent are not cancelled: the connection closes once their replies have arrived. That holds while it is still being established too: they go out when it is up and are answered first. Commands sent while the connection is subscribed are cancelled with "disconnected" instead. So disconnect() does not drop a server that stopped answering: use "command_timeout", or destroy the object. Sending "QUIT" is no substitute: the server closing the connection is reported to "on_error" as a lost connection, and "reconnect" connects again. is_connected Returns true (1) if a connection context is active (including while the connection is being established), false (0) otherwise. has_ssl Class method. Returns true (1) if the module was built with TLS support, false (0) otherwise. if (EV::Redis->has_ssl) { # TLS connections are available } connect_timeout([$ms]) Get or set the connection timeout in milliseconds. Pass 0 to disable. Returns the current value, or undef if never set. Can also be set via constructor. It counts from the start of the connect attempt, once the host name is resolved; commands issued meanwhile do not extend it. Without it, a connect to a unix socket whose server has a full listen queue is retried in a busy loop until the server accepts it. With "tls" it covers the TCP connect only: a stalled handshake is ended by "command_timeout", once a command is outstanding. command_timeout([$ms]) Get or set the command timeout in milliseconds. Pass 0 to disable. Returns the current value, or undef if never set. Can also be set via constructor. It fires when replies are outstanding and nothing has been received for that long; sending more commands does not extend it, but a command too large for one write counts its bytes going out as progress, as the system's socket buffer takes them, so a steady stream of such commands can keep it from firing. That buffer can hold megabytes: on a slow link a large command can still time out while it goes out. (P)SUBSCRIBE and (P)UNSUBSCRIBE are not tracked as outstanding replies; MONITOR is, so a MONITOR connection that sees no traffic for that long is dropped with a timeout. A command that failed with a timeout, or with a lost connection, may still have run on the server, or may still run there later, even after commands sent on the next connection. Repeat only commands that are safe to run twice. When changed while connected, takes effect immediately on the active connection, including re-arming (or stopping, for 0) an already-scheduled timeout for commands currently in flight. on_error([$cb->($errstr)]) Set error callback. With a CODE reference argument, replaces the handler and returns the new handler. With "undef" or without arguments, clears the handler and returns undef; any other value clears it with a warning. Note: Calling without arguments clears the handler. There is no way to read the current handler without clearing it. This applies to all handler methods ("on_error", "on_connect", "on_disconnect", "on_push"). on_connect([$cb->()]) Set connect callback. With a CODE reference argument, replaces the handler and returns the new handler. With "undef" or without arguments, clears the handler and returns undef; any other value clears it with a warning. Commands issued from the callback go ahead of commands waiting in the local queue and past "max_pending", so it suits per-connection setup such as "AUTH" or "SELECT". If a setup command waits for transaction replies, later setup commands wait behind it, ahead of the ordinary queue. Commands still waiting in the local queue when it runs (issued during a reconnect delay, or over "max_pending") go after that setup. One sent while the connection is being established -- right after "new" or "connect", or while a reconnect attempt is connecting -- goes ahead of it, unless both "reconnect" and "resume_waiting_on_reconnect" are on, which make it wait. on_disconnect([$cb->()]) Set disconnect callback, called on both normal and error disconnections. With a CODE reference argument, replaces the handler and returns the new handler. With "undef" or without arguments, clears the handler and returns undef; any other value clears it with a warning. on_push([$cb->($reply)]) Set RESP3 push callback for server-initiated messages (Redis 6.0+). The callback receives the decoded push message as an array reference. With a CODE reference argument, replaces the handler and returns the new handler. With "undef" or without arguments, clears the handler and returns undef; any other value clears it with a warning. When changed while connected, takes effect immediately. $redis->on_push(sub { my ($msg) = @_; # $msg is an array ref, e.g. ['invalidate', ['key1', 'key2']] }); reconnect($enable, $delay_ms, $max_attempts) Configure automatic reconnection. $redis->reconnect(1); # enable with defaults (1s delay, unlimited) $redis->reconnect(1, 0); # enable with immediate reconnect $redis->reconnect(1, 2000); # enable with 2 second delay $redis->reconnect(1, 1000, 5); # enable with 1s delay, max 5 attempts $redis->reconnect(0); # disable $delay_ms defaults to 1000 (1 second). 0 means immediate reconnect: a local server that refuses connections then gets tens of thousands of attempts a second. $max_attempts defaults to 0 (unlimited). When enabled, the client will automatically attempt to reconnect on connection failure or unexpected disconnection. Intentional disconnect() calls will not trigger reconnection. The reconnect is scheduled after "on_error" and "on_disconnect" return, so command() still croaks inside them. A new connection starts fresh: subscriptions, "AUTH", "SELECT" and "HELLO" are not restored, so issue them from "on_connect" (which says what can overtake them). reconnect_enabled Returns true (1) if automatic reconnection is enabled, false (0) otherwise. pending_count Returns the number of commands sent to Redis awaiting responses. Persistent commands (subscribe, psubscribe, monitor) are not included in this count. When called from inside a callback for a reply or a connection error, the count includes the current command (it is decremented after the callback returns); inside one that "skip_pending" runs, it does not. waiting_count Returns the number of commands queued locally (not yet sent to Redis): commands over the "max_pending" limit, commands issued while a reconnect is pending (with "resume_waiting_on_reconnect", also while an attempt is connecting), commands waiting for transaction replies, and later commands queued behind them to keep their order. max_pending($limit) Get or set the maximum number of concurrent commands sent to Redis. Persistent commands (subscribe, psubscribe, monitor) do not count toward the limit; subscribe and psubscribe wait behind it like other commands, and monitor croaks unless the connection is idle. 0 means unlimited (default). When the limit is reached, additional commands are queued locally and sent as responses arrive. Replies still owed by a connection that disconnect() replaced count in "pending_count" but hold no slot of the new connection. waiting_timeout($ms) Get or set the maximum time in milliseconds a command can wait in the local queue. Commands exceeding this timeout are cancelled with "waiting timeout" error. 