How can I delete all keys that don't match certain names with JQ?

Question

I have a huge JSON file with lots of stuff I don't care about, and I want to filter it down to only the few keys I care about, preserving the structure. I won't bother if the same key name might occur in different paths and I get both of them. I gleaned something very close from the answers to this question, it taught me how to delete all properties with certain values, like all null values:

del(..|nulls)

or, more powerfully

del(..|select(. == null))

I searched high and low if I could write a predicate over the name of a property when I am looking at a property. I come from XSLT where I could write something like this:

del(..|select(name|test("^(foo|bar)$")))

where name/1 would be the function that returns the property name or array index number where the current value comes from. But it seems that jq lacks the metadata on its values, so you can only write predicates about their value, and perhaps the type of their value (that's still just a feature of the value), but you cannot inspect the name, or path leading up to it?

I tried to use paths and leaf_paths and stuff like that, but I have no clue what that would do and tested it out to see how this path stuff works, but it seems to find child paths inside an object, not the path leading up to the present value.

So how could this be done, delete everything but a set of key values? I might have found a way here:

walk(
  if type == "object" then 
    with_entries(
      select(    ( .key |test("^(foo|bar|...)$") )
             and ( .value != "" )
             and ( .value != null ) )
    )
  else 
    . 
  end
)

OK, this seems to work. But I still wonder it would be so much easier if we had a way of querying the current property name, array index, or path leading up to the present item being inspected with the simple recusion ..| form.

Answer 1

Here's a solution that uses a specialized variant of walk for efficiency (*). It retains objects all keys of which are removed; only trivial changes are needed if a blacklist or some other criterion (e.g., regexp-based) is given instead. WHITELIST should be a JSON array of the key names to be retained.

jq --argjson whitelist WHITELIST '
def retainKeys($array):
  INDEX($array[]; .) as $keys
  | def r:
      if type == "object" 
      then with_entries( select($keys[.key]) )
      | map_values( r )
      elif type == "array" then map( r )
      else .
      end;
  r;

  retainKeys($whitelist)

' input.json

---

(*) Note for example:

1. the use of INDEX
2. the recursive function, r, has arity 0
3. for objects, the top-level deletion occurs first.

Answer 2

In analogy to your approach using .. and del, you could use paths and delpaths to operate on a stream of path arrays, and delete a given path if not all of its elements meet your conditions.

delpaths([paths | select(all(IN("foo", "bar") or type == "number") | not)])

For the condition I used IN("foo", "bar") but (type == "string" and test("^(foo|bar)$")) would work as well. To also retain array elements (which have numeric indices), I added or type == "number".

Answer 3

Unlike in XML, there's no concept of attributes in jq. You'll need to delete from objects.

To delete an element of an object, you need to use del( obj[ key ] ) (or use with_entries). You can get a stream of the keys of an object using keys[]/keys_unsorted[] and filter out the ones you don't want to delete.

Finally, you need to invert the result of test because you want to delete those that don't match.

---

After fixing these problems, we get the following:

INDEX( "foo", "bar" ) as $keep |
del(
  .. | objects |
  .[
     keys_unsorted[] |
     select( $keep[ . ] | not )
  ]
)

Demo on jqplay

Note that I substituted the regex match with a dictionary lookup. You could use test( "^(?:foo|bar)\\z" ) in lieu of $keep[ . ], but a dictionary lookup should be faster than a regex match. And it should be less error-prone too, considering you misused $ and (...) in lieu of \z and (?:...).

---

The above visits deleted branches for nothing. We can avoid that by using walk instead of ...

INDEX( "foo", "bar" ) as $keep |
walk(
   if type == "object" then 
      del(
         .[
            keys_unsorted[] |
            select( $keep[ . ] | not )
         ]
      )
   else
      .
   end
)

Demo on jqplay

---

Since I mentioned one could use with_entries instead of del, I'll demonstrate.

INDEX( "foo", "bar" ) as $keep |
walk(
   if type == "object" then 
      with_entries( select( $keep[ .key ] ) )
   else
      .
   end
)

Demo on jqplay

Answer 4

Here's a space-efficient, walk-free approach, tailored for the case of a WHITELIST. It uses the so-called "streaming" parser, so the invocation would look like this:

jq -n --stream --argjson whitelist WHITELIST -f program.jq input.json

where WHITELIST is a JSON array of the names of the keys to be deleted, and where program.jq is a file containing the program:

# Input: an array
# Output: the longest head of the array that includes only numbers or items in the dictionary
def acceptable($dict):
  last(label $out
       | foreach .[] as $x ([];
           if ($x|type == "number") or $dict[$x] then . + [$x]
           else ., break $out
           end));

INDEX( $whitelist[]; .) as $dict
| fromstream(inputs
    | if length==2
      then (.[0] | acceptable($dict)) as $p
      | if   ($p|length) == (.[0]|length) - 1 then .[0] = $p | .[1] = {}
        elif ($p|length) <  (.[0]|length) then empty
        else .
        end
      else .
      end )

Note: The reason this is relatively complicated is that it assumes that you want to retain objects all of whose keys have been removed, as illustrated in the following example. If that is not the case, then the required jq program is much simpler.

---

Example:

WHITELIST: '["items", "config", "spec", "setting2", "name"]'

input.json:

{
  "items": [
    {
      "name": "issue1",
      "spec": {
        "config": {
          "setting1": "abc",
          "setting2": {
            "name": "xyz"
          }
        },
        "files": {
          "name": "cde",
          "path": "/home"
        },
        "program": {
          "name": "apache"
        }
      }
    },
    {
      "name": {
        "etc": 0
      }
    }
  ]
}

Output:

{
  "items": [
    {
      "name": "issue1",
      "spec": {
        "config": {
          "setting2": {
            "name": "xyz"
          }
        }
      }
    },
    {
      "name": {}
    }
  ]
}

Answer 5

I am going to put my own tentative answer here.

The thing is, the solution I had already in my question, meaning I can select keys during forward navigation, but I cannot find out the path leading up to the present value.

I looked around in the source code of jq to see how come we cannot inquire the path leading up to the present value, so we could ask for the key string or array index of the present value. And indeed it looks like jq does not track the path while it walks through the input structure.

I think this is actually a huge opportunity forfeited that could be so easily kept track during the tree walk.

This is why I continue thinking that XML with XSLT and XPath is a much more robust data representation and tool chain than JSON. In fact, I find JSON harder to read even than XML. The benefit of the JSON being so close to javascript is really only relevant if - as I do in some cases - I read the JSON as a javascript source code assigning it to a variable, and then instrument it by changing the prototype of the anonymous JSON object so that I have methods to go with them. But changing the prototype is said to cause slowness. Though I don't think it does when setting it for otherwise anonymous JSON objects.

There is JsonPath that tries (by way of the name) to be something like what XPath is for XML. But it is a poor substitute and also has no way to navigate up the parent (or then sibling) axes.

So, in summary, while selecting by key in white or black lists is possible in principle, it is quite hard, because a pretty easy to have feature of a JSON navigation language is not specified and not implemented. Other useful features that could be easily achieved in jq is backward navigation to parent or ancestor of the present value. Currently, if you want to navigate back, you need to capture the ancestor you want to get back to as a variable. It is possible, but jq could be massively improved by keeping track of ancestors and paths.