Why Sponsor Oil? | source | all docs for version 0.11.0 | all versions | oilshell.org
Hay lets you use the syntax of the Oil shell to declare data and interleaved code. It allows the shell to better serve its role as essential glue. For example, these systems all combine Unix processes in various ways:
Slogans:
This doc describes how to use Hay, with motivating examples.
As of 2022, this is a new feature of Oil, and it needs user feedback. Nothing is set in stone, so you can influence the language and its features!
Hay could be used to configure a hypothetical Linux package manager:
# cpython.hay -- A package definition
hay define Package/TASK # define a tree of Hay node types
Package cpython { # a node with attributes, and children
version = '3.9'
url = 'https://python.org'
TASK build { # a child node, with Oil code
./configure
make
}
}
This program evaluates to a JSON tree, which you can consume from programs in any language, including Oil:
{ "type": "Package",
"args": [ "cpython" ],
"attrs": { "version": "3.9", "url": "https://python.org" },
"children": [
{ "type": "TASK",
"args": [ "build" ],
"code_str": " ./configure\n make\n"
}
]
}
That is, a package manager can use the attributes to create a build environment, then execute shell code within it. This is a staged evaluation model.
A goal of Hay is to restore the simplicity of Unix to distributed systems. It's all just code and data!
This means that it's a bit abstract, so here are a few ways of understanding it.
The relation between Hay and Oil is like the relationship between these pairs of languages:
Each of these is 70's-style macro programming — a stringly-typed language generating another stringly-typed language, with all the associated problems.
In contrast, Hay and Oil are really the same language, with the same syntax, and the same Python- and JavaScript-like dynamic types. Hay is just Oil that builds up data instead of executing commands.
(Counterpoint: Ninja is intended for code generation, and it makes sense for Oil to generate simple languages.)
See the Survey of Config Languages on the wiki, which puts them in these categories:
Excerpts:
YAML is a data format that is (surprisingly) the de-facto control plane language for the cloud. It's an approximate superset of JSON.
UCL (universal config language) and HCL (HashiCorp config language) are influenced by the Nginx config file syntax. If you can read any of these languages, you can read Hay.
Nix has a functional language to configure Linux distros. In contrast, Hay is multi-paradigm and imperative.
The Starlark language is a dialect of Python used by the Bazel build system. It uses imperative code to specify build graph variants, and you can use this same pattern in Hay. That is, if statements, for loops, and functions are useful in Starlark and Hay.
Ruby's use of first-class blocks inspired Oil. They're used in systems like Vagrant (VM dev environments) and Rake (a build system).
In Lisp, code and data are expressed with the same syntax, and can be interleaved. G-Expressions in Guix use a staged evaluation model, like Hay.
The biggest difference between Hay and UCL / HCL is that it's embedded in a shell. In other words, Hay languages are internal DSLs, while those languages are external.
This means:
The sections below elaborate on these points.
Hay nodes have a regular structure:
There are two kinds of node with this structure.
(1) SHELL nodes contain unevaluated code, and their type is ALL CAPS.
The code is turned into a string that can be executed elsewhere.
TASK build {
./configure
make
}
# =>
# ... {"code_str": " ./configure\n make\n"}
(2) Attr nodes contain data, and their type starts with a capital letter.
They eagerly evaluate a block in a new stack frame and turn it into an
attributes dict.
Package cpython {
version = '3.9'
}
# =>
# ... {"attrs": {"version": "3.9"}} ...
These blocks have a special rule to allow bare assignments like version = '3.9'. That is, you don't need keywords like const or var.
(3) In contrast to these two types of Hay nodes, Oil builtins that take a block usually evaluate it eagerly:
cd /tmp { # run in a new directory
echo $PWD
}
Builtins are spelled with lower case letters, so SHELL and Attr nodes
won't be confused with them.
So Hay is designed to be used with a staged evaluation model:
These two stages conceptually different, but use the same syntax and evaluator! Again, the evaluator runs in a mode where it builds up data rather than executing commands.
