Migrating to a new Liquidsoap version

In this page, we list the most common catches when migrating to a new version of Liquidsoap.


If you are installing via opam, it can be useful to create a new switch to install the new version of liquidsoap. This will allow to test the new version while keeping the old version around in case you to revert to it.

More generally, we recommend to always keep a version of your script around and also to make sure that you test your new script with a staging environment that is close to production. Streaming issues can build up over time. We do our best to release the most stable possible code but problems can arise from many reasons so, always best to first to a trial run before putting things to production!

From 2.2.x to 2.3.x

Crossfade transitions and track marks

Track marks can now be properly passed through crossfade transitions. This means that you also have to make sure that your transition function is fallible! For instance, this silly transition function:

def transition(_, _) =

Will never terminate!

Typically, to insert a jingle you would do:

def transition(old, new) =
  sequence([old.source, single("/path/to/jingle.mp3"), new.source])


  • There is a new metadata.replaygain function that extracts the replay gain value in dB from the metadata. It handles both r128_track_gain and replaygain_track_gain internally and returns a single unified gain value.

  • The file.replaygain function now takes a new compute parameter: file.replaygain(~id=null(), ~compute=true, ~ratio=50., file_name). The compute parameter determines if gain should be calculated when the metadata does not already contain replaygain tags.

  • The enable_replaygain_metadata function now accepts a compute parameter to control replaygain calculation.

  • The replaygain function no longer takes an ebu_r128 parameter. The signature is now simply: replaygain(~id=null(), s). Previously, ebu_r128 allowed controlling whether EBU R128 or standard replaygain was used. However, EBU R128 data is now extracted directly from metadata when available. So replaygain cannot control the gain type via this parameter anymore.


check_next in playlist operators is now called before the request is resolved, to make it possible to cut out unwanted requests before consuming process time. If you need to see the request’s metadata or if the request resolves into a valid tile, however, you might need to call request.resolve inside your check_next script.


The default port for the Prometheus metrics exporter has changed from 9090 to 9599. As before, you can change it with settings.prometheus.server.port := <your port value>.

From 2.1.x to 2.2.x


The !x notation for getting the value of a reference is now deprecated. You should write x() instead. And x := v is now an alias for x.set(v) (both can be used interchangeably).

Icecast and Shoutcast outputs

output.icecast and output.shoutcast are some of our oldest operators and were in dire need of some cleanup so we did it!

We applied the following changes:

  • You should now use output.icecast only for sending to icecast servers and output.shoutcast only for sending to shoutcast servers. All shared options have been moved to their respective specialized operator.
  • Old icy_metadata argument was renamed to send_icy_metadata and changed to a nullable bool. null means guess.
  • New icy_metadata argument now returns a list of metadata to send with ICY updates.
  • Added a icy_song argument to generate default "song" metadata for ICY updates. Defaults to <artist> - <title> when available, otherwise artist or title if available, otherwise null, meaning don’t add the metadata.
  • Cleaned up and removed parameters that were irrelevant to each operator, i.e. icy_id in output.icecast and etc.
  • Made mount mandatory and name nullable. Use mount as name when name is null.

HLS events

Starting with version 2.2.1, on HLS outputs, on_file_change events are now "created", "updated" and "deleted". This breaking was required to reflect the fact that file changes are now atomic. See this issue for more details.


Starting with version 2.2.4, the cue_cut operator has been removed. Requests cue-in and cue-out processing has been integrated directly into requests resolution. In most cases, you simply can remove the operator from your script. In some cases, you might need to disable cue_in_metadata and cue_out_metadat either when creating new requests or when creating playlist sources.

Harbor HTTP server and SSL support

The API for registering HTTP server endpoint and using SSL was completely rewritten. It should be more flexible and provide node/express like API for registering endpoints and middleware. You can checkout the harbor HTTP documentation for more details. The Https support section also explains the new SSL/TLS API.


We used to have timeout values labelled timeout or timeout_ms, some of these would be integer and in milliseconds, other floating point and in seconds etc. This was pretty confusing so, now all timeout settings and arguments have been unified to be named timeout and hold a floating point value representing a number of seconds.

In most cases, your script will not execute until you have updated your custom timeout values but you should also review all of them to make sure that they follow the new convention.

Metadata overrides

Some metadata overrides have been made to reset on track boundaries. Previously, those were permanent even though they were documented as only applying to the current track. If you need to keep the previous behavior, you can used the persist_overrides parameters (persis_override for cross/crossfade).

The list of concerned metadata is:

  • "liq_fade_out"
  • "liq_fade_skip"
  • "liq_fade_in"
  • "liq_cross_duration"
  • "liq_fade_type"

JSON rendering

The confusing let json.stringify syntax has been removed as it did not provide any feature not already covered by either the json.stringify() function or the generic json() object mapper. Please use either of those now.

Default character encoding in output.{harbor,icecast,shoutcast}

Default encoding for output.harbor, output.icecast and output.shoutcast metadata has been changed to UTF-8 in all cases.

