The Perl Toolchain Summit needs more sponsors. If your company depends on Perl, please support this very important event.


IO::Lambda - non-blocking I/O as lambda calculus


The code below demonstrates execution of parallel HTTP requests

   use strict;
   use IO::Lambda qw(:lambda :func);
   use IO::Socket::INET;

   # this function creates a new lambda object 
   # associated with one socket, and fetches a single URL
   sub http
      my $host = shift;

      # Simple HTTP functions by first sending request to the remote, and
      # then waiting for the response. This sets up a new lambda object with 
      # attached one of many closures that process sequentially
      return lambda {

         # create a socket, and issue a tcp connect
         my $socket = IO::Socket::INET-> new( 
            PeerAddr => $host, 
            PeerPort => 80 

         # Wait until socket become writable. Parameters to writable()
         # are passed using context(). This association is remembered 
         # within the engine. 
         context $socket;

         # writeable sets up a possible event to monitor, when 
         # $socket is writeable, execute the closure.
      writable {
         # The engine discovered we can write, so send the request
         print $socket "GET /index.html HTTP/1.0\r\n\r\n";

         # This variable needs to stay shared across
         # multiple invocations of our readable closure, so 
         # it needs to be outside that closure. Here, it collects
         # whatever the remote returns
         my $buf = ''; 

         # readable registers another event to monitor - 
         # that $socket is readable. Note that we do not 
         # need to set the context again because when we get 
         # here, the engine knows what context this command 
         # took place in, and assumes the same context. 
         # Also note that socket won't be awaited for writable events
         # anymore, and this code won't be executed for this $socket.
      readable {
         # This closure is executed when we can read.
         # Read from the socket. sysread() returns number of
         # bytes read. Zero means EOF, and undef means error, so
         # we stop on these conditions.
         # If we return without registering a follow-up 
         # handler, this return will be processed as the 
         # end of this sequence of events for whoever is 
         # waiting on us.
         return $buf unless 
            sysread( $socket, $buf, 1024, length($buf));

         # We're not done so we need to do this again. 
         # Note that the engine knows that it just 
         # called this closure because $socket was 
         # readable, so it can infer that it is supposed 
         # to set up a callback that will call this 
         # closure when $socket is next readable.

   # Fire up a single lambda and wait until it completes.
   print http('')-> wait;

   # Fire up a lambda that waits for two http requests in parallel.
   # tails() can wait for more than one lambda
   my @hosts = ('', '');
   lambda {
      context map { http($_) } @hosts;
      # tails() asynchronously waits until all lambdas in the context
      # are finished.
      tails { print @_ }
   }-> wait;

   # crawl for all urls in parallel, but keep 10 parallel connections max
   print par(10)-> wait(map { http($_) } @hosts);
   # crawl for all urls sequentially
   print mapcar( curry { http(shift) })-> wait(@hosts);

Note: io and lambda are synonyms - I personally prefer lambda but some find the word slightly inappropriate, hence io. See however "Higher-order functions" to see why it is more lambda than io.


This module is another attempt to fight the horrors of non-blocking I/O. It tries to bring back the simplicity of the declarative programming style, that is only available when one employs threads, coroutines, or co-processes. Usually coding non-blocking I/O for single process, single thread programs requires construction of state machines, often fairly complex, which fact doesn't help the code clarity, and is the reason why the asynchronous I/O programming is often considered 'messy'. Similar to the concept of monads in functional languages, that enforce a certain order of execution over generally orderless functions, IO::Lambda allows writing I/O callbacks in a style that resembles the good old sequential, declarative programming.

The manual begins with code examples, then proceeds to explaining basic assumptions, then finally gets to the complex concepts, where the real fun begins. You can skip directly there ("Stream IO", "Higher-order functions"), where the functional style mixes with I/O. If, on the contrary, you are intimidated by the module's ambitions, you can skip to "Simple use" for a more gentle introduction. Those, who are interested how the module is different from the other I/O frameworks, please continue reading.

Simple use

This section is for those who don't need all of the module's powerful machinery. Simple callback-driven programming examples show how to use the module for unsophisticated tasks, using concepts similar to the other I/O frameworks. It is possible to use the module on this level only, however one must be aware that by doing so, the real power of the higher-order abstraction is not used.

IO::Lambda, like all I/O multiplexing libraries, provides functions for registering callbacks, that in turn are called when a timeout occurs, or when a file handle is ready for reading and/or writing. See below code examples that demonstrate how to program on this level of abstraction.

