mirrors/garage
1
0
Fork 0
mirror of https://git.deuxfleurs.fr/Deuxfleurs/garage.git synced 2024-11-25 01:21:01 +00:00
Code Issues Projects Releases Packages Wiki Activity

Tests written in layout.rs

Browse source 
added staged_parameters to ClusterLayout
removed the serde(default) -> will need a migration function
This commit is contained in:
Mendes 2022-10-10 17:19:25 +02:00
parent 911eb17bd9
commit fcf9ac674a
4 changed files with 105 additions and 143 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
src/db/lib.rs
View file

@ -3,7 +3,7 @@
extern crate tracing; extern crate tracing;
#[cfg(not(any(feature = "lmdb", feature = "sled", feature = "sqlite")))] #[cfg(not(any(feature = "lmdb", feature = "sled", feature = "sqlite")))]
compile_error!("Must activate the Cargo feature for at least one DB engine: lmdb, sled or sqlite."); //compile_error!("Must activate the Cargo feature for at least one DB engine: lmdb, sled or sqlite.");
#[cfg(feature = "lmdb")] #[cfg(feature = "lmdb")]
pub mod lmdb_adapter; pub mod lmdb_adapter;

4
src/garage/cli/layout.rs
View file

@ -190,7 +190,7 @@ pub async fn cmd_show_layout(
println!(); println!();
println!("==== PARAMETERS OF THE LAYOUT COMPUTATION ===="); println!("==== PARAMETERS OF THE LAYOUT COMPUTATION ====");
println!("Zone redundancy: {}", layout.parameters.get().zone_redundancy); println!("Zone redundancy: {}", layout.staged_parameters.get().zone_redundancy);
println!(); println!();
// this will print the stats of what partitions // this will print the stats of what partitions
@ -270,7 +270,7 @@ pub async fn cmd_config_layout(
println!("The zone redundancy must be at least 1."); println!("The zone redundancy must be at least 1.");
} }
else { else {
layout.parameters.update(LayoutParameters{ zone_redundancy: r }); layout.staged_parameters.update(LayoutParameters{ zone_redundancy: r });
println!("The new zone redundancy has been saved ({}).", r); println!("The new zone redundancy has been saved ({}).", r);
} }
} }

14
src/rpc/graph_algo.rs
View file

@ -419,17 +419,3 @@ fn cycles_of_1_forest(forest: &[Option<usize>]) -> Vec<Vec<usize>> {
} }
//====================================================================================
//====================================================================================
//====================================================================================
//====================================================================================
//====================================================================================
//====================================================================================
#[cfg(test)]
mod tests {
use super::*;
}