Here's a description of the result of Hay evaluation (the first stage).
# The source may be "cpython.hay"
FileResult = (source Str, children List[NodeResult])
NodeResult =
# package cpython { version = '3.9' }
Attr (type Str,
args List[Str],
attrs Map[Str, Any],
children List[NodeResult])
# TASK build { ./configure; make }
| Shell(type Str,
args List[Str],
location_str Str,
location_start_line Int,
code_str Str)
Notes:
You can put Hay blocks and normal shell code in the same file. Retrieve the
result of Hay evaluation with the _hay() function.
# myscript.oil
hay define Rule
Rule mylib.o {
inputs = ['mylib.c']
# not recommended, but allowed
echo 'hi'
ls /tmp/$(whoami)
}
echo 'bye' # other shell code
const result = _hay()
json write (result)
In this case, there are no restrictions on the commands you can run.
You can put hay definitions in their own file:
# my-config.hay
Rule mylib.o {
inputs = ['mylib.c']
}
echo 'hi' # allowed for debugging
# ls /tmp/$(whoami) would fail due to restrictions on hay evaluation
In this case, you can use echo and write, but the interpreted is
restricted (see below).
Parse it with parse_hay(), and evaluate it with eval_hay():
# my-evaluator.oil
hay define Rule # node types for the file
const h = parse_hay('build.hay')
const result = eval_hay(h)
json write (result)
# =>
# {
# "children": [
# { "type": "Rule",
# "args": ["mylib.o"],
# "attrs": {"inputs": ["mylib.c"]}
# }
# ]
# }
Instead of creating separate files, you can also use the hay eval builtin:
hay define Rule
hay eval :result { # assign to the variable 'result'
Rule mylib.o {
inputs = ['mylib.c']
}
}
json write (result) # same as above
This is mainly for testing and demos.
The "restrictions" are not a security boundary! (They could be, but we're not making promises now.)
Even with eval_hay() and hay eval, the config file is evaluated in the
same interpreter. But the following restrictions apply:
echo and write aren't allowed
.hay file can't invoke shopt to change global shell
options.hay file can't mutate your locals
setglobal!In summary, Hay evaluation is restricted to prevent basic mistakes, but your code isn't completely separate from the evaluated Hay file.
If you want to evaluate untrusted code, use a separate process, and run it in a container or VM.
Here is a list of all the mechanisms mentioned.
hay
hay define to define node types.hay pp to pretty print the node types.hay reset to delete both the node types and the current evaluation
result.hay eval :result { ... } to evaluate in restricted mode, and put the
result in a variable.haynode builtin is run when types like
Package and TASK are invoked. That is, all node types are aliases for
this same builtin.parse_hay() parses a file, just as bin/oil does.eval_hay() evaluates the parsed file in restricted mode, like hay eval._hay() retrieves the current result
_ because it's a "register" mutated by the
interpreter.Hay is parsed and evaluated with option group oil:all, which includes
parse_proc and parse_equals.
Why would you want to interleave data and code? One reason is to naturally express variants of a configuration. Here are some examples.
Build variants. There are many variants of the Oil binary:
dbg and opt. the compiler optimization level, and whether debug symbols
are included.asan and ubsan. Dynamic analysis with Clang sanitizers.-D GC_EVERY_ALLOC. Make a build that helps debug the garbage collector.So the Ninja build graph to produce these binaries is shaped similarly, but it varies with compiler and linker flags.
Service variants. A common problem in distributed systems is how to develop and debug services locally.
Do your service dependencies live in the cloud, or are they run locally? What about state? Common variants:
local. Part or all of the service runs locally, so you may pass flags like
--auth-service localhost:8001 to binaries.staging. A complete copy of the service, in a different cloud, with a
different database.prod. The live instance running with user data.Again, these collections of services are all shaped similarly, but the flags vary based on where binaries are physically running.