Combination of two timeouts and an IO_READ event
   use IO::Lambda qw(:constants);

   my $obj = IO::Lambda-> new;

   # Either 3 or time + 3 will do. See "Time" section for more info
   $obj-> watch_timer( 3, sub { print "I've slept 3 seconds!\n" });

   # I/O flags is a combination of IO_READ, IO_WRITE, and IO_EXCEPTION.
   # Timeout is either 5 or time + 5, too.
   $obj-> watch_io( IO_READ, \*STDIN, 5, sub {
        my ( $self, $ok) = @_;
        print $ok ?
            "stdin is readable!\n" : 
            "stdin is not readable within 5 seconds\n";

   # main event loop is stopped when there are no lambdas and no 
   # pending events
Waiting for another lambda to complete
   use IO::Lambda;

   my $a = IO::Lambda-> new;
   $a-> watch_timer( 3, sub { print "I've slept 3 seconds!\n" });

   my $b = IO::Lambda-> new;
   # A lambda can wait for more than one event or lambda.
   # A lambda can be awaited by more than one lambda.
   $b-> watch_lambda( $a, sub { print "lambda #1 is finished!\n"});


Example: reading lines from a filehandle

Given $filehandle is non-blocking, the following code creates a lambda object (later, simply a lambda) that reads from the handle until EOF or an error occured. Here, getline (see "Stream IO" below) constructs a lambda that reads a single line from a filehandle.

    use IO::Lambda qw(:all);

    sub my_reader
       my $filehandle = shift;
       lambda {
           context getline, $filehandle, \(my $buf = '');
       tail {
           my ( $string, $error) = @_;
           if ( $error) {
               warn "error: $error\n";
           } else {
               print $string;
               return again;

Assume we have two socket connections, and sockets are non-blocking - read from both of them in parallel. The following code creates a lambda that reads from two readers:

    sub my_reader_all
        my @filehandles = @_;
        lambda {
            context map { my_reader($_) } @filehandles;
            tails { print "all is finished\n" };

    my_reader_all( $socket1, $socket2)-> wait;

Non-blocking HTTP client

Given a socket, create a lambda that implements the HTTP protocol

    use IO::Lambda qw(:all);
    use IO::Socket;
    use HTTP::Request;

    sub talk
        my $req    = shift;
        my $socket = IO::Socket::INET-> new( PeerAddr => '', PeerPort => 80);

        lambda {
            context $socket;
            writable {
                # connected
                print $socket "GET ", $req-> uri, "\r\n\r\n";
                my $buf = '';
                readable {
                    sysread $socket, $buf, 1024, length($buf) or return $buf;
                    again; # wait for reading and re-do the block

Connect and talk to the remote

    $request = HTTP::Request-> new( GET => '');

    my $q = talk( $request );
    print $q-> wait; # will print content of $buf

Connect two parallel connections: by explicitly waiting for each

    $q = lambda {
        context talk($request);
        tail { print shift };
        context talk($request2);
        tail { print shift };
    $q-> wait;

Connect two parallel connections: by waiting for all

    $q = lambda {
        context talk($request1), talk($request2);
        tails { print for @_ };
    $q-> wait;

Teach our simple http request to redirect by wrapping talk(). talk_redirect() will have exactly the same properties as talk() does

    sub talk_redirect
        my $req = shift;
        lambda {
            context talk( $req);
            tail {
                my $res = HTTP::Response-> parse( shift );
                return $res unless $res-> code == 302;

                $req-> uri( $res-> uri);
                context talk( $req);

Full example code

    use strict;
    use IO::Lambda qw(:lambda);
    use IO::Socket::INET;

    sub get
        my ( $socket, $url) = @_;
        lambda {
            context $socket;
        writable {
            print $socket "GET $url HTTP/1.0\r\n\r\n";
            my $buf = '';
        readable {
            my $n = sysread( $socket, $buf, 1024, length($buf));
            return "read error:$!" unless defined $n;
            return $buf unless $n;

    sub get_parallel
        my @hosts = @_;

        lambda {
           context map { get(
              IO::Socket::INET-> new(
                  PeerAddr => $_, 
                  PeerPort => 80 
              ), '/index.html') } @hosts;
           tails {
              join("\n\n\n", @_ )

    print get_parallel('', '')-> wait;

See tests and additional examples in directory eg/ for more information.


Events and states

A lambda is an IO::Lambda object, that waits for I/O and timeout events, and for events generated when other lambdas are completed. On each such event a callback is executed. The result of the execution is saved, and passed on to the next callback, when the next event arrives.

Life cycle of a lambda goes through three modes: passive, waiting, and stopped. A lambda that is just created, or was later reset with reset call, is in the passive state. When the lambda gets started, the only executed code will be the callback associated with the lambda:

    $q = lambda { print "hello world!\n" };
    # not printed anything yet
    $q-> wait; # <- here it will

Lambdas are usually not started explicitly. Usually, the function that can wait for a lambda, starts it too. wait, the synchronous waiter, and tail/tails, the asynchronous ones, start passive lambdas when called. A lambda is finished when there are no more events to listen to. The lambda in the example above will finish right after print statement.

Lambda can listen to events by calling conditions, that internally subscribe the lambda object to the corresponding file handles, timers, and other lambdas. Most of the expressive power of IO::Lambda lies in the conditions, such as readable, writable, timeout. Conditions are different from normal perl subroutines in the way how they receive their parameters. The only parameter they receive in the normal way, is the associated callback, while all other parameters are passed to it through the alternate stack, by the explicit context call.

