Merging records

In Nickel, the basic building blocks for data are records (objects in JSON or attribute sets in Nix). Merging is a fundamental built-in operation whose role is to combine records together. Fields common to several records will be themselves recursively merged if possible, following the semantics described in this document.

Merging is useful to compose small and logical blocks into a potentially complex final configuration, making it more manageable. A merge is performed by the & operator.

Merge is a symmetric operation. More pedantically, it is commutative. In practice, this means that order doesn't matter, and left & right is the same thing as right & left. When the operands need to be distinguished, as we will see for default values for example, the idea is to use special annotations (metadata) instead of relying on the order of the operands.

Warning: At the time of writing, Nickel's version is 1.0. Custom merge functions are planned for a future version. They are not detailed here yet. See the associated technical document RFC001.

The section describes the behavior and use-cases of merge, by considering the following situations:

Simple merge (no common fields)

Merging two records with no common fields results in a record with the fields from both operands. For example, {foo = 1, bar = "bar"} & {baz =false} evaluates to {foo = 1, bar = "bar", baz = false}.

Specification

Formally, if we write the left operand as

left = {
  field_left_1 = value_left_1,
  ..,
  field_left_n = value_left_n,
}

and the right operand as

right = {
  field_right_1 = value_right_1,
  ..,
  field_right_k = value_right_k
}

then the merge left & right evaluates to the record:

{
  field_left_1 = value_left_1,
  ..,
  field_left_n = value_left_n,
  field_right_1 = value_right_1,
  ..,
  field_right_k = value_right_k
}

In other words, left & right is the union of left and right.

Examples

Split

You can split a configuration into subdomains:

# file: server.ncl
{
    host_name = "example",
    host = "example.org",
    ip_addr = "0.0.0.0",
}

# file: firewall.ncl
{
    enable_firewall = true,
    open_ports = [23, 80, 443],
}

# file: network.ncl
let server = import "server.ncl" in
let firewall = import "firewall.ncl" in
server & firewall

This gives:

{
    host_name = "example",
    host = "example.org",
    ip_addr = "0.0.0.0",
    enable_firewall = true,
    open_ports = [23, 80, 443],
}

Extension

Given a configuration, you can use merge to add new fields:

# file: safe-network.ncl
let base = import "network.ncl" in
base & {use_iptables = true}

Recursive merge (with common fields)

When the two operands have fields in common, those fields are recursively merged. For example:

{
  top_left = 1,
  common = {left = "left"}}
& {
  top_right = 2,
  common = {right = "right"}
}

Evaluates to the record:

{
  top_left = 1,
  top_right = 2,
  common = {left = "left", right = "right"}
}

When one or both of the common fields are not records, the merge will fail unless one of the following condition hold:

  • They are both of the same primitive data type (Number, Bool, EnumTag, or String) and they are equal
  • They are both enum variants, and their tags are equal. In this case, the arguments are merged recursively: that is, 'Tag arg1 & 'Tag arg2 is 'Tag (arg1 & arg2).
  • They are both arrays, and they are equal (checked by generating an application of the delayed contract std.contract.Equal)
  • They are both equal to null

Specification

Formally, let's write the left operand as

left = {
  field_left_1 = value_left_1,
  ..,
  field_left_n = value_left_n,
  common_1 = common_vleft_1,
  ..,
  common_m = common_vleft_m,
}

and the right operand as

right = {
  field_right_1 = value_right_1,
  ..,
  field_right_k = value_right_k
  common_1 = common_vright_1,
  ..,
  common_m = common_vright_m,
}

where the field_left_i and field_right_j are distinct for all i and j. Then the merge left & right evaluates to the record:

{
  field_left_1 = value_left_1,
  ..,
  field_left_n = value_left_n,
  field_right_1 = value_right_1,
  ..,
  field_right_k = value_right_k
  common_1 = common_vleft_1 & common_vright_1,
  ..,
  common_m = common_vleft_m & common_vright_m,
}

For two values v1 and v2, if at least one value is not a record, then

v1 & v2 = v1               if (type_of(v1) is Number, Bool, String, Enum or
                                v1 == null)
                              AND v1 == v2

          v2 | Equal v1    if type_of(v1) is Array and type_of(v2) is Array

          _|_              otherwise (indicates failure)

