Agreed, as I absolutely like that schema is enforced. I think the semantics I bring up here is how “empty/non-existent” is represented in the actual DB storage. The rails codebase I’m migrating from enforced jsonb schema, but represented empty (which is different from null), as truly empty in the storage.
But Ecto was awesome enough for me to get to the representation I desired. Here’s the custom type I created, which was inspired by the links I shared earlier. Only disadvantage of using the map type is that I cannot represent incremental additions to the stored data. Its all or nothing.

defmodule Ecto.CompactEmbed do
  @moduledoc """
  Custom ecto type for embedded map types that get compacted on saving to data.
  "Raw" ecto embedded schemas include all fields in the json representation in the database even if the field is not specified.
  This compacts the json fields by removing all null fields or applying the supplied `:on_compact` option.

  For more discussion see https://forum.elixirforum.com/t/ecto-embedded-schema-storing-extra-default-values-for-fields-that-arent-specified/35739

  ## Options

    * `:schema` (*Required*) - The (embedded) schema module would otherwise have been specified as the field type.
      By default the schema's changeset/2 function is used to cast the data if provided.
      Otherwise the data is casted to only include the fields defined.

    * `:on_compact` - The function that is invoked to compact the map data before writing to the database.
      This must be function with an arity of 1 that takes the schema map data and compacts/modifies it.
      An MFA tuple may be passed as well.  For example `on_compact: {Utils, :remove_empty}` or `on_compact: {Utils, :remove_empty, []}`
      If this is not specified, all null fields will be removed by default

  ## Example:
    # Define schema of compacted embed
    defmodule Asset do
      use Ecto.Schema

      @primary_key false

      embedded_schema do
        field :url, :string
        field :filename, :string
        field :mime_type, :string
        field :size, :integer
        field :width, :integer
        field :height, :integer
      end
    end

    # Use in schema field
    defmodule Attachment do
      use Ecto.Schema

      schema "attachments" do
        field :asset, Ecto.CompactEmbed, schema: Asset
      end
  """

  # Loosely modeled after:
  #   https://github.com/elixir-ecto/ecto/blob/master/lib/ecto/embedded.ex
  #   https://github.com/mathieuprog/polymorphic_embed/blob/master/lib/polymorphic_embed.ex

  use Ecto.ParameterizedType

  alias __MODULE__

  @type t :: %CompactEmbed{
          schema: Ecto.Schema.t(),
          on_compact: fun() | {module(), atom()} | {module(), atom(), []}
        }
  defstruct [:schema, :on_compact]

  ## Type behavior

  @impl true
  def type(_params), do: :map

  @impl true
  def init(opts) do
    schema = Keyword.fetch!(opts, :schema)
    on_compact = opts[:on_compact]

    on_compact =
      case on_compact do
        nil ->
          &Utils.Presence.sans_nil(&1)

        {mod, fun} when is_atom(mod) and is_atom(fun) ->
          fn map ->
            apply(mod, fun, [map])
          end

        {mod, fun, []} when is_atom(mod) and is_atom(fun) ->
          fn map ->
            apply(mod, fun, [map])
          end

        compacter when is_function(compacter, 1) ->
          compacter

        _ ->
          raise ArgumentError,
                "invalid `:on_compact` option for #{inspect(on_compact)}. " <>
                  "This must be a function with an arity of 1 that takes a map and returns the compacted/modified map." <>
                  "An MFA may be passed as well.  For example `on_compact: {Utils, :remove_empty}` or `on_compact: {Utils, :remove_empty, []}`"
      end

    %CompactEmbed{schema: schema, on_compact: on_compact}
  end

  @impl true
  def cast(nil, _), do: {:ok, nil}

  def cast(attrs, %{schema: schema}) do
    schema
    |> cast_to_changeset(attrs)
    |> case do
      %{valid?: true} = changeset ->
        {:ok, Ecto.Changeset.apply_changes(changeset)}

      changeset ->
        {:error, build_errors(changeset)}
    end
  end

  @impl true
  def load(nil, _loader, _params), do: {:ok, nil}

  def load(data, _loader, %{schema: schema}) do
    struct =
      cast_to_changeset(schema, data)
      |> Ecto.Changeset.apply_changes()

    {:ok, struct}
  end

  @impl true
  def dump(nil, _dumber, _params), do: {:ok, nil}

  def dump(data, _dumper, %{on_compact: on_compact}) do
    case Ecto.Type.dump(:map, data) do
      {:ok, map} -> {:ok, compact_map(map, on_compact)}
      error -> error
    end
  end

  @impl true
  def embed_as(_, _), do: :dump

  @impl true
  def equal?(term1, term2, %{on_compact: on_compact}) when is_map(term1) and is_map(term2) do
    case MapDiff.diff(compact_map(term1, on_compact), compact_map(term2, on_compact)) do
      %{changed: :equal} -> true
      _ -> false
    end
  end

  def equal?(term1, term2, _params), do: term1 == term2

  ## Privates

  defp cast_to_changeset(%schema{} = struct, attrs) do
    if function_exported?(schema, :changeset, 2) do
      schema.changeset(struct, attrs)
    else
      embed_fields = schema.__schema__(:embeds)
      non_embed_fields = schema.__schema__(:fields) -- embed_fields

      Ecto.Changeset.cast(struct, attrs, non_embed_fields)
      |> cast_embeds_to_changeset(embed_fields)
    end
  end

  defp cast_to_changeset(module, attrs) when is_atom(module) do
    cast_to_changeset(struct(module), attrs)
  end

  defp cast_embeds_to_changeset(changeset, embed_fields) do
    Enum.reduce(embed_fields, changeset, fn embed_field, changeset ->
      Ecto.Changeset.cast_embed(
        changeset,
        embed_field,
        with: fn embed_struct, data ->
          cast_to_changeset(embed_struct, data)
        end
      )
    end)
  end

  defp compact_map(%_module{} = struct, on_compact) do
    struct
    |> Map.from_struct()
    |> compact_map(on_compact)
  end

  defp compact_map(map, on_compact) when is_map(map), do: on_compact.(map)

  defp build_errors(%{errors: errors, changes: changes}) do
    Enum.reduce(changes, errors, fn {field, value}, all_errors ->
      case value do
        %Ecto.Changeset{} = changeset ->
          Keyword.merge([{field, {"is invalid", changeset.errors}}], all_errors)

        _ ->
          all_errors
      end
    end)
  end
end

Oh yes! I meant to contribute this as a hex package. Only hesitation is that I haven’t released a hex package before. It ended up being a couple of modules that took heavy inspiration from Ecto embeds, and it’s been working great. Let me see if I can find guidance on releasing hex packages so I can release as version 0.0.1 and refine as others try it out

I haven’t measured. But jsonb (i.e. binary json) columns store not only “null” as values (which are likely optimized), but also the corresponding map keys (i.e. field names) for every record. These keys don’t hold any useful information, so still a waste. My particular case above is just one permutation of several factors i.e. number of fields in embeds and number of records. And I’m migrating data from another platform where json was “structured” but didn’t have the unneeded fields stored in the columns. So that’s driving my expectations here. I can measure and report something quantitative, but on a fundamental level, feels like having only the specified fields in the data is not a stretch of expectation