In the example below, lambda watches for file handle readability:

    $q = lambda {
        context \*SOCKET;
        readable { print "I'm readable!\n"; }
        # here is nothing printed yet
    # and here is nothing printed yet

Such lambda, when started, will switch to the waiting state, which means that it will be waiting for the socket. The lambda will finish only after the callback associated with readable condition is called. Of course, new event listeners can be created inside all callbacks, on each state. This fact constitutes another large benefit of IO::Lambda, as it allows to program FSMs dynamically.

The new event listeners can be created either by explicitly calling condition, or by restarting the last condition with the again call. For example, code

     readable { 
        print 1;
        again if int rand(2)

prints indeterminable number of ones.


All callbacks associated with a lambda object (further on, merely lambda) execute in one, private context, also associated to the lambda. The context here means that all conditions register callbacks on an implicitly given lambda object, and keep the passed parameters on the context stack. The fact that the context is preserved between states, helps building terser code with series of IO calls:

    context \*SOCKET;
    writable {
    readable {

is actually the shorter form for

    context \*SOCKET;
    writable {
    context \*SOCKET; # <-- context here is retained from one frame up
    readable {

And as the context is bound to the current closure, the current lambda object is too, in this property. The code above is actually

    my $self = this;
    context \*SOCKET;
    writable {
    this $self;      # <-- object reference is retained here
    context \*SOCKET;
    readable {

this can be used if more than one lambda needs to be accessed. In which case,

    this $object;
    context @context;

is the same as

    this $object, @context;

which means that explicitly setting this will always clear the context.

Data and execution flow

A lambda is initially called with some arguments passed from the outside. These arguments can be stored using the call method; wait and tail also issue call internally, thus replacing any previous data stored by call. Inside the lambda these arguments are available as @_.

Whatever is returned by a condition callback (including the lambda condition itself), will be passed further on as @_ to the next callback, or to the outside, if the lambda is finished. The result of the finished lambda is available by peek method, that returns either all array of data available in the array context, or first item in the array otherwise. wait returns the same data as peek does.

When more than one lambda watches for another lambda, the latter will get its last callback results passed to all the watchers. However, when a lambda creates more than one state that derive from the current state, a forking behaviour of sorts, the latest stored results gets overwritten by the first executed callback, so constructions such as

    readable  { 1 + shift };
    writable { 2 + shift };

will eventually return 3, but whether it will be 1+2 or 2+1, is undefined.

wait is not the only function that synchronises input and output data. wait_for_all method waits for all lambdas, including the caller, to finish. It returns collected results of all the objects in a single list. wait_for_any method waits for at least one lambda, from the list of passed lambdas (again, including the caller), to finish. It returns list of finished objects as soon as possible.


Timers and I/O timeouts can be given not only in the timeout values, as it usually is in event libraries, but also as deadlines in (fractional) seconds since epoch. This decision, strange at first sight, actually helps a lot when a total execution time is to be tracked. For example, the following code reads as many bytes as possible from a socket within 5 seconds:

   lambda {
       my $buf = '';
       context $socket, time + 5;
       readable {
           if ( shift ) {
               return again if sysread $socket, $buf, 1024, length($buf);
           } else {
               print "oops! a timeout\n";

Rewriting the same code with readable semantics that accepts time as a timeout instead, would be not that elegant:

   lambda {
       my $buf = '';
       my $time_left = 5;
       my $now = time;
       context $socket, $time_left;
       readable {
           if ( shift ) {
               if (sysread $socket, $buf, 1024, length($buf)) {
                   $time_left -= (time - $now);
                   $now = time;
                   context $socket, $time_left;
                   return again;
           } else {
               print "oops! a timeout\n";

However, the exact opposite is true for timeout. The following two lines both sleep 5 seconds:

   lambda { context 5;        timeout {} }
   lambda { context time + 5; timeout {} }

Internally, timers use Time::HiRes::time that gives the fractional number of seconds. This however is not required for the caller, because when high-res timers are not used, timeouts will simply be less precise, and will jitter plus-minus half a second.


All conditions receive their parameters from the context stack, or simply the context. The only parameter passed to them by using perl call, is the callback itself. Conditions can also be called without a callback, in which case, they will pass further data that otherwise would be passed as @_ to the callback. Thus, a condition can be called either as

    readable { .. code ... }


    &readable(); # no callback
    &readable;   # DANGEROUS!! same as &readable(@_)

Conditions can either be used after explicit exporting

   use IO::Lambda qw(:lambda);
   lambda { ... }

or by using the package syntax,

   use IO::Lambda;
   IO::Lambda::lambda { ... };

Note: If you know concept of continuation-passing style, this is exactly how conditions work, except that closures are used instead of continuations (Brock Wilcox:thanks!) .


