From 8722e27600ddc3ba869cebc388214f3fabb0ca57 Mon Sep 17 00:00:00 2001 From: Quentin Date: Mon, 23 Nov 2020 17:49:21 +0100 Subject: [PATCH 1/3] CRDT doc --- src/model/bucket_table.rs | 3 +++ src/table/crdt.rs | 32 ++++++++++++++++++++++++++++---- 2 files changed, 31 insertions(+), 4 deletions(-) diff --git a/src/model/bucket_table.rs b/src/model/bucket_table.rs index b7f24d71..11f103ff 100644 --- a/src/model/bucket_table.rs +++ b/src/model/bucket_table.rs @@ -8,6 +8,9 @@ use garage_util::error::Error; use crate::key_table::PermissionSet; +// We import the same file but in its version 0.1.0. +// We can then access v0.1.0 data structures. +// We use them to perform migrations. use model010::bucket_table as prev; #[derive(PartialEq, Clone, Debug, Serialize, Deserialize)] diff --git a/src/table/crdt.rs b/src/table/crdt.rs index 2b903cf0..3e5e8317 100644 --- a/src/table/crdt.rs +++ b/src/table/crdt.rs @@ -2,7 +2,26 @@ use serde::{Deserialize, Serialize}; use garage_util::data::*; +/// Conflict-free replicated data type (CRDT) +/// +/// CRDT are a type of data structures that do not require coordination. +/// In other words, we can edit them in parallel, we will always +/// find a way to merge it. +/// +/// A general example is a counter. Its initial value is 0. +/// Alice and Bob get a copy of the counter. +/// Alice does +1 on her copy, she reads 1. +/// Bob does +3 on his copy, he reads 3. +/// Now, it is easy to merge their counters, order does not count: +/// we always get 4. +/// +/// Learn more about CRDT [on Wikipedia](https://en.wikipedia.org/wiki/Conflict-free_replicated_data_type) pub trait CRDT { + /// Merge the two datastructures according to the CRDT rules + /// + /// # Arguments + /// + /// * `other` - the other copy of the CRDT fn merge(&mut self, other: &Self); } @@ -19,6 +38,9 @@ where // ---- LWW Register ---- +/// Last Write Win (LWW) +/// +/// LWW is a very simple #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)] pub struct LWW { ts: u64, @@ -64,8 +86,9 @@ where } } -// ---- Boolean (true as absorbing state) ---- - +/// Boolean +/// +/// with True as absorbing state #[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq)] pub struct Bool(bool); @@ -87,8 +110,9 @@ impl CRDT for Bool { } } -// ---- LWW Map ---- - +/// Last Write Win Map +/// +/// #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)] pub struct LWWMap { vals: Vec<(K, u64, V)>, From cbd10c1b0a1325fbd7cf91a43ea0044ba1aae409 Mon Sep 17 00:00:00 2001 From: Quentin Date: Mon, 23 Nov 2020 18:17:48 +0100 Subject: [PATCH 2/3] Add some doc on LWW --- src/table/crdt.rs | 31 ++++++++++++++++++++++++++++++- 1 file changed, 30 insertions(+), 1 deletion(-) diff --git a/src/table/crdt.rs b/src/table/crdt.rs index 3e5e8317..b0323a66 100644 --- a/src/table/crdt.rs +++ b/src/table/crdt.rs @@ -40,7 +40,22 @@ where /// Last Write Win (LWW) /// -/// LWW is a very simple +/// LWW is based on time, the most recent write wins. +/// As multiple computers clocks are always desynchronized, +/// when operations are close enough, it is equivalent to +/// take one copy and drop the other one. +/// +/// Given that clocks are not too desynchronized, this assumption +/// is enough for most cases, as there is few chance that two humans +/// coordonate themself faster than the time difference between two NTP servers. +/// +/// As a more concret example, let's suppose you want to upload a file +/// with the same key (path) in the same bucket at the very same time. +/// For each request, the file will be timestamped by the receiving server +/// and may differ from what you observed with your atomic clock! +/// +/// This scheme is used by AWS S3 or Soundcloud and often without knowing +/// in entreprise when reconciliating databases with ad-hoc scripts. #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)] pub struct LWW { ts: u64, @@ -51,22 +66,36 @@ impl LWW where T: CRDT, { + /// Creates a new CRDT + /// + /// CRDT's internal timestamp is set with current node's clock. pub fn new(value: T) -> Self { Self { ts: now_msec(), v: value, } } + + /// Build a new CRDT from a previous non-compatible one + /// + /// Compared to new, the CRDT's timestamp is not set to now + /// but must be set to the previous, non-compatible, CRDT's timestamp. pub fn migrate_from_raw(ts: u64, value: T) -> Self { Self { ts, v: value } } + + /// Update the LWW CRDT while keeping some causal ordering. pub fn update(&mut self, new_value: T) { self.ts = std::cmp::max(self.ts + 1, now_msec()); self.v = new_value; } + + /// Get the CRDT value pub fn get(&self) -> &T { &self.v } + + /// Get a mutable value for the CRDT pub fn get_mut(&mut self) -> &mut T { &mut self.v } From 5c6c067b0ce4b9042ccfcdda5d4d31e208f00a27 Mon Sep 17 00:00:00 2001 From: Alex Auvolat Date: Sat, 12 Dec 2020 17:06:40 +0100 Subject: [PATCH 3/3] More documentation on CRDTs (we should probably extract this to a standalone crate!) --- src/table/crdt.rs | 164 ++++++++++++++++++++++++++++++++++++++-------- 1 file changed, 138 insertions(+), 26 deletions(-) diff --git a/src/table/crdt.rs b/src/table/crdt.rs index b0323a66..386e478b 100644 --- a/src/table/crdt.rs +++ b/src/table/crdt.rs @@ -1,30 +1,48 @@ +//! This package provides a simple implementation of conflict-free replicated data types (CRDTs) +//! +//! CRDTs are a type of data structures that do not require coordination. In other words, we can +//! edit them in parallel, we will always find a way to merge it. +//! +//! A general example is a counter. Its initial value is 0. Alice and Bob get a copy of the +//! counter. Alice does +1 on her copy, she reads 1. Bob does +3 on his copy, he reads 3. Now, +//! it is easy to merge their counters, order does not count: we always get 4. +//! +//! Learn more about CRDT [on Wikipedia](https://en.wikipedia.org/wiki/Conflict-free_replicated_data_type) + use serde::{Deserialize, Serialize}; use garage_util::data::*; -/// Conflict-free replicated data type (CRDT) +/// Definition of a CRDT - all CRDT Rust types implement this. /// -/// CRDT are a type of data structures that do not require coordination. -/// In other words, we can edit them in parallel, we will always -/// find a way to merge it. +/// A CRDT is defined as a merge operator that respects a certain set of axioms. /// -/// A general example is a counter. Its initial value is 0. -/// Alice and Bob get a copy of the counter. -/// Alice does +1 on her copy, she reads 1. -/// Bob does +3 on his copy, he reads 3. -/// Now, it is easy to merge their counters, order does not count: -/// we always get 4. +/// In particular, the merge operator must be commutative, associative, +/// idempotent, and monotonic. +/// In other words, if `a`, `b` and `c` are CRDTs, and `⊔` denotes the merge operator, +/// the following axioms must apply: /// -/// Learn more about CRDT [on Wikipedia](https://en.wikipedia.org/wiki/Conflict-free_replicated_data_type) +/// ```text +/// a ⊔ b = b ⊔ a (commutativity) +/// (a ⊔ b) ⊔ c = a ⊔ (b ⊔ c) (associativity) +/// (a ⊔ b) ⊔ b = a ⊔ b (idempotence) +/// ``` +/// +/// Moreover, the relationship `≥` defined by `a ≥ b ⇔ ∃c. a = b ⊔ c` must be a partial order. +/// This implies a few properties such as: if `a ⊔ b ≠ a`, then there is no `c` such that `(a ⊔ b) ⊔ c = a`, +/// as this would imply a cycle in the partial order. pub trait CRDT { - /// Merge the two datastructures according to the CRDT rules + /// Merge the two datastructures according to the CRDT rules. + /// `self` is modified to contain the merged CRDT value. `other` is not modified. /// /// # Arguments /// - /// * `other` - the other copy of the CRDT + /// * `other` - the other CRDT we wish to merge with fn merge(&mut self, other: &Self); } +/// All types that implement `Ord` (a total order) also implement a trivial CRDT +/// defined by the merge rule: `a ⊔ b = max(a, b)`. impl CRDT for T where T: Ord + Clone, @@ -40,11 +58,24 @@ where /// Last Write Win (LWW) /// -/// LWW is based on time, the most recent write wins. +/// An LWW CRDT associates a timestamp with a value, in