228
src/rpc/layout.rs
View file

@ -30,8 +30,8 @@ pub struct ClusterLayout {
//This attribute is only used to retain the previously computed partition size, //This attribute is only used to retain the previously computed partition size,
//to know to what extent does it change with the layout update. //to know to what extent does it change with the layout update.
#[serde(default="default_partition_size")]
pub partition_size: u32, pub partition_size: u32,
pub parameters: LayoutParameters,
pub roles: LwwMap<Uuid, NodeRoleV>, pub roles: LwwMap<Uuid, NodeRoleV>,
@ -49,20 +49,11 @@ pub struct ClusterLayout {
pub ring_assignation_data: Vec<CompactNodeType>, pub ring_assignation_data: Vec<CompactNodeType>,
/// Role changes which are staged for the next version of the layout /// Role changes which are staged for the next version of the layout
#[serde(default="default_layout_parameters")] pub staged_parameters: Lww<LayoutParameters>,
pub parameters: Lww<LayoutParameters>,
pub staging: LwwMap<Uuid, NodeRoleV>, pub staging: LwwMap<Uuid, NodeRoleV>,
pub staging_hash: Hash, pub staging_hash: Hash,
} }
fn default_partition_size() -> u32{
0
}
fn default_layout_parameters() -> Lww<LayoutParameters>{
Lww::<LayoutParameters>::new(LayoutParameters{ zone_redundancy: 1})
}
///This struct is used to set the parameters to be used in the assignation computation ///This struct is used to set the parameters to be used in the assignation computation
///algorithm. It is stored as a Crdt. ///algorithm. It is stored as a Crdt.
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)] #[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
@ -124,8 +115,8 @@ impl ClusterLayout {
//We set the default zone redundancy to be equal to the replication factor, //We set the default zone redundancy to be equal to the replication factor,
//i.e. as strict as possible. //i.e. as strict as possible.
let default_parameters = Lww::<LayoutParameters>::new( let parameters = LayoutParameters{ zone_redundancy: replication_factor};
LayoutParameters{ zone_redundancy: replication_factor}); let staged_parameters = Lww::<LayoutParameters>::new(parameters.clone());
let empty_lwwmap = LwwMap::new(); let empty_lwwmap = LwwMap::new();
let empty_lwwmap_hash = blake2sum(&rmp_to_vec_all_named(&empty_lwwmap).unwrap()[..]); let empty_lwwmap_hash = blake2sum(&rmp_to_vec_all_named(&empty_lwwmap).unwrap()[..]);
@ -137,7 +128,8 @@ impl ClusterLayout {
roles: LwwMap::new(), roles: LwwMap::new(),
node_id_vec: Vec::new(), node_id_vec: Vec::new(),
ring_assignation_data: Vec::new(), ring_assignation_data: Vec::new(),
parameters: default_parameters, parameters,
staged_parameters,
staging: empty_lwwmap, staging: empty_lwwmap,
staging_hash: empty_lwwmap_hash, staging_hash: empty_lwwmap_hash,
} }
@ -150,8 +142,8 @@ impl ClusterLayout {
true true
} }
Ordering::Equal => { Ordering::Equal => {
let param_changed = self.parameters.get() != other.parameters.get(); let param_changed = self.staged_parameters.get() != other.staged_parameters.get();
self.parameters.merge(&other.parameters); self.staged_parameters.merge(&other.staged_parameters);
self.staging.merge(&other.staging); self.staging.merge(&other.staging);
@ -330,7 +322,7 @@ To know the correct value of the new layout version, invoke `garage layout show`
let zones_of_p = nodes_of_p.iter() let zones_of_p = nodes_of_p.iter()
.map(|n| self.get_node_zone(&self.node_id_vec[*n as usize]) .map(|n| self.get_node_zone(&self.node_id_vec[*n as usize])
.expect("Zone not found.")); .expect("Zone not found."));
let redundancy = self.parameters.get().zone_redundancy; let redundancy = self.parameters.zone_redundancy;
if zones_of_p.unique().count() < redundancy { if zones_of_p.unique().count() < redundancy {
return false; return false;
} }