Creates a new IO::Lambda object.


Same as lambda.

readable($filehandle, $deadline = undef)

Executes either when $filehandle becomes readable, or after $deadline. Passes one argument, which is either TRUE if the handle is readable, or FALSE if time is expired. If deadline is undef, then no timeout is registered, that means that it will never be called with FALSE.

writable($filehandle, $deadline = undef)

Exactly same as readable, but executes when $filehandle becomes writable.

rwx($flags, $filehandle, $deadline = undef)

Executes either when $filehandle satisfies any of the condition in $flags, or after $deadline. $flags is a combination of three integer constants, IO_READ, IO_WRITE, and IO_EXCEPTION, that are imported by

   use IO::Lambda qw(:constants);

Passes one argument, which is either a combination of the same IO_XXX flags, that report which conditions the handle satisfied, or 0 if time is expired. If deadline is undef, no timeout is registered, i.e. will never return 0.


Executes after $deadline. $deadline cannot be undef.

tail($lambda, @parameters)

Issues $lambda-> call(@parameters), then waits for the $lambda to complete. Since call can only be done on inactive lambdas, will fail if @parameters is not empty and $lambda is already running.

By default, tail resets lambda if is was alredy finished. This behavior can be changed by manipulating autorestart property.


Executes when all objects in @lambdas are finished, returns the collected, unordered results of the objects.


Same as tails, but the results are ordered.


Executes either when all objects in @lambdas are finished, or $deadline expires. Returns lambdas that were successfully executed during the allotted time.

context @ctx

If called with no parameters, returns the current context, otherwise replaces the current context with @ctx. It is thus not possible (not that it is practical anyway) to clear the context with this call. If really needed, use this(this) syntax.

this $this, @ctx

If called with no parameters, returns the current lambda. Otherwise, replaces both the current lambda and the current context. Can be useful either when juggling with several lambdas, or as a convenience hack over my variables, for example,

    this lambda { ... };
    this-> wait;

instead of

    my $q = lambda { ... };
    $q-> wait;
condition $lambda, $callback, $method, $name

Helper function for creating conditions, either from lambdas or from lambda constructors.

Example: convert existing getline constructor into a condition:

   sub gl(&) { getline-> call(context)-> condition( shift, \&gl, 'gl') }
   context $fh, $buf, $deadline;
   gl { ... }


These are functions to jump to previous callback frames to a previously saved context.

again([$frame, [@context]])

Restarts the frame with the context. If $frame is given, jumps to the frame previously returned by a restartable call, resetting the context to its previous state too. If @context is given, it is used instead.

All the conditions above, excluding lambda, are restartable with again call (see start for restarting a lambda). The code

   context $obj1;
   tail {
       return if $null++;
       context $obj2;

is thus equivalent to

   context $obj1;
   tail {
       context $obj2;

again passes the current context to the condition.

If $frame is provided, then it is treated as result of previous restartable call. It contains data sufficient to restarting another call, instead of the current. See Frames for details.


Save a frame. restartable can generate unique $name itself if not given. All frames are deleted when a lambda is stopped.


    my $counter = 0;
    context lambda { $counter += 1 };
    tail {
        return if 12 == shift;
        my $frame = restartable;
        context lambda { $counter += 10 };
        tail {

restartable records the current content on the lambda, and again switches it back so that again call goes to the first tail instead of the second.


Deletes existing saved frame. Can be used to clean up eventual circular references (see below).

get_frame, set_frame(@frame), swap_frame(@frame)

A lower level frame accessors that save and restore all contexts. Do not use directly because it can easily used to inadvertedly create circular references, where @frame points to a callback while the callback via the closure mechanisms holds a reference to the @frame variable (Thanks to Ben Tilly for bringing up the issue).

Stream IO

The whole point of this module is to help building protocols of arbitrary complexity in a clear, consequent programming style. Consider how perl's low-level sysread and syswrite relate to its higher-level readline, where the latter not only does the buffering, but also recognizes $/ as input record separator. The section above described lower-level lambda I/O conditions, that are only useful for sysread and syswrite. This section tells about higher-level lambdas that relate to these low-level ones, as the aforementioned readline relates to sysread.

All functions in this section return the lambda, that does the actual work. Not unlike as a class constructor returns a newly created class instance, these functions return newly created lambdas. Such functions will be further referred as lambda constructors, or simply constructors. Therefore, constructors are documented here as having two inputs and one output, as for example a function sysreader is a function that takes 0 parameters, always returns a new lambda, and this lambda, in turn, takes four parameters and returns two. This constructor will be described as

    # sysreader() :: ($fh,$$buf,$length,$deadline) -> ($result,$error)

Since all stream I/O lambdas return same set of scalars, the return type will be further on referred as ioresult:

    # ioresult    :: ($result, $error)
    # sysreader() :: ($fh,$$buf,$length,$deadline) -> ioresult

ioresult's first scalar is defined on success, and is not otherwise. In the latter case, the second scalar contains the error, usually either $! or 'timeout' (if $deadline was set).

