Merging User-Data Sections

Overview

This was implemented because it has been a common feature request that there be a way to specify how cloud-config yaml “dictionaries” provided as user-data are merged together when there are multiple yamls to merge together (say when performing an #include).

Since previously the merging algorithm was very simple and would only overwrite and not append lists, or strings, and so on it was decided to create a new and improved way to merge dictionaries (and their contained objects) together in a way that is customizable, thus allowing for users who provide cloud-config user-data to determine exactly how their objects will be merged.

For example.

#cloud-config (1)
run_cmd:
  - bash1
  - bash2

#cloud-config (2)
run_cmd:
  - bash3
  - bash4

The previous way of merging the two objects above would result in a final cloud-config object that contains the following.

#cloud-config (merged)
run_cmd:
  - bash3
  - bash4

Typically this is not what users want; instead they would likely prefer:

#cloud-config (merged)
run_cmd:
  - bash1
  - bash2
  - bash3
  - bash4

This way makes it easier to combine the various cloud-config objects you have into a more useful list, thus reducing duplication necessary to accomplish the same result with the previous method.

Customizability

Because the above merging algorithm may not always be desired (just as the previous merging algorithm was not always the preferred one), the concept of customized merging was introduced through ‘merge classes’.

A merge class is a class definition which provides functions that can be used to merge a given type with another given type.

An example of one of these merging classes is the following:

class Merger(object):
    def __init__(self, merger, opts):
        self._merger = merger
        self._overwrite = 'overwrite' in opts

    # This merging algorithm will attempt to merge with
    # another dictionary, on encountering any other type of object
    # it will not merge with said object, but will instead return
    # the original value
    #
    # On encountering a dictionary, it will create a new dictionary
    # composed of the original and the one to merge with, if 'overwrite'
    # is enabled then keys that exist in the original will be overwritten
    # by keys in the one to merge with (and associated values). Otherwise
    # if not in overwrite mode the 2 conflicting keys themselves will
    # be merged.
    def _on_dict(self, value, merge_with):
        if not isinstance(merge_with, (dict)):
            return value
        merged = dict(value)
        for (k, v) in merge_with.items():
            if k in merged:
                if not self._overwrite:
                    merged[k] = self._merger.merge(merged[k], v)
                else:
                    merged[k] = v
            else:
                merged[k] = v
        return merged

As you can see there is a ‘_on_dict’ method here that will be given a source value and a value to merge with. The result will be the merged object. This code itself is called by another merging class which ‘directs’ the merging to happen by analyzing the types of the objects to merge and attempting to find a know object that will merge that type. I will avoid pasting that here, but it can be found in the mergers/__init__.py file (see LookupMerger and UnknownMerger).

So following the typical cloud-init way of allowing source code to be downloaded and used dynamically, it is possible for users to inject there own merging files to handle specific types of merging as they choose (the basic ones included will handle lists, dicts, and strings). Note how each merge can have options associated with it which affect how the merging is performed, for example a dictionary merger can be told to overwrite instead of attempt to merge, or a string merger can be told to append strings instead of discarding other strings to merge with.

How to activate

There are a few ways to activate the merging algorithms, and to customize them for your own usage.

  1. The first way involves the usage of MIME messages in cloud-init to specify multipart documents (this is one way in which multiple cloud-config is joined together into a single cloud-config). Two new headers are looked for, both of which can define the way merging is done (the first header to exist wins). These new headers (in lookup order) are ‘Merge-Type’ and ‘X-Merge-Type’. The value should be a string which will satisfy the new merging format defintion (see below for this format).
  2. The second way is actually specifying the merge-type in the body of the cloud-config dictionary. There are 2 ways to specify this, either as a string or as a dictionary (see format below). The keys that are looked up for this definition are the following (in order), ‘merge_how’, ‘merge_type’.

String format

The string format that is expected is the following.

classname1(option1,option2)+classname2(option3,option4)....

The class name there will be connected to class names used when looking for the class that can be used to merge and options provided will be given to the class on construction of that class.

For example, the default string that is used when none is provided is the following:

list()+dict()+str()

Dictionary format

A dictionary can be used when it specifies the same information as the string format (i.e. the second option above), for example:

{'merge_how': [{'name': 'list', 'settings': ['extend']},
               {'name': 'dict', 'settings': []},
               {'name': 'str', 'settings': ['append']}]}

This would be the equivalent format for default string format but in dictionary form instead of string form.

Specifying multiple types and its effect

Now you may be asking yourself, if I specify a merge-type header or dictionary for every cloud-config that I provide, what exactly happens?

The answer is that when merging, a stack of ‘merging classes’ is kept, the first one on that stack is the default merging classes, this set of mergers will be used when the first cloud-config is merged with the initial empty cloud-config dictionary. If the cloud-config that was just merged provided a set of merging classes (via the above formats) then those merging classes will be pushed onto the stack. Now if there is a second cloud-config to be merged then the merging classes from the cloud-config before the first will be used (not the default) and so on. This way a cloud-config can decide how it will merge with a cloud-config dictionary coming after it.

Other uses

In addition to being used for merging user-data sections, the default merging algorithm for merging ‘conf.d’ yaml files (which form an initial yaml config for cloud-init) was also changed to use this mechanism so its full benefits (and customization) can also be used there as well. Other places that used the previous merging are also, similarly, now extensible (metadata merging, for example).

Note, however, that merge algorithms are not used across types of configuration. As was the case before merging was implemented, user-data will overwrite conf.d configuration without merging.