IPEP 13: Updating the Message Spec

Min RK edited this page Jul 15, 2014 · 14 revisions
Clone this wiki locally
Status Implemented
Author Min RK <benjaminrk@gmail.com>
Created April 6, 2013
Updated April 8, 2014
Implementation PR 3190, 4536

See also: IPEP 24: completion and object_info for updates to completion and object_info

We've learned a lot in our first 18 months with the message spec, and we got a remarkable amount right. But we didn't get everything right, and we want to clean some things up prior to 1.0.

Renamed message types

A few names need to change, because they are Python-specific, and our message spec is not.

NOTE: Since these names have notebook format implications, they will not be changed until 3.0

  • pyin will become execute_input
  • pyout will become execute_result
  • pyerr will become error

pyout vs display_data

pyout and display_data are identical messages, with pyout having an additional prompt_number field. It was proposed to merge these two, but preserving the distinction between result and side effect at the message level is desirable.

display metadata

The display_data message currently has a metadata field, which is not yet used for anything. We propose some basic usage guidelines for the metadata field, and some associated changes to the Python API and notebook file format.

In general, there is metadata associated with each output. This is the metadata key of the display_data message. We propose the convention that sub-dicts of the metadata dict, under the same mime-type keys as the data dictionary, should be interpreted as metadata specific to that particular output.

In general, this will mean that rich frontends will want to implement logic similar to the following to handle metadata of a particular output:

data = message['content']['data']
metadata = message['content']['metadata']


for mimetype, mime_data in data.items():
    mime_md = metadata.get(mimetype)
    if mime_md:

In particular, the javascript handlers for each mimetype in the IPython Notebook now receive both the data and the metadata for each individual mimetype, so

OutputArea.prototype.append_html = function (html, element)

has become

OutputArea.prototype.append_html = function (html, md, element)

Where html is the content of data['text/html'], and md is the content of metadata['text/html'].

Python API changes

  • The publish_foo methods from core.displaypub are to be removed. The display_foo methods are our public API, and display_foo(raw=True) has always been an alias to publish_foo, so no functionality is lost.
  • add two kwargs to display:
    • raw: display({'mime': rawdata}, raw=True), so that display is consistent with display_foo(raw=True)
    • metadata: display(obj, metadata={'key': value}), adding a public API to the metadata dict.
  • display_foo methods add an extra, optional metadata key. If metadata is specified, it will be placed in a sub-dict, keyed by the appropriate mimetype for foo.
  • display_foo will no longer result in also displaying the plaintext representation of objects.
  • type formatters, whether registered via formatter.for_type, or _repr_foo_ methods may return either the raw data currently supported, or a tuple of length two, where the first element is the raw data and the second element is a metadata dict. This metadata will be added to the resulting message's metadata under the same mimetype key as the display data itself.
  • metadata is optional - no metadata key must ever be required to be defined. It should always be allowed that the metadata dict is empty.

Notebook format changes

The only change to the notebook format is that each output in a cell's outputs list now has an extra metadata key, which has the same keys as the output itself.

Note: The shortname json, png transforms made to the output dict were a bad decision, and will be changed to proper mime-types in the future.

Third-party use of metadata

The metadata field is a black-box dict, which you can write to. It is expected that third-party code, such as rich javascript plugins, will want to communicate some extra information via the metadata field. It is appropriate for developers to write their info to the metadata field, but some basic practices should be followed to avoid conflict:

  • third-parties should define their own key in the metadata, and only write there.

Actually, that's about it. Just the one guideline to avoid namespace conflicts.

Application: image shapes

Images currently cannot include shape information, which means that one must use raw HTML to display an image with a specific shape. Using the above proposal, we can now support image sizes in the PNG format by defining the following metadata keys:

  "image/png" : {
    "width": 100,
    "height": 200

The same for image/jpeg

The Image display class supports these metadata via its width and height attributes. If specified, the display size of the image will be specified.