Before describing the actual functions, consider the code that may benefit from using them. Let's take a lambda that needs to implement a very simple HTTP/0.9 request:

   lambda {
       my $handle = shift;
       my $buf = '';
       context getline, $handle, \$buf;
   tail {
       my $req = shift;
       die "bad request" unless $req =~ m[GET (.*)$]i;
       do_request($handle, $1);

getline reads from $handle to $buf, and wakes up when a new line is there. However, what if we need, for example, HTTPS instead of HTTP, where reading from a socket may involve some writing, and of course some waiting? Then the first default parameter to getline has to be replaced. By default,

   context getline, $handle, \$buf;

is the same as

   my $reader = sysreader;        
   context getline($reader), $handle, \$buf;

where sysreader creates a lambda $reader, that given $handle, awaits when it becomes readable, and reads from it. getline, in turn, repeatedly calls $reader, until the whole line is read.

Thus, we call

   context getline(https_reader), $handle, \$buf;

instead, that should conform to sysreader signature:

   sub https_reader
       lambda {
           my ( $fh, $buf, $length, $deadline) = @_;
           # read from SSL socket
           return $error ? (undef, $error) : $data;

I'm not showing the actual implementation of a HTTPS reader (if you're curious, look at IO::Lambda::HTTP::HTTPS ), but the idea is that inside that reader, it is perfectly fine to do any number of read and write operations, and wait for their completion too, as long as the upper-level lambda will sooner or later gets the data. getline (or, rather, readbuf that getline is based on) won't care about internal states of the reader.

Check out t/06_stream.t that emulates reading and writing implemented in this fashion.

These functions are imported with

   use IO::Lambda qw(:stream);
sysreader() :: ($fh, $$buf, $length, $deadline) -> ioresult

Creates a lambda that accepts all the parameters used by sysread (except $offset though), plus $deadline. The lambda tries to read $length bytes from $fh into $buf, when $fh becomes available for reading. If $deadline expires, fails with 'timeout' error. On successful read, returns number of bytes read, or $! otherwise.

syswriter() :: ($fh, $$buf, $length, $offset, $deadline) -> ioresult

Creates a lambda that accepts all the parameters used by syswrite plus $deadline. The lambda tries to write $length bytes to $fh from $buf from $offset, when $fh becomes available for writing. If $deadline expires, fails with 'timeout' error. On successful write, returns number of bytes written, or $! otherwise.

readbuf($reader = sysreader()) :: ($fh, $$buf, $cond, $deadline) -> ioresult

Creates a lambda that is able to perform buffered reads from $fh, either using custom lambda reader, or using one newly generated by sysreader. The lambda, when called, reads continually from $fh into $buf, and either fails on timeout, I/O error, or end of file, or succeeds if $cond condition matches.

The condition $cond is a "smart match" of sorts, and can be one of:


The lambda will succeed when $buf is exactly $cond bytes long.


The lambda will succeed when $cond matches the content of $buf. Note that readbuf saves and restores value of pos($$buf), so use of \G is encouraged here.

coderef :: ($buf -> BOOL)

The lambda succeeds if coderef called with $buf returns true value.


The lambda will succeed on end of file. Note that for all other conditions end of file is reported as an error, with literal "eof" string.

writebuf($writer) :: ($fh, $$buf, $length, $offset, $deadline) -> ioresult

Creates a lambda that is able to perform buffered writes to $fh, either using custom lambda writer, or using one generated by syswriter. That writer lambda, in turn, writes continually $buf (from $offset, $length bytes) and either fails on timeout or I/O error, or succeeds when $length bytes are written successfully.

If $length is undefined, buffer is continuously checked if it got new data. This feature can be used to implement concurrent writes.

getline($reader) :: ($fh, $$buf, $deadline) -> ioresult

Same as readbuf, but succeeds when a string of bytes ended by a newline is read.

Higher-order functions

Functions described in this section justify the lambda in IO::Lambda. Named deliberately after the classic function names, they provide a similar interface.

These function are imported with

   use IO::Lambda qw(:func);
mapcar($lambda) :: @p -> @r

Given a $lambda, creates another lambda, that accepts array @p, and sequentially executes $lambda with each parameter from the array. The lambda returns results collected from the executed lambdas.

   print mapcar( lambda { 1 + shift })-> wait(1..5);

mapcar can be used for organizing simple loops:

   mapcar(curry { sendmail(shift) })-> wait(@email_addresses);
filter($lambda) :: @p -> @r

Given a $lambda, creates another lambda, that accepts array @p, and sequentially executes $lambda with each parameter from the array. Depending on the result of the execution, parameters are either returned, or not returned back to the caller.

   print filter(lambda { shift() % 2 })-> wait(1..5);
fold($lambda) :: @b -> @c; $lambda :: ($a,@b) -> @c

