crdt post
All checks were successful
ci/woodpecker/push/deploy Pipeline was successful

This commit is contained in:
Michael Zhang 2024-12-17 16:38:38 -06:00
parent 9c8c3588c1
commit 6e82102c1b
3 changed files with 299 additions and 2 deletions

3
.vscode/settings.json vendored Normal file
View file

@ -0,0 +1,3 @@
{
"npm.packageManager": "pnpm"
}

View file

@ -0,0 +1,294 @@
---
title: CRDTs in Agda
slug: 2024-12-16-crdt
date: 2024-12-16
tags: ["agda", "formalization"]
---
Today during my flight back from Boston, I did a small formalization of [conflict-free replicated data types, or CRDTs][crdt] (thanks Sun Country for not having wifi...).
CRDTs are prevalent in the data synchronization world, and enable each peer to track their own changes, then merge those changes without conflict.
One of the nice guarantees of this system (and also a requirement of the data structures!) is that changes can be applied in any order, and they must eventually become consistent.
I'll present a way of modeling this using Agda, along with two simple example formalizations of data structures.
[crdt]: https://en.wikipedia.org/wiki/Conflict-free_replicated_data_type
> [!admonition: NOTE]
> As usual, this is a [Literate agda][lagda] file.
> This means it is real code that type-checks.
> Click on any element to navigate to its definition!
[lagda]: https://agda.readthedocs.io/en/latest/tools/literate-programming.html
<details>
<summary>Imports</summary>
```
module 2024-12-16-crdt where
open import Agda.Primitive
open import Data.List using (List; _∷_; [])
open import Data.Nat using (; suc)
open import Data.Product
open import Data.Bool using (true; false)
open import Data.Unit using ()
open import Data.Empty
open import Function
open import Relation.Binary.Bundles
open import Relation.Binary.Definitions
open import Relation.Binary.Consequences
open import Relation.Binary.Structures
import Relation.Binary.PropositionalEquality as Eq
open import Relation.Binary.PropositionalEquality hiding (isEquivalence; isDecEquivalence)
open Eq.≡-Reasoning
open import Relation.Nullary.Decidable
open import Relation.Nullary.Negation
open import Relation.Nullary.Reflects
```
</details>
## Definition
For this post, I'm doing an operation-based CRDT.
Since the C in CRDT stands for conflict-free:
- The `apply` function is commutative.
We should be free to merge changes in any order.
This also guarantees eventual consistency (as long as all changes are replicated properly), since no matter what the changes are going to be, as long as they can commute arbitrarily, we can just re-order them.
- The `apply` function is total.
We can't have any errors being thrown.
All states must merge cleanly.
(NOTE: This is a protocol-level concept. We can still have conflicts in applications, they would just have to appear as a valid state in the CRDT.)
Let's define our data structure:
<details>
<summary>Some boilerplate for abstracting over time</summary>
```
-- ignore the levels, they are there so `Time` can be abstracted over
module _ { c time time₁ time₂ : Level} (TimeOrder : StrictTotalOrder time time₁ time₂) where
open StrictTotalOrder (TimeOrder)
using (compare; _<_; _≈_; _>_; asym; irrefl; isDecEquivalence)
renaming (Carrier to Time)
open IsDecEquivalence (isDecEquivalence) using () renaming (sym to symEqv)
TCompare : Time → Time → Set (time₁ ⊔ time₂)
TCompare t1 t2 = Tri (t1 < t2) (t1 t2) (t1 > t2)
```
</details>
```
-- This data structure represents a proof that the type A is a CRDT
record isCRDT {c : Level} (A : Set c) : Set (time ⊔ c ⊔ lsuc ) where
field
-- Some type representing all possible operations
op : Set
-- The no-op. Having this lets me only require apply2