For example:

def png_with_shape(img, w, h):
    md = dict(width=w, height=h)
    return png_image_data(img), md

would result in the display_data message:

    "data" : {
        "image/png" : "iVBOR...",
    "metadata" : {
        "image/png" : {
            "height": 4,
            "width": 10

Backward compatibility

This does add a key to the notebook format, and adds an interpretation of a particular subset of the metadata, but these are safely ignored. In this way, notebook frontends that do not support the metadata gracefully degrade, simply losing some layout information, but no data.

multiple objects

Nothing will be done on this issue for this IPEP

We currently have no native mechanism for presenting containers of objects (e.g. a list of images). The only thing we can do is display each individual image via `map(display, list_of_images)`. Nor can we have any representation of an object that might contain multiple elements (such as a pair of images to be displayed side-by-side), or objects for which multiple components *should* be displayed (e.g. LaTeX + PNG). For all of these cases, the only option is to do the entire HTML rendering kernel-side, and use raw HTML-reprs. This sucks quite a bit, but I don't have a good answer for it.

Perhaps JSON reprs / js widgets are going to be the only way to do this sort of thing.


Add a password boolean key to input_request, to indicate that frontends should not echo input.

Message headers

Message headers will have the protocol version (5.0 as of IPython 3.0) as a string. If the message header lacks a version key, it should be assumed to be 4.1, the last version before this update.


  • version keys in kernel_info_reply messages will be strings, not version_info-style lists.
  • A banner key is added, for the kernel's contribution to console frontend banners.
  • instead of ipython_version, implementation and implementation_version keys are added.

The IPython kernel's kernel_info_reply in IPython 3.0:

  "protocol_version" : "5.0.0",
  "implementation" : "ipython",
  "implementation_version" : "3.0.0",
  "language" : "python",
  "language_version" : "3.4.0",
  "banner" : "Python 3.4.0 (/usr/local/bin/python)\nIPython 3.0.0\n",


The mime-type structure used in display_data is useful in more contexts than just display messages. It will be adopted in other places where representation information is requested, specifically:

  • object_info_reply
  • user_expressions
  • pager payload

A mime-bundle takes the form:

  "status" : "ok",
  "data" : {
    "mime-type" : "mime-type-data",
  "metadata" : {
    "mime-type" : {
      "metadata" : "about mime-type",
  • if status == 'ok', then 'data' and 'metadata' must be defined.
  • if status == 'error', then 'ename', 'evalue', and 'etraceback' must be defined, as described in the message spec.

In some contexts, the status = 'error' case may not make sense, but status=='ok' should always be checked before working with data.


IPEP 24: completion and object_info describes turning object_info_reply with a mime-bundle message.

JSON data

The `application/json`` key in a mime bundle stores JSON-serialized data, rather than JSON-able data. This results in serializing this data twice, and having to deserialize it twice on the frontend. The JSON key will now be unserialized, rather than the serialized form.

user_expressions / user_variables

Since variable names are expressions, user_variables will be removed, and only user_expressions remaining.

The user_expressions and user_variables keys in an execute_request currently only support plaintext reprs, and have bad custom representations of errors. These will be replaced with mime-bundles (like display).


The entire Payload model will be revisited in a separate IPEP.

Currently, display payloads list the full object path of the source, which doesn't make any sense. It should be a simple identifier, so that moving the implementation does not need to be reflected in the frontend.

pager mime-bundle

The pager payload will also become a mime-bundle, instead of having just a 'text' key.

An example execute_reply:

  "status" : "ok",
  "execution_count" : 4,
  "user_expressions" : {
    "foo" : {
      "status": "ok",
      "data" : {
        "text/plain" : "5",
      "metadata" : {},
  "payloads" : [
      "source" : "pager",
      "status" : "ok",
      "data" : {
        "text/plain" : "pager data...",
      "metadata" : {},