Example

# file: udp.ncl
{
    # same as firewall = {open_ports = {udp = [...]}},
    firewall.open_ports.udp = [12345,12346],
}

# file: tcp.ncl
{
    # same as firewall = {open_ports = {tcp = [...]}},
    firewall.open_ports.tcp = [23, 80, 443],
}

# firewall.ncl
let udp = import "udp.ncl" in
let tcp = import "tcp.ncl" in
udp & tcp

In the above example, we merge two records, both with a field firewall. On both sides, the value of firewall is again a record, which is therefore merged. The same process happens one layer below, on the common field open_ports, to result in the final record:

{
    firewall = {
      open_ports = {
        udp = [12345, 12346],
        tcp = [23, 80, 443],
      }
    }
}

Merging records with metadata

Metadata can be attached to values thanks to the | operator. Metadata currently include contract annotation, default value, merge priority, optionality annotation, and documentation. We describe in this section how metadata interact with merging.

Note that metadata can only be syntactically attached to record fields, except for type and contract annotations, which can appear anywhere in a freestanding expression such as (x | Num) + 1. However, a contract annotation outside a record field isn't considered metadata (it's a mere contract check) and doesn't behave the same with respect to merging. In particular, {foo | Num = 1} can behave differently from {foo = (1 | Num)} when merged. See the contracts section for more details.

Optional fields

A field can be marked as optional using the optional annotation:

# schema.ncl
{
  Command = {
    command
      | String,
    arg_type
      | [| 'String, 'Number |],
    alias
      | String
      | optional,
  },
}

Optional fields are intended to be used in record contracts to mean that a field might be defined, but is not mandatory. An optional field initially doesn't have a definition optional-with-value. A value is provided by applying the contract to - or merging it with - a record which defines a value for this field:

{
  command = "exit",
  arg_type = 'String,
  alias = "e",
} | Command

Once an optional field becomes defined, it acts just like a regular field. Any attached contract will be applied as well (here, String on alias). An optional field also becomes a regular field as soon as it is merged with the same field that doesn't have the optional annotation. This holds even if the other field doesn't have a definition:

> {foo = 1, bar | optional} & {bar | optional}
{ foo = 1, }

> {foo = 1, bar | optional} & {bar}
error: missing definition for `bar`
  ┌─ <repl-input-1>:1:121{foo = 1, bar | optional} & {bar}----------^^^--------------------
  │  │         │
  │  │         required here
  │  in this record

In the second example, bar isn't optional on the right-hand side of the merge. Although the right-hand side bar doesn't have a definition, the resulting bar field isn't optional anymore.

As long as an optional field doesn't have a value, it will be invisible to record operations. Optional fields without a value don't show up in std.record.fields, they won't make std.record.values throw a missing field definition error, etc.

> let Contract = {foo = 1, bar | optional}

> std.record.values Contract
[ 1 ]

> std.record.has_field "bar" Contract
false

Optional fields can still be discovered through metadata queries (run nickel help query or type :help query in the REPL for more information) or generated documentation.

field, but this nullifies the effect of optional: what you get is just a regular field

Merge priorities

Priorities are specified using the priority annotation, followed by a number literal. There are also two other special priority annotations, the lowest priority default, and the highest priority force. Both are written without the priority keyword.

Priorities dictate which values take precedence over other values. By default, values are given the priority 0. Values with the same priority are recursively merged as specified in this document, which can mean failure if the values can't be meaningfully combined:

> {foo = 1} & {foo = 2}
error: non mergeable terms
  ┌─ <repl-input-5>:1:91{foo = 1} & {foo = 2}-------^-----------^-
  │  │      │           │
  │  │      │           with this expression
  │  │      cannot merge this expression
  │  originally merged here
  │
  = Merge operands have the same merge priority but they can't be combined.
  = Both values are of type Number but they aren't equal.
  = Number values can only be merged if they are equal

If the priorities differ, the value with the highest priority simply erases the other:

> {foo | priority 1 = 1} & {foo = 2}
{ foo | priority 1 = 1, }

> {foo | priority -1 = 1} & {foo = 2}
{ foo = 2, }

The priorities are ordered in the following way:

  • default is the lowest priority
  • numeral priorities are ordered as usual numbers (priorities can be any valid Nickel number, including fractions and negative values)
  • force is the highest priority

Default values

A default annotation can be used to provide a base value, but have it be overridable through merging. For example, {foo | default = 1} & {foo = 2} evaluates to {foo = 2}. A default value is just a special case of a priority: the lowest possible one.

