Skip to content

Commit 953567d

Browse files
authored
Document config system (#935)
1 parent 29e920d commit 953567d

File tree

3 files changed

+174
-1
lines changed

3 files changed

+174
-1
lines changed

crates/stackable-operator/src/config/fragment.rs

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -109,7 +109,7 @@ pub trait FromFragment: Sized {
109109
/// For complex structs, this should be a variant of `Self` where each field is replaced by its respective `Fragment` type. This can be derived using
110110
/// [`Fragment`].
111111
type Fragment;
112-
/// A variant of [`Self::Fragment`] that is used when the container already provides a to indicate that a value is optional.
112+
/// A variant of [`Self::Fragment`] that is used when the container already provides a way to indicate that a value is optional.
113113
///
114114
/// For example, there's no use marking a value as [`Option`]al again if the value is already contained in an `Option`.
115115
///

crates/stackable-operator/src/config/merge.rs

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,3 +1,5 @@
1+
//! Automatically merges objects *deeply*, especially fragments.
2+
13
use k8s_openapi::{
24
api::core::v1::{NodeAffinity, PodAffinity, PodAntiAffinity, PodTemplateSpec},
35
apimachinery::pkg::{api::resource::Quantity, apis::meta::v1::LabelSelector},
Lines changed: 171 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -1,2 +1,173 @@
1+
//! The Stacklet Configuration System™©®️️️ (SCS).
2+
//!
3+
//! # But oh god why is this monstrosity a thing?
4+
//!
5+
//! Products are complicated. They need to be supplied many kinds of configuration.
6+
//! Some of it applies to the whole installation (Stacklet). Some of it applies only to one [role](`Role`).
7+
//! Some of it applies only to a subset of the instances of that role (we call this a [`RoleGroup`]).
8+
//!
9+
//! We (usually) don't know at what level it makes sense to apply a given piece of configuration, but we also
10+
//! don't want to force users to repeat themselves constantly! Instead, we model the configuration as a tree:
11+
//!
12+
//! ```yaml
13+
//! stacklet1:
14+
//! role1:
15+
//! group1:
16+
//! group2:
17+
//! role2:
18+
//! group3:
19+
//! group4:
20+
//! stacklet2:
21+
//! role3:
22+
//! group5:
23+
//! ```
24+
//!
25+
//! Where only the leaves (*groups*) are actually realized into running products, but every level inherits
26+
//! the configuration of its parents. So `group1` would inherit any keys from `role1` (and, transitively, `stacklet1`),
27+
//! unless it overrides them.
28+
//!
29+
//! We also want to *validate* that the configuration actually makes sense, but only once we have the fully realized
30+
//! configuration for a given rolegroup.
31+
//!
32+
//! However, in practice, living in a fully typed land like Rust makes this slightly awkward. We end up having to choose from
33+
//! a few awkward options:
34+
//!
35+
//! 1. Give up on type safety until we're done merging - Type safety is nice, and we still need to produce a schema for
36+
//! Kubernetes to validate against.
37+
//! 2. Give on distinguishing between pre- and post-validation types - Type safety is nice, and it gets error-prone having to memorize
38+
//! which [`Option::unwrap`]s are completely benign, and which are going to bring down the whole cluster. And, uh, good luck trying
39+
//! to *change* that in either direction.
40+
//! 3. Write *separate* types for the pre- and post-validation states - That's a lot of tedious code to have to write twice, and that's not
41+
//! even counting the validation ([parsing]) and inheritance routines! That's not really stuff you want to get wrong!
42+
//!
43+
//! So far, none of those options look particularly great. 3 would probably be the least unworkable path, but...
44+
//! But then again, uh, we have a compiler. What if we could just make it do the hard work?
45+
//!
46+
//! # Okay, but how does it work?
47+
//!
48+
//! The SCS™©®️️️ is split into two subsystems: [`fragment`] and [`merge`].
49+
//!
50+
//! ## Uhhhh, fragments?
51+
//!
52+
//! The [`Fragment`] macro implements option 3 from above for you. You define the final validated type,
53+
//! and it generates a "Fragment mirror type", where all fields are replaced by [`Option`]al counterparts.
54+
//!
55+
//! For example,
56+
//!
57+
//! ```
58+
//! # use stackable_operator::config::fragment::Fragment;
59+
//! #[derive(Fragment)]
60+
//! struct Foo {
61+
//! bar: String,
62+
//! baz: u8,
63+
//! }
64+
//! ```
65+
//!
66+
//! generates this:
67+
//!
68+
//! ```
69+
//! struct FooFragment {
70+
//! bar: Option<String>,
71+