Given a $lambda, returns another lambda that accepts array @b, and runs pairwise its members through $lambda. Results of repeated execution of $lambda is returned.

   print fold( lambda { $_[0] + $_[1] } )-> wait( 1..4 );
curry(@a -> $l) :: @a -> @b

curry accepts a function that returns a lambda, and possible parameters to it. Returns a new lambda, that will execute the inner lambda, and returns its result as is. For example,

   context $lambda, $a, $b, $c;
   tail { ... }

where $lambda accepts three parameters, can be rewritten as

   $m = curry { $lambda, $a, $b };
   context $m, $c;
   tail { ... }

Another example, tie readbuf with a filehandle and buffer:

   my $readbuf = curry { readbuf, $fh, \(my $buf = '') };
seq() :: @a -> @b

Creates a new lambda that executes all lambdas passed to it in @a sequentially, one after another. The lambda returns results collected from the executed lambdas.

  sub seq { mapcar curry { shift }}
  print seq-> wait( map { my $k = $_; lambda { $k } } 1..5);
par($max = 0) :: @a -> @b

Given a limit $max, returns a new lambda that accepts lambdas in @a to be executed in parallel, but so that number of lambdas that run simultaneously never goes higher than the limit. The lambda returns results collected from the executed lambdas.

If $max is undefined or 0, behaves similar to a lambda version of tails, i.e., all of the lambdas are run in parallel.

The code below prints 123, then sleeps, then 456, then sleeps, then 789.

  par(3)-> wait( map {
      my $k = $_;
      lambda {
          context 0.5;
          timeout { print $k, "\n" }
  } 1..9);

Object API

This section lists methods of IO::Lambda class. Note that by design all lambda-style functionality is also available for object-style programming. Together with the fact that lambda syntax is not exported by default, it thus leaves a place for possible implementations of user-defined syntax, either with or without lambdas, on top of the object API, without accessing the internals.

The object API is mostly targeted to developers that need to connect third-party asynchronous event libraries with the lambda interface.

new($class, $start)

Creates new IO::Lambda object in the passive state. $start will be called once, after the lambda gets active.

watch_io($flags, $handle, $deadline, $callback, $cancel)

Registers an IO event listener that calls $callback either after $handle satisfies condition of $flags ( a combination of IO_READ, IO_WRITE, and IO_EXCEPTION bits), or after $deadline time is passed. If $deadline is undef, watches for the file handle indefinitely.

The callback is called with first parameter as integer set of IO_XXX flags, or 0 if the callback was timed out. Other parameters, as it is the case with the other callbacks, are passed the result of the last called callback attached to the same lambda. The result of this callback will then be stored and passed on to the next callback in the same fashion.

If the event is cancelled with cancel_event, then $cancel callback is executed. The result of this callback will be stored and passed on, in the same manner as results and parameters to $callback.

watch_timer($deadline, $callback, $cancel)

Registers a timer listener that calls $callback after $deadline time.

watch_lambda($lambda, $callback, $cancel)

Registers a listener that calls $callback after $lambda, a IO::Lambda object is finished. If $lambda is in passive state, it is started first.


Reports whether lambda is stopped or not.


Reports whether lambda has any registered callbacks left or not.


Reports if lambda wasn't run yet. Is true when the lambda is in a state after either new or reset are called.


Reports if lambda was run.


Cancels all watchers and switches the lambda to the passive state. If there are any lambdas that watch for this object, these will be called first.


If set, gives permission to watchers to reset the lambda if it becomes stopped. tail does that when needed, other watchers are allowed to do that too. Is set by default.


At any given time, returns stored data that are either passed in by call if the lambda is in the passive state, or stored result of execution of the latest callback.


Starts a passive lambda. Can be used for effective restart of the whole lambda; the only requirement is that the lambda should have no pending events.

call @args

Stores @args internally, to be passed on to the first callback. Only works in passive state, croaks otherwise. If called multiple times, arguments from the previous calls are overwritten.

terminate @args

Cancels all watchers and resets lambda to the stopped state. If there are any lambdas that watch for this object, these will be notified first. @args will be stored and available for later calls by peek.


Cancels all watchers and resets lambda to the stopped state. Does the same to all lambdas the caller lambda watches after, recursively. Useful where explicit, long-lived lambdas shouldn't be subject to the global destruction, which kills objects in random order; destroy kills them in some order, at least.

wait @args

Waits for the caller lambda to finish, returns the result of peek. If the object was in passive state, calls call(@args), otherwise @args are not used.

wait_for_all @lambdas

Waits for caller lambda and @lambdas to finish. Returns collection of peek results for all objects. The results are unordered.

wait_for_any @lambdas

Waits for at least one lambda from the list of caller lambda and @lambdas to finish. Returns list of finished objects.

yield $nonblocking = 0

Runs one round of dispatching events. Returns 1 if there are more events in internal queues, 0 otherwise. If $NONBLOCKING is set, exits as soon as possible, otherwise waits for events; this feature can be used for organizing event loops without wait/run calls.