Forcing values

Dually, values with the force annotation are given the highest priority. Such a value can never be overridden, and will either take precedence over another value or be tentatively merged if the other value has priority force as well.

Specification

Each field definition foo = val is assigned a priority p(val). When merging two common fields value_left and value_right, the result is either the one with the highest priority (which overrides the other), or the two are tentatively recursively merged if the priorities are equal. Without loss of generality, consider the simple case of two records with only one common field:

{common = left} & {common = right}
= {
  common = left          if p(left) > p(right)
           right         if p(left) < p(right)
           left & right  if p(left) = p(right)
}

Example

Let us stick to our firewall example. Thanks to default values, we can set the most restrictive configuration by default, but have it still be overridden if needed.

Let us first try without default values:

let base = {
  firewall.enabled = true,
  firewall.type = "iptables",
  firewall.open_ports = [21, 80, 443],
} in
let patch = {
  firewall.enabled = false,
  server.host.options = "TLS",
} in
base & patch

Because merging is meant to be symmetric, Nickel is unable to know which value to pick between enabled = true and enabled = false for the firewall, and thus it will fail:

error: non mergeable terms
   ┌─ repl-input-8:2:22
   │
 2 │   firewall.enabled = true,
   │                      ^^^^ cannot merge this expression
   ·
 7 │   firewall.enabled = false,
   │                      ^^^^^ with this expression
   ·
10 │ base & patch
   │ ------------ originally merged here
   │
   = Merge operands have the same merge priority but they can't be combined.
   = Both values are of type Bool but they aren't equal.
   = Bool values can only be merged if they are equal

We can use default values to give the priority to the right side:

let base = {
  firewall.enabled | default = true,
  firewall.type | default = "iptables",
  firewall.open_ports | default = [21, 80, 443],
} in
let patch = {
  firewall.enabled = false,
  server.host.options = "TLS",
} in
base & patch

This evaluates to:

{
  firewall = {
    enabled = false,
    open_ports = [21, 80, 443],
    type = "iptables",
  },
  server = {
    host = {
      "options" = "TLS",
    }
  }
}

Contracts

Note: see the correctness section and the contracts section for a thorough introduction to contracts in Nickel.

Fields may have contracts attached, either directly, as in {foo | Number = 1}, or propagated from an annotation higher up, as in {foo = 1} | {foo | Number}. In both cases, foo must satisfy the contract Number. What happens if the value of foo is altered in a subsequent merge? For example:

  • Should {foo | default | Number = 1} & {foo = "bar"} succeed, although foo would be a string in the final result?
  • Should {foo | {subfield | String} = {subfield = "a"}} & {foo.other_subfield = 1} succeed, although a closed contract {subfield | String} is attached to foo, and the final result would have an additional field other_subfield ?

Nickel chooses to answer no to both. In general, when a contract is attached to a field foo, merging ensures that whatever this field is merged with, including being dropped in favor of another value, the final value for foo has to respect the contract as well or the evaluation will fail accordingly.

This is only true for contracts attached directly to record fields (either directly, or coming from an enclosing record contract). In particular, {foo | Number = 1} & {foo | force = "bar"} will fail, but {foo = (1 | Number)} & {foo | force = "bar"} will succeed. In the latter case, the contract is not considered to be field metadata, but a local contract check, which is not propagated by merging.

Specification

Consider two operands with one field each, which is the same on both side, and respective contracts Left1, .., Leftn and Right1, .., Rightk attached

left = {
  common | Left1
         | ..
         | Leftn
}

and

right = {
  common | Right1
         | ..
         | Rightk
}

Then the common field of left & right will be checked against Left1, .., Leftn, Right1, .., Rightk.

The accumulated contracts are applied lazily: as long as the field's value isn't requested, contracts are accumulated but not yet applied. This makes it possible to build a value piecewise, whereas the intermediate values don't necessarily satisfy the contract. For example:

let FooContract = {
  required_field1,
  required_field2,
} in
{ foo | FooContract}
& { foo.required_field1 = "here" }
& { foo.required_field2 = "here" }

Running the above program succeeds, even though the intermediate value {foo.required_field1 = "here"} doesn't respect FooContract (it misses the field required_field2).