Legacy systems used to expect ISO-8859-1 (also known as latin1) for metadata inserted into mp3 streams via the icy mechanism.

It seems that, nowadays, most software expect UTF-8 out of the box, including for legacy systems that previously assumed other encodings. Therefore, by changing this default value, we try to match expectations of the largest number of users of our software.

If you are using one of these outputs, make sure to test this assumptions with your listners’ clients. If needed, the characters encoding can be set to a different value using the operator’s parameters.

Decoder names

Decoder names have been converted to lowercase. If you were relying on specific settings for decoders priority/ordering, you will need to convert them to lowercase, for instance:




Actually, because of the above change in references, this even becomes:

settings.decoder.decoders := ["ffmpeg"]


Add file-based operators do not support strftime type conversions out of the box anymore. Instead, you should use explicit conversions using time.string. This means that this script:

output.file("/path/to/file%H%M%S.wav", ...)


output.file({time.string("/path/to/file%H%M%S.wav")}, ...)

Other breaking changes

  • reopen_on_error and reopen_on_metadata in output.file an related outputs are now callbacks.
  • request.duration now returns a nullable float, null being value returned when the request duration could not be computed.
  • getenv (resp. setenv) has been renamed to environment.get (resp. environment.set).

From 2.0.x to 2.1.x

Regular expressions

First-class regular expression are introduced and are used to replace the following operators:

  • string.match(pattern=<regexp>, <string> is replaced by: r/<regexp>/.test(<string>)
  • string.extract(pattern=<regexp>, <string>) is replaced by: r/<regexp>/.exec(<string>)
  • string.replace(pattern=<regexp>, <string>) is replaced by: r/<regexp>/g.replace(<string>)
  • string.split(separator=<regexp>, <string>) is replaced by: r/<regexp>/.split(<string>)

Partial application

In order to improve performance, avoid some programming errors and simplify the code, the support for partial application of functions was removed (from our experience it was not used much anyway). This means that you should now provide all required arguments for functions. The behavior corresponding to partial application can of course still be achieved by explicitly abstracting (with fun(x) -> ...) over some arguments.

For instance, suppose that we defined the addition function with two arguments with

def add(x,y) =
  x + y

and defined the successor function by partially applying it to the first argument

suc = add(1)

We now need to explicitly provide the second argument, and the suc function should now be defined as

suc = fun(x) -> add(1, x)


def suc(x) =
  add(1, x)

JSON parsing

JSON parsing was greatly improved and is now much more user-friendly. You can check out our detailed presentation here.

Runtime evaluation

Runtime evaluation of strings has been re-implemented as a type-safe eval let decoration. You can now do:

let eval x = "[1,2,3]"

And, just like with JSON parsing, the recommended use is with a type annotation:

let eval (x: [int]) = "[1,2,3]"

Deprecations and breaking changes

  • The argument streams_info of output.file.hls is now a record.
  • Deprecated argument timeout of http.* operators.
  • source.on_metadata and source.on_track now return a source as this was the case in previous versions, and associated handlers are triggered only when the returned source is pulled
  • output.youtube.live renamed output.youtube.live.rtmp, remove bitrate and quality arguments and added a single encoder argument to allow stream copy and more.
  • list.mem_assoc is replaced by list.assoc.mem
  • timeout argument in http.* operators is replaced by timeout_ms.
  • request.ready is replaced by request.resolved

From 1.4.x to 2.0.0


audio_to_stereo should not be required in most situations anymore. liquidsoap can handle channels conversions transparently now!

auth function in input.harbor

The type of the auth function in input.harbor has changed. Where before, you would do:

def auth(user, password) =

You would now do:

def auth(params)
  user     = params.user
  password = params.password

Type errors with lists of sources

Now that sources have their own methods, the actual list of methods attached to each source can vary from one to the next. For instance, playlist has a reload method but input.http does not. This currently confuses the type checker and leads to errors that look like this:

At script.liq, line xxx, char yyy-zzz:
Error 5: this value has type
  _ * source(audio=?A, video=?B, midi=?C)
    time : () -> float,
    shutdown : () -> unit,
    fallible : bool,
    skip : () -> unit,
    seek : (float) -> float,
    is_active : () -> bool,
    is_up : () -> bool,
    log :
    {level : (() -> int?).{set : ((int) -> unit)}
    self_sync : () -> bool,
    duration : () -> float,
    elapsed : () -> float,
    remaining : () -> float,
    on_track : ((([string * string]) -> unit)) -> unit,
    on_leave : ((() -> unit)) -> unit,
    on_shutdown : ((() -> unit)) -> unit,
    on_metadata : ((([string * string]) -> unit)) -> unit,
    is_ready : () -> bool,
    id : () -> string,
    selected : (() -> source(audio=?D, video=?E, midi=?F)?)
but it should be a subtype of the type of the value at radio.liq, line 122, char 2-21
  _ * _.{reload : _}

In such cases, we recommend to give a little nudge to the typechecker by using the (s:source) type annotation where a list of source is causing the issue. For instance:

s = fallback([

This tells the type checker not to worry about the source methods and just focus on what matters, that they are actually sources.. 🙂

Http input and operators

In order to provide as much compatibility as possible with the different HTTP protocols and implementation, we have decided to delegate HTTP support to external libraries which have large scale support and implementation. This means that, if you have installed liquidsoap using opam:

  • You need to install the ocurl package to enable all HTTP request operators, http.get, http.post, http.put, http.delete and http.head
  • You need to install the ffmpeg package (version 1.0.0 or above) to enable input.http
  • You do not need to install the ssl package anymore to enable their https counter-part. These operators have been deprecated.