This model can be referred to as "graph metaprogramming" or "staged programming". In Oil, it's done with dynamically typed data like integers and dictionaries. In contrast, systems like CMake and autotools are more stringly typed.
The following examples are meant to be "evocative"; they're not based on real code. Again, user feedback can improve them!
Conditionals can go on the inside of a block:
Service auth.example.com { # node taking a block
if (variant === 'local') { # condition
port = 8001
} else {
port = 80
}
}
Or on the outside:
Service web { # node
root = '/home/www'
}
if (variant === 'local') { # condition
Service auth-local { # node
port = 8001
}
}
Iteration can also go on the inside of a block:
Rule foo.o { # node
inputs = [] # populate with all .cc files except one
# variables ending with _ are "hidden" from block evaluation
for name_ in *.cc {
if name_ !== 'skipped.cc' {
_ append(inputs, name_)
}
}
}
Or on the outside:
for name_ in *.cc { # loop
Rule $(basename $name_ .cc).o { # node
inputs = [name_]
}
}
procProcs can wrap blocks:
proc myrule(name) {
# needed for blocks to use variables higher on the stack
shopt --set dynamic_scope {
Rule dbg/$name.o { # node
inputs = ["$name.c"]
flags = ['-O0']
}
Rule opt/$name.o { # node
inputs = ["$name.c"]
flags = ['-O2']
}
}
}
myrule foo # call proc
myrule bar # call proc
Or they can be invoked from within blocks:
proc set-port(port_num, :out) {
setref out = "localhost:$port_num"
}
Service foo { # node
set-port 80 :p1 # call proc
set-port 81 :p2 # call proc
}
The general pattern is:
./my-evaluator.oil my-config.hay | json read :result
The evaluator does the following:
hay definemy-config.hay with parse_hay()eval_hay()Then a separate Oil processes reads this JSON and executes application code.
TODO: Show code example.
In Python, you would:
subprocess module to invoke ./my-evaluator.oil my-config.hay.json module to parse the result.TODO: Show code example.
.d DirsDebian has a pattern of splitting configuration into a directory of concatenated files. It's easier for shell scripts to add to a directory than add to a file.
This can be done with an evaluator that simply enumerates all files:
var results = []
for path in myconfig.d/*.hay {
const code = parse_hay(path)
const result = eval(hay)
_ append(results, result)
}
# Now iterate through results
TODO: Example of using xargs -P to spawn processes with parse_hay() and
eval_hay(). Then merge the JSON results.
Assigning attributes and invoking procs can look similar:
Package grep {
version = '1.0' # An attribute?
version 1.0 # or call proc 'version'?
}
The first style is better for typed data like integers and dictionaries. The
latter style isn't useful here, but it could be if version 1.0 created
complex Hay nodes.
Hay nodes shouldn't take flags or --. Flags are for key-value pairs, and
blocks are better for expressing such data.
No:
Package --version 1.0 grep {
license = 'GPL'
}
Yes:
Package grep {
version = '1.0'
license = 'GPL'
}
Superficially, dicts and blocks are similar:
Package grep {
mydict = {name: 'value'} # a dict
mynode foo { # a node taking a block
name = 'value'
}
}
Use dicts in cases where you don't know the names or types up front, like
files = {'README.md': true, '__init__.py': false}
Use blocks when there's a schema. Blocks are also different because:
if statements and for loops in them.TASK build; TASK test within a block, creating multiple
objects of the same type.Hay files are parsed as Oil, not OSH. That includes SHELL nodes:
TASK build {
cp @deps /tmp # Oil splicing syntax
}
If you want to use POSIX shell or bash, use two arguments, the second of which is a multi-line string:
TASK build '''
cp "${deps[@]}" /tmp
'''
The Oil style gives you static parsing, which catches some errors earlier.
hay proc for arbitrary schema validation, including JSON schemaPlease send feedback about Hay. It will inform and prioritize this work!