@ -384,7 +376,8 @@ impl ClusterLayout {
//changes in the layout. We retrieve the old_assignation reframed with the new ids //changes in the layout. We retrieve the old_assignation reframed with the new ids
let old_assignation_opt = self.update_node_id_vec()?; let old_assignation_opt = self.update_node_id_vec()?;
let redundancy = self.parameters.get().zone_redundancy; let redundancy = self.staged_parameters.get().zone_redundancy;
let mut msg = Message::new(); let mut msg = Message::new();
msg.push(format!("Computation of a new cluster layout where partitions are \ msg.push(format!("Computation of a new cluster layout where partitions are \
@ -417,13 +410,15 @@ impl ClusterLayout {
if old_assignation_opt != None { if old_assignation_opt != None {
msg.push(format!("Given the replication and redundancy constraint, the \ msg.push(format!("Given the replication and redundancy constraint, the \
optimal size of a partition is {}. In the previous layout, it used to \ optimal size of a partition is {}. In the previous layout, it used to \
be {}.", partition_size, self.partition_size)); be {} (the zone redundancy was {}).", partition_size, self.partition_size,
self.parameters.zone_redundancy));
} }
else { else {
msg.push(format!("Given the replication and redundancy constraints, the \ msg.push(format!("Given the replication and redundancy constraints, the \
optimal size of a partition is {}.", partition_size)); optimal size of a partition is {}.", partition_size));
} }
self.partition_size = partition_size; self.partition_size = partition_size;
self.parameters = self.staged_parameters.get().clone();
if partition_size < 100 { if partition_size < 100 {
msg.push("WARNING: The partition size is low (< 100), you might consider to \ msg.push("WARNING: The partition size is low (< 100), you might consider to \
@ -511,6 +506,10 @@ impl ClusterLayout {
//We write the ring //We write the ring
self.ring_assignation_data = Vec::<CompactNodeType>::new(); self.ring_assignation_data = Vec::<CompactNodeType>::new();
if !self.check() {
return Err(Error::Message("Critical error: The computed layout happens to be incorrect".into()));
}
Ok(Some(old_assignation)) Ok(Some(old_assignation))
} }
@ -585,7 +584,7 @@ impl ClusterLayout {
self.useful_nodes().len()); self.useful_nodes().len());
let mut g= Graph::<FlowEdge>::new(&vertices); let mut g= Graph::<FlowEdge>::new(&vertices);
let nb_zones = zone_to_id.len(); let nb_zones = zone_to_id.len();
let redundancy = self.parameters.get().zone_redundancy; let redundancy = self.staged_parameters.get().zone_redundancy;
for p in 0..NB_PARTITIONS { for p in 0..NB_PARTITIONS {
g.add_edge(Vertex::Source, Vertex::Pup(p), redundancy as u32)?; g.add_edge(Vertex::Source, Vertex::Pup(p), redundancy as u32)?;
g.add_edge(Vertex::Source, Vertex::Pdown(p), (self.replication_factor - redundancy) as u32)?; g.add_edge(Vertex::Source, Vertex::Pdown(p), (self.replication_factor - redundancy) as u32)?;
@ -800,96 +799,80 @@ impl ClusterLayout {
#[cfg(test)] #[cfg(test)]
mod tests { mod tests {
use super::*; use super::{*,Error};
use std::io::*; use std::cmp::min;
// use itertools::Itertools;
/*
fn check_assignation(cl: &ClusterLayout) {
//Check that input data has the right format
let nb_partitions = 1usize << PARTITION_BITS;
assert!(cl.ring_assignation_data.len() == nb_partitions * cl.replication_factor);
//Check that is is a correct assignation with zone redundancy
let rf = cl.replication_factor;
for i in 0..nb_partitions {
assert!(
rf == cl.ring_assignation_data[rf * i..rf * (i + 1)]
.iter()
.map(|nod| node_zone[*nod as usize].clone())
.unique()
.count()
);
}
let nb_nodes = cl.node_id_vec.len(); //This function checks that the partition size S computed is at least better than the