//! baz: Option<u8>,
72+
//! }
73+
//! ```
74+
//!
75+
//! Additionally, it provides the [`validate`] function, which lets you turn your `FooFragment` back into a `Foo`
76+
//! (while also making sure that the contents actually make sense).
77+
//!
78+
//! Fragments can also be *nested*, as long as the whole hierarchy has fragments. In this case, the fragment of the substruct will be used,
79+
//! instead of wrapping it in an Option. For example, this:
80+
//!
81+
//! ```
82+
//! # use stackable_operator::config::fragment::Fragment;
83+
//! #[derive(Fragment)]
84+
//! struct Foo {
85+
//! bar: Bar,
86+
//! }
87+
//!
88+
//! #[derive(Fragment)]
89+
//! struct Bar {
90+
//! baz: String,
91+
//! }
92+
//! ```
93+
//!
94+
//! generates this:
95+
//!
96+
//! ```
97+
//! struct FooFragment {
98+
//! bar: BarFragment,
99+
//! }
100+
//!
101+
//! struct BarFragment {
102+
//! baz: Option<String>,
103+
//! }
104+
//! ```
105+
//!
106+
//! rather than wrapping `Bar` as an option, like this:
107+
//!
108+
//! ```
109+
//! struct FooFragment {
110+
//! bar: Option<Bar>,
111+
//! }
112+
//!
113+
//! struct Bar {
114+
//! baz: String,
115+
//! }
116+
//! // BarFragment would be irrelevant here
117+
//! ```
118+
//!
119+
//! ### How does it actually know whether to use a subfragment or an [`Option`]?
120+
//!
121+
//! That's (kind of) a trick question! [`Fragment`] actually has no idea about what an [`Option`] even is!
122+
//! It always uses [`FromFragment::Fragment`]. A type can opt into the [`Option`] treatment by implementing
123+
//! [`Atomic`], which is a marker trait for leaf types that cannot be merged any further.
124+
//!
125+
//! ### And what about defaults? That seems like a pretty big oversight.
126+
//!
127+
//! The Fragment system doesn't natively support default values! Instead, this comes "for free" with the merge system (below).
128+
//! One benefit of this is that the same `Fragment` type can support different default values in different contexts
129+
//! (for example: different defaults in different rolegroups).
130+
//!
131+
//! ### Can I customize my `Fragment` types?
132+
//!
133+
//! Attributes can be applied to the generated types using the `#[fragment_attrs]` attribute. For example,
134+
//! `#[fragment_attrs(derive(Default))]` applies `#[derive(Default)]` to the `Fragment` type.
135+
//!
136+
//! ## And what about merging? So far, those fragments seem pretty useless...
137+
//!
138+
//! This is where the [`Merge`] macro (and trait) comes in! It is designed to be applied to the `Fragment` types (see above),
139+
//! and merges their contents field-by-field, deeply (as in: [`merge`] will recurse into substructs, and merge *their* keys in turn).
140+
//!
141+
//! Just like for `Fragment`s, types can opt out of being merged using the [`Atomic`] trait. This is useful both for "primitive" values
142+
//! (like [`String`], the recursion needs to end *somewhere*, after all), and for values that don't really make sense to merge
143+
//! (like a set of search query parameters).
144+
//!
145+
//! # Fine, how do I actually use it, then?
146+
//!
147+
//! For declarations (in CRDs):
148+
//! - Apply `#[derive(Fragment)] #[fragment_attrs(derive(Merge))]` for your product configuration (and any of its nested types).
149+
//! - DON'T: `#[derive(Fragment, Merge)]`
150+
//! - Pretty much always derive deserialization and defaulting on the `Fragment`, not the validated type:
151+
//! - DO: `#[fragment_attrs(derive(Serialize, Deserialize, Default, JsonSchema))]`
152+
//! - DON'T: `#[derive(Fragment, Serialize, Deserialize, Default, JsonSchema)]`
153+
//! - Refer to the `Fragment` type in CRDs, not the validated type.
154+
//! - Implementing [`Atomic`] if something doesn't make sense to merge.
155+
//! - Define the "validated form" of your configuration: only make fields [`Option`]al if [`None`] is actually a legal value.
156+
//!
157+
//! For runtime code:
158+
//! - Validate and merge with [`RoleGroup::validate_config`] for CRDs, otherwise [`merge`] manually and then validate with [`validate`].
159+
//! - Validate as soon as possible, user code should never read the contents of `Fragment`s.
160+
//! - Defaults are just another layer to be [`merge`]d.
161+
//!
162+
//! [parsing]: https://lexi-lambda.github.io/blog/2019/11/05/parse-don-t-validate/
163+
//! [`merge`]: Merge::merge
164+
1165
pub mod fragment;
2166
pub mod merge;
167+
168+
#[cfg(doc)]
169+
use crate::role_utils::{Role, RoleGroup};
170+
#[cfg(doc)]
171+
use fragment::{validate, Fragment, FromFragment};
172+
#[cfg(doc)]
173+
use merge::{Atomic, Merge};

0 commit comments

Comments
 (0)