If we try to observe the intermediate result (deep_seq recursively forces the evaluation of its first argument and proceeds with evaluating the second argument), we do get a contract violation error:

> let FooContract = {
    required_field1,
    required_field2,
  }
  in
 
  let intermediate =
    { foo | FooContract }
    & { foo.required_field1 = "here" }
  in
  
  intermediate
  & { foo.required_field2 = "here" }
  |> std.deep_seq intermediate
error: missing definition for `required_field2`
     ┌─ <repl-input-8>:3:53 │     required_field2,
     │     ^^^^^^^^^^^^^^^ required here
     ·
   9& { foo.required_field1 = "here" }------------------------ in this record
     │
[...]

Example

We might want to require that only non-privileged port numbers are used in a configuration. This could for example be done as follows:

let Port
  | doc "A valid port number"
  =
    std.contract.from_predicate
      (
        fun value =>
          std.is_number value
          && value % 1 == 0
          && value >= 0
          && value <= 65535
      )
  in

let GreaterThan
  | doc "A number greater than the parameter"
  = fun x => std.contract.from_predicate (fun value => value > x)
  in

{
  port
    | GreaterThan 1024
    | default
    = 8080,
}
& {
  port | Port = 80,
}

Because 80 would be less than 1024, this fails at evaluation:

error: contract broken by the value of `port`
   ┌─ example.ncl:27:17
   │
22 │     | GreaterThan 1024
   │       ---------------- expected type
   ·
27 │   port | Port = 80,
   │                 ^^ applied to this expression

Not exported

A field can be marked as not exported, using the not_exported annotation. Such a field will behave like a regular field during evaluation, but it will be ignored during serialization: such a field won't appear in the output of nickel export nor in the output of std.serialize. It won't be evaluated upon serialization either.

For example, say we want to add some high-level configuration field to a modular configuration, from which other fields are derived:

# hello-service.ncl
{
  greeter
    | String
    | not_exported
    | default
    = "world",

  systemd.services.hello = {
    wantedBy = ["multi-user.target"],
    serviceConfig.ExecStart = "/usr/bin/hello -g'Hello, %{greeter}!'",
  },
}

greeter can then be customized without interfering with the final output:

$ nickel export <<< '(import "hello-service.ncl") & {greeter = "country"}'
{
  "systemd": {
    "services": {
      "hello": {
        "serviceConfig": {
          "ExecStart": "/usr/bin/hello -g'Hello, country!'"
        },
        "wantedBy": [
          "multi-user.target"
        ]
      }
    }
  }
}

Documentation

Documentation is attached via the doc keyword. Merging propagates documentation. Documentation can be retrieved through nickel query, the :query command inside the REPL, or when using the LSP. For example, we can query foo by running nickel query --field foo config.ncl on:

# config.ncl
{
  foo | doc "Some documentation"
      | default = {}
} & {
  foo.field = null,
}
$ nickel query --field foo config.ncl
• documentation: Some documentation

Available fields
• field

If both sides have documentation, the behavior is unspecified, as merging two distinct blobs of text doesn't make sense in general. Currently, Nickel will randomly keep one of the two.

Recursive overriding

We've seen that default values are useful to override a single field with a different value. The combo of merging, recursive records and merge priorities together provides the capability of recursive overriding, which is a powerful tool, and the subject of this section.

In Nickel, records are recursive by default, in order to easily express dependencies between the different fields of a configuration. Concretely, you can refer to other fields of a record from within the same record:

let base_config = {
  version | default = "20.09",
  input.url | default = "nixpkgs/nixos-%{version}",
} in
base_config

Here, version references the input field transparently. This configuration evaluates to:

{
  version = "20.09",
  input = {url = "nixpkgs/nixos-20.09"},
}

In this case, when we override the default value of version, the fields that depend on version -- here, input -- will also be updated automatically. This is what we mean by recursive overriding. For example, base_config & {version = "unstable"} will evaluate to:

{
  version = "unstable",
  input = {url = "nixpkgs/nixos-unstable"},
}

Example

Here is another variation of recursive overriding on the firewall example:

let security = {
  firewall.open_proto.http | default = true,
  firewall.open_proto.https | default = true,
  firewall.open_proto.ftp | default = true,
  firewall.open_ports =
    []
    @ (if firewall.open_proto.ftp then [21] else [])
    @ (if firewall.open_proto.http then [80] else [])
    @ (if firewall.open_proto.https then [443] else []),
}
in # => security.firewall.open_ports = [21, 80, 443]
security & { firewall.open_proto.ftp = false } # => firewall.open_ports = [80, 443]

Here, security.firewall.open_ports is [21, 80, 443]. But in the final configuration, firewall.open_ports will be [80, 443].


  1. You can actually provide a value (default or not) for an optional