//Check optimality //one given by a very naive algorithm. To do so, we try to run the naive algorithm
let node_nb_part = (0..nb_nodes) //assuming a partion size of S+1. If we succed, it means that the optimal assignation
.map(|i| { //was not optimal. The naive algorithm is the following :
cl.ring_assignation_data //- we compute the max number of partitions associated to every node, capped at the
.iter() //partition number. It gives the number of tokens of every node.
.filter(|x| **x == i as u8) //- every zone has a number of tokens equal to the sum of the tokens of its nodes.
.count() //- we cycle over the partitions and associate zone tokens while respecting the
}) //zone redundancy constraint.
.collect::<Vec<_>>(); //NOTE: the naive algorithm is not optimal. Counter example:
//take nb_partition = 3 ; replication_factor = 5; redundancy = 4;
//number of tokens by zone : (A, 4), (B,1), (C,4), (D, 4), (E, 2)
//With these parameters, the naive algo fails, whereas there is a solution:
//(A,A,C,D,E) , (A,B,C,D,D) (A,C,C,D,E)
fn check_against_naive(cl: &ClusterLayout) -> Result<bool,Error> {
let over_size = cl.partition_size +1;
let mut zone_token = HashMap::<String, usize>::new();
let nb_partitions = 1usize << PARTITION_BITS;
let (zones, zone_to_id) = cl.generate_useful_zone_ids()?;
if zones.is_empty() {
return Ok(false);
}
let zone_vec = node_zone.iter().unique().collect::<Vec<_>>(); for z in zones.iter() {
let zone_nb_part = zone_vec zone_token.insert(z.clone(), 0);
.iter() }
.map(|z| { for uuid in cl.useful_nodes().iter() {
cl.ring_assignation_data let z = cl.get_node_zone(uuid)?;
.iter() let c = cl.get_node_capacity(uuid)?;
.filter(|x| node_zone[**x as usize] == **z) zone_token.insert(z.clone(), zone_token[&z] + min(nb_partitions , (c/over_size) as usize));
.count() }
})
.collect::<Vec<_>>(); //For every partition, we count the number of zone already associated and
//the name of the last zone associated
//Check optimality of the zone assignation : would it be better for the let mut id_zone_token = vec![0; zones.len()];
//node_capacity/node_partitions ratio to change the assignation of a partition for (z,t) in zone_token.iter() {
id_zone_token[zone_to_id[z]] = *t;
}
if let Some(idmin) = (0..nb_nodes).min_by(|i, j| { let mut nb_token = vec![0; nb_partitions];
(node_capacity[*i] * node_nb_part[*j] as u32) let mut last_zone = vec![zones.len(); nb_partitions];
.cmp(&(node_capacity[*j] * node_nb_part[*i] as u32))
}) { let mut curr_zone = 0;
if let Some(idnew) = (0..nb_nodes)
.filter(|i| { let redundancy = cl.parameters.zone_redundancy;
if let Some(p) = zone_vec.iter().position(|z| **z == node_zone[*i]) {
zone_nb_part[p] < nb_partitions for replic in 0..cl.replication_factor {
} else { for p in 0..nb_partitions {
false while id_zone_token[curr_zone] == 0 ||
} (last_zone[p] == curr_zone
}) && redundancy - nb_token[p] <= cl.replication_factor - replic) {
.max_by(|i, j| { curr_zone += 1;
(node_capacity[*i] * (node_nb_part[*j] as u32 + 1)) if curr_zone >= zones.len() {
.cmp(&(node_capacity[*j] * (node_nb_part[*i] as u32 + 1))) return Ok(true);
}) { }
assert!( }
node_capacity[idmin] * (node_nb_part[idnew] as u32 + 1) id_zone_token[curr_zone] -= 1;
>= node_capacity[idnew] * node_nb_part[idmin] as u32 if last_zone[p] != curr_zone {
); nb_token[p] += 1;
} last_zone[p] = curr_zone;
} }
}
}
return Ok(false);
}
//In every zone, check optimality of the nod assignation
for z in zone_vec {
let node_of_z_iter = (0..nb_nodes).filter(|id| node_zone[*id] == *z);
if let Some(idmin) = node_of_z_iter.clone().min_by(|i, j| {
(node_capacity[*i] * node_nb_part[*j] as u32)
.cmp(&(node_capacity[*j] * node_nb_part[*i] as u32))
}) {
if let Some(idnew) = node_of_z_iter.min_by(|i, j| {
(node_capacity[*i] * (node_nb_part[*j] as u32 + 1))
.cmp(&(node_capacity[*j] * (node_nb_part[*i] as u32 + 1)))