noOp : op
-- Apply 2 operations to a state, along with their timestamps
-- Note that there is no error state for this.
-- All operations must be cleanly applied.
apply2 : (o1 o2 : op) (t1 t2 : Time) → A → A
-- The property that applying changes must be commutative.
comm : (a : A) → (o1 o2 : op) (t1 t2 : Time) → apply2 o1 o2 t1 t2 a ≡ apply2 o2 o1 t2 t1 a
```
I think in reality it would be the job of `op` to store the time, rather than the entire `isCRDT`, but this is the way I wrote it :sob:
Another note is, I required `apply2` rather than just `apply` taking a single operation, because I have to express the fact that the data structure can choose to reorder my operations however it wants.
## Example: counter
Here is a simple example of a counter. The counter's only operation is _increment_, and the merging operation is trivial.
```
module _ where
open isCRDT
-- A counter is represented by just a natural number
Counter =
private
data Op : Set where
nop : Op
inc : Op
CounterIsCRDT : isCRDT Counter
CounterIsCRDT .op = Op
CounterIsCRDT .noOp = nop
```
> [!admonition: NOTE]
> Because the rest of the code all lies inside this module, future code blocks are going to start with more indentation :(
Now we need to define apply and commutativity:
```
CounterIsCRDT .apply2 = apply2' where
apply2' : Op → Op → Time → Time → Counter → Counter
apply2' nop nop t1 t2 c = c -- no change
apply2' nop inc t1 t2 c = suc c -- just add one
apply2' inc nop t1 t2 c = suc c -- just add one
apply2' inc inc t1 t2 c = suc (suc c) -- two inc's adds two
CounterIsCRDT .comm = comm' where
apply2' = CounterIsCRDT .apply2
-- All cases are trivial, since the operations are symmetric by definition
comm' : (a : Counter) (o1 o2 : Op) (t1 t2 : Time)
→ apply2' o1 o2 t1 t2 a ≡ apply2' o2 o1 t2 t1 a
comm' c nop nop t1 t2 = refl
comm' c nop inc t1 t2 = refl
comm' c inc nop t1 t2 = refl
comm' c inc inc t1 t2 = refl
```
## Example: last-write-win map
Here's a slightly more complicated example: last-write-win maps.
Instead of using a built-in map module, I'm going to define some properties that a map should have.
Importantly, I'll require that there be some way of breaking ties between equal keys.
<details>
<summary>Boilerplate for setting up maps</summary>
```
record isMap (M K V : Set c) : Set (c ⊔ time) where
constructor mkIsMap
field
mget : M → K → V
mext : (m1 m2 : M) → ((k : K) → mget m1 k ≡ mget m2 k) → m1 ≡ m2
minsert : (k : K) (v : V) (m : M) → Σ M (λ m' → mget m' k ≡ v)
-- These functions are used to arbitrarily order two choices in case the time _and_ keys collide
-- It must be deterministic
mpick : (kvt1 kvt2 : K × V × Time) → (K × V × Time) × (K × V × Time)
mpickcomm : (kvt1 kvt2 : K × V × Time) → mpick kvt1 kvt2 ≡ mpick kvt2 kvt1
MapType : (K V : Set c) → Set (lsuc c ⊔ time)
MapType K V = Σ (Set c) (λ M → isMap M K V)
module _ (K V : Set c) (Map : MapType K V) where
open isCRDT
M = Map .proj₁
open isMap (Map .proj₂)
```
</details>
It's time to define operations and apply.
```
private
data Op : Set c where
nop : Op
insert : K → V → Op
-- "deterministic pair" is how it's used, but this is really just two (K, V) pairs
record DetPair (K V : Set c) : Set c where
field
k₁ k₂ : K
v₁ v₂ : V
module _ (k1 : K) (v1 : V) (t1 : Time) (k2 : K) (v2 : V) (t2 : Time) where
-- This helper takes the time comparison result and decides how to order the
-- resulting pair
mkDetPairHelper : TCompare t1 t2 → DetPair K V
mkDetPairHelper (tri< a ¬b ¬c) =
record { k₁ = k1 ; k₂ = k2 ; v₁ = v1 ; v₂ = v2 }
mkDetPairHelper (tri≈ ¬a b ¬c) =
let arb = mpick (k1 , v1 , t1) (k2 , v2 , t2) in
record
{ k₁ = arb .proj₁ .proj₁