Enters the event loop and doesn't exit until there are no registered events. Can be also called as package method.

bind $cancel, @args

Creates an event record that contains the lambda and @args, and returns it. The lambda won't finish until this event is returned with resolve. $cancel is an optional callback that will be called when the event is cancelled; the callback is passed two parameters, the lambda and the cancelled event record.

bind can be called several times on a single lambda; each event requires individual resolve.

resolve $event

Removes $event from the internal waiting list. If a lambda has no more events to wait, notifies eventual lambdas that wait to the objects, and then stops.

Note that resolve doesn't provide any means to call associated callbacks, which is intentional.

intercept $condition [ $state = '*' ] $coderef

Installs a $coderef as an overriding hook for a condition callback, where condition is tail, readable, writable, etc. Whenever a condition callback is being called, the $coderef hook will be called instead, that should be able to analyze the call, and allow or deny it the further processing.

$state, if omitted, is equivalent to '*', that means that checks on lambda state are omitted too. Setting $state to undef is allowed though, and will match when the lambda state is also undefined (which it is by default).

There can exist more than one intercept handlers, stacked on top of each other. If $coderef is undef, the last registered hook is removed.


    my $q = lambda { ... tail { ... }};
    $q-> intercept( tail => sub {
        if ( stars are aligned right) {
            # pass
            return this-> super(@_);
        } else {
            return 'not right';

See also state, super, and override.

override $condition [ $state = '*' ] $coderef

Installs a $coderef as an overriding hook for a condition - tail, readable, writable, etc, possibly with a named state. Whenever a lambda calls one of these condition, the $coderef hook will be called instead, that should be able to analyze the call, and allow or deny it the further processing.

$state, if omitted, is equivalent to '*', that means that checks on lambda state are omitted too. Setting $state to undef is allowed though, and will match when the lambda state is also undefined (which it is by default).

There can exist more than one override handlers, stacked on top of each other. If $coderef is undef, the last registered hook is removed.


    my $q = lambda { ... tail { ... }};
    $q-> override( tail => sub {
        if ( stars are aligned right) {
            # pass
            this-> super;
        } else {
            # deny and rewrite result
            return tail { 'not right' }

See also state, super, and intercept.


Analogous to Perl's SUPER, but on the condition level, this method is designed to be called from overridden conditions to call the original condition or callback.

There is a slight difference in the call syntax, depending on whether it is being called from inside an override or intercept callback. The intercept'ed callback will call the previous callback right away, and may call it with parameters directly. The override callback will only call the condition registration routine itself, not the callback, and therefore is called without parameters. See intercept and override for examples of use.

state $state

A helper function for explicit naming of condition calls. The function stores the $state string on the current lambda; this string can be used in calls to intercept and override to identify a particular condition or a callback.

The recommended use of the method is when a lambda contains more than one condition of a certain type; for example the code

   tail {
   tail {

is therefore better to be written as

   state A => tail {
   state B => tail {

Exceptions and backtrace

In addition to the normal call stack as reported by the caller builtin, it can be useful also to access execution information of the thread of events, when a lambda waits for another, which in turn waits for another, etc. The following functions deal with backtrace information and exceptions, that propagate through thread of events.

catch $coderef, $event

Registers $coderef on $event, that is called when $event is aborted via either cancel_event, cancel_all_event, or terminate:

   my $resource = acquire;
   context lambda { .. $resource .. };
   catch {
      $resource-> free;
   } tail {
      $resource-> free;

catch must be invoked after a condition, but in the syntax above that means that catch should lexically come before it. If undesirable, use explicit event reference:

   my $event = tail { ... };
   catch   { ... }, $event;
autocatch $event

Prefixes a condition, so that it is called even if cancelled. However, immediately after the call the exception is rethrown. Can be used in the following fashion:

    context lambda ...;
    autocatch tail {
        print "aborted\n" if this-> is_cancelling;
        .. finalize ...

Returns true if running within a catch block.


To be called only from within a catch block. Calls the normal callback that would be called if the event wouldn't be cancelled. @param is passed to the callback.


Terminates the current lambda, then propagates @error to the immediate caller lambdas. They will have a chance to catch the exception with catch later, and re-throw by calling throw again. The default action is to propagate the exception further.

When there are no caller lambdas, a sigthrow callback is called ( analog: die outside eval calls $SIG{__DIE__} ).

sigthrow($callback :: ($lambda, @error))

Retrieves and sets a callback that is invoked when throw is called on lambda that no lambdas wait for. By default, is empty. When invoked, is passed the lambda, and parameters passed to throw.


Returns event records that watch for the lambda.


Returns event records that corresponds to the lambdas this lambda watches.


Returns a IO::Lambda::Backtrace object that represents thread of events which leads to the current lambda. See IO::Lambda::Backtrace for more.