}) {
assert!(
node_capacity[idmin] * (node_nb_part[idnew] as u32 + 1)
>= node_capacity[idnew] * node_nb_part[idmin] as u32
);
}
}
}
}
*/
fn show_msg(msg : &Message) { fn show_msg(msg : &Message) {
for s in msg.iter(){ for s in msg.iter(){
println!("{}",s); println!("{}",s);
@ -901,6 +884,7 @@ mod tests {
node_id_vec: &Vec<u8>, node_id_vec: &Vec<u8>,
node_capacity_vec: &Vec<u32>, node_capacity_vec: &Vec<u32>,
node_zone_vec: &Vec<String>, node_zone_vec: &Vec<String>,
zone_redundancy: usize
) { ) {
for i in 0..node_id_vec.len() { for i in 0..node_id_vec.len() {
if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) { if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) {
@ -917,11 +901,11 @@ mod tests {
); );
cl.roles.merge(&update); cl.roles.merge(&update);
} }
cl.staged_parameters = Lww::<LayoutParameters>::new(LayoutParameters{zone_redundancy});
} }
#[test] #[test]
fn test_assignation() { fn test_assignation() {
std::io::stdout().flush().ok().expect("Could not flush stdout");
let mut node_id_vec = vec![1, 2, 3]; let mut node_id_vec = vec![1, 2, 3];
let mut node_capacity_vec = vec![4000, 1000, 2000]; let mut node_capacity_vec = vec![4000, 1000, 2000];
let mut node_zone_vec = vec!["A", "B", "C"] let mut node_zone_vec = vec!["A", "B", "C"]
@ -929,22 +913,11 @@ mod tests {
.map(|x| x.to_string()) .map(|x| x.to_string())
.collect(); .collect();
let mut cl = ClusterLayout { let mut cl = ClusterLayout::new(3);
node_id_vec: vec![], update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 3);
show_msg(&cl.calculate_partition_assignation().unwrap());
roles: LwwMap::new(),
replication_factor: 3,
zone_redundancy: 1,
partition_size: 0,
ring_assignation_data: vec![],
version: 0,
staging: LwwMap::new(),
staging_hash: blake2sum(&rmp_to_vec_all_named(&LwwMap::<Uuid, NodeRoleV>::new()).unwrap()[..]),
};
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
assert!(cl.check()); assert!(cl.check());
assert!(matches!(check_against_naive(&cl), Ok(true)));
node_id_vec = vec![1, 2, 3, 4, 5, 6, 7, 8, 9]; node_id_vec = vec![1, 2, 3, 4, 5, 6, 7, 8, 9];
node_capacity_vec = vec![4000, 1000, 1000, 3000, 1000, 1000, 2000, 10000, 2000]; node_capacity_vec = vec![4000, 1000, 1000, 3000, 1000, 1000, 2000, 10000, 2000];
@ -952,19 +925,22 @@ mod tests {
.into_iter() .into_iter()
.map(|x| x.to_string()) .map(|x| x.to_string())
.collect(); .collect();
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec); update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 2);
show_msg(&cl.calculate_partition_assignation(3,3).unwrap()); show_msg(&cl.calculate_partition_assignation().unwrap());
assert!(cl.check()); assert!(cl.check());
assert!(matches!(check_against_naive(&cl), Ok(true)));
node_capacity_vec = vec![4000, 1000, 2000, 7000, 1000, 1000, 2000, 10000, 2000]; node_capacity_vec = vec![4000, 1000, 2000, 7000, 1000, 1000, 2000, 10000, 2000];
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec); update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 3);
show_msg(&cl.calculate_partition_assignation(3,3).unwrap()); show_msg(&cl.calculate_partition_assignation().unwrap());
assert!(cl.check()); assert!(cl.check());
assert!(matches!(check_against_naive(&cl), Ok(true)));
node_capacity_vec = vec![4000000, 4000000, 2000000, 7000000, 1000000, 9000000, 2000000, 10000, 2000000]; node_capacity_vec = vec![4000000, 4000000, 2000000, 7000000, 1000000, 9000000, 2000000, 10000, 2000000];
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec); update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec, 1);
show_msg(&cl.calculate_partition_assignation(3,1).unwrap()); show_msg(&cl.calculate_partition_assignation().unwrap());
assert!(cl.check()); assert!(cl.check());
assert!(matches!(check_against_naive(&cl), Ok(true)));
} }
} }