order to implement a +/// time-based reconciliation rule: the most recent write wins. +/// For completeness, the LWW reconciliation rule must also be defined for two LWW CRDTs +/// with the same timestamp but different values. +/// +/// In our case, we add the constraint that the value that is wrapped inside the LWW CRDT must +/// itself be a CRDT: in the case when the timestamp does not allow us to decide on which value to +/// keep, the merge rule of the inner CRDT is applied on the wrapped values. (Note that all types +/// that implement the `Ord` trait get a default CRDT implemetnation that keeps the maximum value. +/// This enables us to use LWW directly with primitive data types such as numbers or strings. It is +/// generally desirable in this case to never explicitly produce LWW values with the same timestamp +/// but different inner values, as the rule to keep the maximum value isn't generally the desired +/// semantics.) +/// /// As multiple computers clocks are always desynchronized, /// when operations are close enough, it is equivalent to /// take one copy and drop the other one. -/// +/// /// Given that clocks are not too desynchronized, this assumption /// is enough for most cases, as there is few chance that two humans /// coordonate themself faster than the time difference between two NTP servers. @@ -85,6 +116,12 @@ where } /// Update the LWW CRDT while keeping some causal ordering. + /// + /// The timestamp of the LWW CRDT is updated to be the current node's clock + /// at time of update, or the previous timestamp + 1 if that's bigger, + /// so that the new timestamp is always strictly larger than the previous one. + /// This ensures that merging the update with the old value will result in keeping + /// the updated value. pub fn update(&mut self, new_value: T) { self.ts = std::cmp::max(self.ts + 1, now_msec()); self.v = new_value; @@ -95,7 +132,20 @@ where &self.v } - /// Get a mutable value for the CRDT + /// Get a mutable reference to the CRDT's value + /// + /// This is usefull to mutate the inside value without changing the LWW timestamp. + /// When such mutation is done, the merge between two LWW values is done using the inner + /// CRDT's merge operation. This is usefull in the case where the inner CRDT is a large + /// data type, such as a map, and we only want to change a single item in the map. + /// To do this, we can produce a "CRDT delta", i.e. a LWW that contains only the modification. + /// This delta consists in a LWW with the same timestamp, and the map + /// inside only contains the updated value. + /// The advantage of such a delta is that it is much smaller than the whole map. + /// + /// Avoid using this if the inner data type is a primitive type such as a number or a string, + /// as you will then rely on the merge function defined on `Ord` types by keeping the maximum + /// of both values. pub fn get_mut(&mut self) -> &mut T { &mut self.v } @@ -115,19 +165,20 @@ where } } -/// Boolean -/// -/// with True as absorbing state +/// Boolean, where `true` is an absorbing state #[derive(Clone, Copy, Debug, Serialize, Deserialize, PartialEq)] pub struct Bool(bool); impl Bool { + /// Create a new boolean with the specified value pub fn new(b: bool) -> Self { Self(b) } + /// Set the boolean to true pub fn set(&mut self) { self.0 = true; } + /// Get the boolean value pub fn get(&self) -> bool { self.0 } @@ -141,7 +192,21 @@ impl CRDT for Bool { /// Last Write Win Map /// +/// This types defines a CRDT for a map from keys to values. +/// The values have an associated timestamp, such that the last written value +/// takes precedence over previous ones. As for the simpler `LWW` type, the value +/// type `V` is also required to implement the CRDT trait. +/// We do not encourage mutating the values associated with a given key +/// without updating the timestamp, in fact at the moment we do not provide a `.get_mut()` +/// method that would allow that. /// +/// Internally, the map is stored as a vector of keys and values, sorted