Included modules


Various sub-modules can be controlled with the single environment variable, IO_LAMBDA_DEBUG, which is treated as a comma-separated list of modules. For example,

      env IO_LAMBDA_DEBUG=io=2,http perl

displays I/O debug messages from IO::Lambda (with extra verbosity) and from IO::Lambda::HTTP. IO::Lambda responds for the following keys:


Prints debugging information about file and timeout asynchronous events.


Print debugging information about event flow of lambda objects, where one object waits for another, lambda being cancelled, finished, etc.


Increase verbosity of lambda by storing information about which line invoked object creation and subscription. See IO::Lambda::Backtrace for more.


If set, fatal errors dump the stack trace.


Sets loop module, one of: Select, AnyEvent, Prima, POE.

Keys recognized for the other modules: select,dbi,http,https,signal,message,thread,fork,poll,flock.

Online information

Project homepage:

Mailing list: io-lambda-general at, thanks to sourceforge. Subscribe by visiting


  • A single-process TCP client and server; server echoes back everything is sent by the client. 500 connections sequentially created, instructed to send a single line to the server, and destroyed.

                             2.4GHz x86-64 linux 1.2GHz win32
      Lambda/select              0.697            7.468
      Lambda/select, optimized   0.257            5.273
      Lambda/AnyEvent            0.648            8.175
      Lambda/AnyEvent, optimized                  7.087
      Raw sockets using select   0.149            4.859
      POE/select, components     1.185           12.306
      POE/select, raw sockets    0.382            6.233
      POE/select, optimized      0.770            7.510

    See benchmarking code in eg/bench.


There are many async libraries readily available from CPAN. IO::Lambda is yet another one. How is it different from the existing tools? Why use it? To answer these questions, I need to show the evolution of async libraries, to explain how they grew from simple tools to complex frameworks.

First, all async libraries are based on OS-level syscalls, like select, poll, epoll, kqueue, and Win32::WaitForMultipleObjects. The first layer provides access to exactly these facilities: there are IO::Select, IO::Epoll, IO::Kqueue etc. I won't go deeper into describing pros and cons for programming on this level, this should be obvious.

Perl modules of the next abstraction layer are often characterised by portability and event loops. While the modules of the first layer are seldom portable, and have no event loops, the second layer modules strive to be OS-independent, and use callbacks to ease the otherwise convoluted ways async I/O would be programmed. These modules mostly populate the "asynchronous input-output programming frameworks" niche in the perl world. The examples are many: IO::Events, EV, AnyEvent, IO::NonBlocking, IO::Multiplex, to name a few.

Finally, there's the third layer of complexity, which, before IO::Lambda, had a single representative: POE (now, to the best of my knowledge, IO::Async also partially falls in this category). Modules of the third layer are based on concepts from the second, but introduce a powerful tool to help the programming of complex protocols, something that isn't available in the second layer modules: finite state machines (FSMs). The FSMs reduce programming complexity, for example, of intricate network protocols, that are best modelled as a set of states in a logical circuit. Also, the third layer modules are agnostic of the event loop module: the programmer is (almost) free to choose the event loop backend, such as native select, Gtk, EV, Prima, or AnyEvent, depending on the nature of the task.

IO::Lambda allows the programmer to build protocols of arbitrary complexity, and is also based on event loops, callbacks, and is portable. It differs from POE in the way the FSMs are declared. Where POE requires an explicit switch from one state to another, using f.ex. post or yield commands, IO::Lambda incorporates the switching directly into the program syntax. Consider POE code:

   POE::Session-> create(
       inline_states => {
           state1 => sub { 
              print "state1\n";
              $_[ KERNEL]-> yield("state2");
           state2 => sub {
              print "state2\n";

and the corresponding IO::Lambda code (state1 and state2 are conditions, they need to be declared separately):

    lambda {
       state1 {
          print "state1\n";
       state2 {
          print "state2\n";

In IO::Lambda, the programming style is (deliberately) not much different from the declarative

    print "state1\n";
    print "state2\n";

as much as the nature of asynchronous programming allows that.

To sum up, the intended use of IO::Lambda is for areas where simple callback-based libraries require lots of additional work, and where state machines are beneficial. Complex protocols like HTTP, parallel execution of several tasks, strict control of task and protocol hierarchy - this is the domain where IO::Lambda works best.


This work is partially sponsored by capmon ApS.

This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.


Dmitry Karasik, <>.

I wish to thank those who helped me:

Ben Tilly for providing thorough comments to the code in the synopsis, bringing up various important issues, valuable discussions, for his patience and dedicated collaboration.

David A. Golden for discussions about names, and his propositions to rename some terms into more appropriate, such as "read" to "readable", and "predicate" to "condition". Rocco Caputo for optimizing the POE benchmark script. Randal L. Schwartz, Brock Wilcox, and zby@perlmonks helped me to understand how the documentation for the module could be made better.

All the good people on and perl conferences, who invested their time into understanding the module.