0 means unlimited (default). Returns the current value as an integer (0 when unset). With "reconnect" and "resume_waiting_on_reconnect" it also covers commands issued while a connect is in progress. Time spent held only for outstanding transaction replies (see Reply order note) does not count. resume_waiting_on_reconnect($bool) Get or set whether waiting commands are preserved on disconnect and resumed after reconnection. Default is false (waiting commands cancelled on disconnect and on each failed connect attempt). priority($priority) Get or set the priority for the underlying libev IO watchers. Higher priority watchers are invoked before lower priority ones when multiple watchers are pending. Valid range is -2 (lowest) to +2 (highest), with 0 being the default. Values outside this range are clamped automatically. Can be changed at any time, including while connected. $redis->priority(1); # higher priority $redis->priority(-1); # lower priority $redis->priority(99); # clamped to 2 my $prio = $redis->priority; # get current priority keepalive($seconds) Get or set the TCP keepalive interval in seconds (at most 32767; see the constructor option for platform limits). When set, the OS sends periodic probes on idle connections to detect dead peers. 0 means disabled (default). When set to a positive value while connected over TCP, takes effect immediately, and croaks, keeping the old value, if the system refuses it. Setting to 0 while connected records the preference for future connections but does not disable keepalives on the current socket. prefer_ipv4($bool) Get or set IPv4 preference for DNS resolution, already the default (see "new"). Mutually exclusive with "prefer_ipv6" (setting one clears the other). Takes effect on the next connection. prefer_ipv6($bool) Get or set IPv6 preference for DNS resolution. Mutually exclusive with "prefer_ipv4" (setting one clears the other). Takes effect on the next connection. source_addr($addr) Get or set the local source address to bind to when connecting. This is useful on multi-homed hosts to control which network interface is used. Pass "undef" to clear. Takes effect on the next TCP connection (has no effect on Unix socket connections). tcp_user_timeout($ms) Get or set the TCP user timeout in milliseconds. This controls how long transmitted data may remain unacknowledged before the connection is dropped. 0 means use the OS default. Takes effect on the next connection. Applies to TCP connections only; where the system lacks the option, TCP connects fail with an error. cloexec($bool) Get or set the close-on-exec flag for the Redis socket. When enabled, the socket is automatically closed in child processes after fork+exec. Enabled by default. Takes effect on the next connection. reuseaddr($bool) Get or set SO_REUSEADDR on the Redis socket. Allows rebinding to an address still in TIME_WAIT state. Disabled by default. Takes effect on the next connection. skip_waiting Cancel only waiting (not yet sent) command callbacks. Each callback is invoked with "(undef, "skipped")". In-flight commands continue normally. Commands issued by those callbacks are not cancelled. skip_pending Cancel all pending and waiting command callbacks. Each Perl callback is invoked immediately with "(undef, "skipped")". For pending commands, the internal hiredis tracking entry remains until a reply arrives (which is then discarded); no second callback fires. Commands issued by those callbacks, and callbacks already running, are not cancelled. can($method) Returns code reference if method is available, undef otherwise. Methods installed via AUTOLOAD (Redis commands) will return true after first call. DESTRUCTION BEHAVIOR When an EV::Redis object is destroyed (goes out of scope or is explicitly undefined) while commands are still pending or waiting, hiredis invokes all pending command callbacks with a disconnect error, and EV::Redis invokes all waiting queue callbacks with "disconnected". This ensures callbacks are not orphaned. An object still alive at global destruction (a package variable, or one kept by a reference cycle) runs no callbacks. For predictable cleanup, explicitly disconnect before destruction: $redis->disconnect; # waiting callbacks get "disconnected" undef $redis; # pending callbacks get "disconnected" Or use skip methods to cancel with a specific error message: $redis->skip_pending; # Invokes callbacks with (undef, "skipped") $redis->skip_waiting; $redis->disconnect; undef $redis; Circular references: If your callbacks close over the $redis variable, this creates a reference cycle ($redis -> object -> callback -> $redis) that prevents garbage collection. Break the cycle before the object goes out of scope by clearing callbacks: $redis->on_error(undef); $redis->on_connect(undef); $redis->on_disconnect(undef); $redis->on_push(undef); BENCHMARKS Measured on Linux with Unix socket connection, 100,000 commands with 100-byte values, Perl 5.40, Redis 8.x ("bench/benchmark.pl" in the source repository, "BENCH_COMMANDS=100000"): Pipeline SET ~107K ops/sec Pipeline GET ~112K ops/sec Mixed workload ~112K ops/sec Fire-and-forget SET ~655K ops/sec Sequential round-trip ~39K ops/sec (SET+GET pairs) Fire-and-forget mode (no callback) is roughly 6x faster than callback mode due to zero Perl-side overhead per command. Pipeline throughput is bounded by the event loop round-trip, not by hiredis or the network. Flow control ("max_pending") has minimal impact at reasonable limits: unlimited ~180K ops/sec max_pending=500 ~186K ops/sec max_pending=100 ~146K ops/sec Run "perl bench/benchmark.pl" in a checkout of the repository for full results. Set "BENCH_COMMANDS" and "BENCH_VALUE_SIZE" environment variables to customize; the default of 10,000 commands gives different rates. AUTHOR Daisuke Murase (typester) (original EV::Hiredis) vividsnow COPYRIGHT AND LICENSE Copyright (c) 2013 Daisuke Murase, 2026 vividsnow. All rights reserved. This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.