; v₁ = arb .proj₁ .proj₂ .proj₁
; k₂ = arb .proj₂ .proj₁
; v₂ = arb .proj₂ .proj₂ .proj₁
}
mkDetPairHelper (tri> ¬a ¬b c) =
record { k₁ = k2 ; k₂ = k1 ; v₁ = v2 ; v₂ = v1 }
mkDetPair : DetPair K V
mkDetPair = mkDetPairHelper (compare t1 t2)
-- Define single apply
apply : Op → M → M
apply nop m = m
apply (insert k v) m = minsert k v m .proj₁
apply2' : Op → Op → Time → Time → M → M
apply2' nop nop t1 t2 m = m
apply2' o1 nop t1 t2 m = apply o1 m -- single apply is used
apply2' nop o2 t1 t2 m = apply o2 m -- in the NOP cases
apply2' (insert k1 v1) (insert k2 v2) t1 t2 m =
let p = mkDetPair k1 v1 t1 k2 v2 t2 in
apply (insert (p .k₂) (p .v₂)) (apply (insert (p .k₁) (p .v₁)) m)
where open DetPair
```
Commutativity is now just a matter of going back and verifying that those cases hold.
For the case where I `mpick`ed an arbitrary order, I used the required `mpickcomm` property to show that it commutes.
```
comm' : (m : M) (o1 o2 : Op) (t1 t2 : Time) → apply2' o1 o2 t1 t2 m ≡ apply2' o2 o1 t2 t1 m
comm' m nop nop t1 t2 = refl
comm' m (insert _ _) nop t1 t2 = refl
comm' m nop (insert _ _) t1 t2 = refl
comm' m (insert k1 v1) (insert k2 v2) t1 t2 = cong f (detPairLemma k1 k2 v1 v2 t1 t2) where
f : (dp : DetPair K V) → M
f dp = apply (insert (dp .k₂) (dp .v₂)) (apply (insert (dp .k₁) (dp .v₁)) m)
where open DetPair
module _ (k1 k2 : K) (v1 v2 : V) (t1 t2 : Time) where
detPairLemmaHelper : (tc1 : TCompare t1 t2) (tc2 : TCompare t2 t1) → mkDetPairHelper k1 v1 t1 k2 v2 t2 tc1 ≡ mkDetPairHelper k2 v2 t2 k1 v1 t1 tc2
-- Uninteresting cases.
-- Let me know if you know of a way to avoid having to list these cases!
detPairLemmaHelper (tri< a ¬b ¬c) (tri< a ¬b ¬c) = -elim (asym a a)
detPairLemmaHelper (tri< a ¬b ¬c) (tri ¬a b ¬c) = -elim (irrefl (symEqv b) a)
detPairLemmaHelper (tri≈ ¬a b ¬c) (tri< a ¬b ¬c) = -elim (irrefl (symEqv b) a)
detPairLemmaHelper (tri≈ ¬a b ¬c) (tri> ¬a₁ ¬b c) = ⊥-elim (irrefl b c)
detPairLemmaHelper (tri> ¬a ¬b c) (tri≈ ¬a₁ b ¬c) = ⊥-elim (irrefl b c)
detPairLemmaHelper (tri> ¬a ¬b c) (tri> ¬a₁ ¬b₁ c₁) = ⊥-elim (asym c c₁)
-- Interesting cases.
detPairLemmaHelper (tri< a ¬b ¬c) (tri> ¬a ¬b₁ c) = refl
detPairLemmaHelper (tri≈ ¬a b ¬c) (tri≈ ¬a₁ b₁ ¬c₁) = cong pair pickComm where
kvt1 = k1 , v1 , t1
kvt2 = k2 , v2 , t2
pair : (K × V × Time) × (K × V × Time) → DetPair K V
pair ((k1 , v1 , t1) , (k2 , v2 , t2)) = record { k₁ = k1 ; k₂ = k2 ; v₁ = v1 ; v₂ = v2 }
pickComm : mpick kvt1 kvt2 ≡ mpick kvt2 kvt1
pickComm = mpickcomm kvt1 kvt2
detPairLemmaHelper (tri> ¬a ¬b c) (tri< a ¬b ¬c) = refl
detPairLemma : mkDetPair k1 v1 t1 k2 v2 t2 ≡ mkDetPair k2 v2 t2 k1 v1 t1
detPairLemma = detPairLemmaHelper (compare t1 t2) (compare t2 t1)
```
Finally, to wrap up:
```
LWWMapIsCRDT : isCRDT M
LWWMapIsCRDT .op = Op
LWWMapIsCRDT .noOp = nop
LWWMapIsCRDT .apply2 = apply2'
LWWMapIsCRDT .comm = comm'
```
## Conclusion
If I were to use this in a real product, I could write `{-# COMPILE ... #-}` and give
translations of all the constructs into Javascript, and then import this as a library.
Anyway, this is just a toy implementation.
Let me know how this could be improved!
As an aside, if you're interested in theorem proving (Lean/Agda), make sure to check out [Advent of Proof 2024][AOP]!
There have been plenty of cool problems so far.
[aop]: https://typesig.pl/lean/agda/events/2024/12/12/advent-of-proof-2024.html

View file

@ -4,7 +4,7 @@
--blue: #5175be;
--module: purple;
--inductive-constructor: #008b00;
--comment: #b22222;
--comment: #a7a7a7;
--string: var(--comment);
}
@ -15,7 +15,7 @@
--blue: #9999FF;
--module: hsl(276, 66%, 69%);
--inductive-constructor: #32bb32;
--comment: hsl(0, 78%, 66%);
--comment: hsl(0, 0%, 47%);
}
}