by ascending key order. +/// This is why the key type `K` must implement `Ord` (and also to ensure a unique serialization, +/// such that two values can be compared for equality based on their hashes). As a consequence, +/// insertions take `O(n)` time. This means that LWWMap should be used for reasonably small maps. +/// However, note that even if we were using a more efficient data structure such as a `BTreeMap`, +/// the serialization cost `O(n)` would still have to be paid at each modification, so we are +/// actually not losing anything here. #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)] pub struct LWWMap { vals: Vec<(K, u64, V)>, @@ -152,21 +217,35 @@ where K: Ord, V: CRDT, { + /// Create a new empty map CRDT pub fn new() -> Self { Self { vals: vec![] } } + /// Used to migrate from a map defined in an incompatible format. This produces + /// a map that contains a single item with the specified timestamp (copied from + /// the incompatible format). Do this as many times as you have items to migrate, + /// and put them all together using the CRDT merge operator. pub fn migrate_from_raw_item(k: K, ts: u64, v: V) -> Self { Self { vals: vec![(k, ts, v)], } } - pub fn take_and_clear(&mut self) -> Self { - let vals = std::mem::replace(&mut self.vals, vec![]); - Self { vals } - } - pub fn clear(&mut self) { - self.vals.clear(); - } + /// Returns a map that contains a single mapping from the specified key to the specified value. + /// This map is a mutator, or a delta-CRDT, such that when it is merged with the original map, + /// the previous value will be replaced with the one specified here. + /// The timestamp in the provided mutator is set to the maximum of the current system's clock + /// and 1 + the previous value's timestamp (if there is one), so that the new value will always + /// take precedence (LWW rule). + /// + /// Typically, to update the value associated to a key in the map, you would do the following: + /// + /// ``` + /// let my_update = my_crdt.update_mutator(key_to_modify, new_value); + /// my_crdt.merge(&my_update); + /// ``` + /// + /// However extracting the mutator on its own and only sending that on the network is very + /// interesting as it is much smaller than the whole map. pub fn update_mutator(&self, k: K, new_v: V) -> Self { let new_vals = match self.vals.binary_search_by(|(k2, _, _)| k2.cmp(&k)) { Ok(i) => { @@ -178,12 +257,45 @@ where }; Self { vals: new_vals } } + /// Takes all of the values of the map and returns them. The current map is reset to the + /// empty map. This is very usefull to produce in-place a new map that contains only a delta + /// that modifies a certain value: + /// + /// ``` + /// let mut a = get_my_crdt_value(); + /// let old_a = a.take_and_clear(); + /// a.merge(&old_a.update_mutator(key_to_modify, new_value)); + /// put_my_crdt_value(a); + /// ``` + /// + /// Of course in this simple example we could have written simply + /// `pyt_my_crdt_value(a.update_mutator(key_to_modify, new_value))`, + /// but in the case where the map is a field in a struct for instance (as is always the case), + /// this becomes very handy: + /// + /// ``` + /// let mut a = get_my_crdt_value(); + /// let old_a_map = a.map_field.take_and_clear(); + /// a.map_field.merge(&old_a_map.update_mutator(key_to_modify, new_value)); + /// put_my_crdt_value(a); + /// ``` + pub fn take_and_clear(&mut self) -> Self { + let vals = std::mem::replace(&mut self.vals, vec![]); + Self { vals } + } + /// Removes all values from the map + pub fn clear(&mut self) { + self.vals.clear(); + } + /// Get a reference to the value assigned to a key pub fn get(&self, k: &K) -> Option<&V> { match self.vals.binary_search_by(|(k2, _, _)| k2.cmp(&k)) { Ok(i) => Some(&self.vals[i].2), Err(_) => None, } } + /// Gets a reference to all of the items, as a slice. Usefull to iterate on all map values. + /// In most case you will want to ignore the timestamp (second item of the tuple). pub fn items(&self) -> &[(K, u64, V)] { &self.vals[..] }