The parameters for cross transitions was changed to take advantage of the new module system. Instead of passing multiple arguments related to the ending and starting track, those are regrouped into a single record. So, if you had a transition like this:

def transition(
  ending_dB_level, starting_dB_level,
  ending_metadata, starting_metadata,
  ending_source,   starting_source) =

You would now do:

def transition(ending, starting) =
  # Now you can use:
  #  - ending.db_level, ending.metadata, ending.source
  #  - starting.db_level, starting.metadata, starting.source


Settings are now exported as records. Where you would before write:

set("decoder.decoders", ["MAD", "FFMPEG"])

You can now write:

settings.decoder.decoders.set(["MAD", "FFMPEG"])

Likewise, to get a setting’s value you can now do:

current_decoders = settings.decoder.decoders()

This provides many good features, in particular type-safety.

For convenience, we have added shorter versions of the most used settings. These are all shortcuts to their respective settings values:


The register operator could not be adapted to this new API and had to be removed, however, backward-compatible set and get operators are provided. Make sure to replace them as they should be removed in a future version.

Metadata insertion

The function insert_metadata does not return a pair anymore, but a source with a method named insert_metadata. This means that you should change the code

fs = insert_metadata(s)
# The function to insert metadata
f = fst(ms)
# The source with inserted metadata
s = snd(ms)
# Using the function
f([("artist", "Bob")])
# Using the source


s = insert_metadata(s)
# Using the function
s.insert_metadata([("artist", "Bob")])
# Using the source

Request-based queueing

Queueing for request-based sources has been simplified. The default_duration and length have been removed in favor of a simpler implementation. You can now pass a prefetch parameter which tells the source how many requests should be queued in advance.

Should you need more advanced queueing strategy, request.dynamic.list and request.dynamic now export functions to retrieve and set their own queue of requests.

JSON import/export

json_of has been renamed json.stringify and of_json has been renamed json.parse.

JSON export has been enhanced with a new generic json object export. Associative lists of type (string, 'a) are now exported as lists. See our JSON documentation page for more details.

Convenience functions have been added to convert metadata to and from JSON object format: metadata.json.stringify and metadata.json.parse.

Returned types from output operators

Starting with liquidsoap 2.0.0, output operators return the empty value () while they previously returned a source.

This helps enforce the fact that outputs should be end-points of your scripting graphs. However, in some cases, this can cause issues while migrating old scripts, in particular if the returned value of an output was used in the script.

The way to fix this is to apply your operator to the source directly underneath the output. For instance, the following clock assignment:

s = ...

clock.assign_new([output.icecast(..., s)])

Should now be written:

s = ...

clock.assign_new([s], ...)

output.icecast(..., s)

Deprecated operators

Some operators have been deprecated. For most of them, we provide a backward-compatible support but it is good practice to update your script. You should see logs in your script when running deprecated operatords. Here’s a list of the most important ones:

  • playlist.safe is replaced by: playlist(mksafe(..))
  • playlist.once is replaced by: playlist, setting reload_mode argument to "never" and loop to false
  • rewrite_metadata should be rewritten using metadata.map
  • fade.initial and fade.final are not needed anymore
  • get_process_output is replaced by: process.read
  • get_process_lines is replaced by: process.read.lines
  • test_process is replaced by: process.test
  • system is replaced by: process.run
  • add_timeout is replaced by: thread.run.recurrent
  • on_blank is replaced by: blank.detect
  • skip_blank is replaced by: blank.skip
  • eat_blank is replaced by: blank.eat
  • strip_blank is replaced by: blank.strip
  • which is replaced by: file.which
  • register_flow: flow is no longer maintained
  • empty is replaced by: source.fail
  • file.unlink is replaced by: file.remove
  • string.utf8.escape is replaced by: string.escape
  • metadata.map is replaced by: metadata.map

Windows build

The windows binary is statically built and, for this reason, we cannot enable both the %ffmpeg encoder and any encoder that uses the same underlying libraries, for instance libmp3lame for mp3 encoding. The technical reason is that both libraries import the same C symbols, which makes compilation fail.

The %ffmpeg encoder provides all the functionalities of the internal encoders that conflict with it along with many more format we do not support otherwise. For this reason, it was decided to enable the %ffmpeg encoder and disable all other encoders.

This means that, if you were previously using a different encoder than %ffmpeg, you will need to adapt your script to use it. For instance, for mp3 encoding with variable bitrate:

%ffmpeg(format="mp3", %audio(codec="libmp3lame", q=7))