2021-11-09 11:24:04 +00:00
|
|
|
use std::cmp::Ordering;
|
2022-05-01 07:57:05 +00:00
|
|
|
use std::collections::HashMap;
|
2022-09-21 12:39:59 +00:00
|
|
|
use std::collections::HashSet;
|
|
|
|
|
|
|
|
use hex::ToHex;
|
2022-09-22 17:30:01 +00:00
|
|
|
use itertools::Itertools;
|
2021-11-09 11:24:04 +00:00
|
|
|
|
|
|
|
use serde::{Deserialize, Serialize};
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
use garage_util::crdt::{AutoCrdt, Crdt, LwwMap, Lww};
|
2021-11-09 11:24:04 +00:00
|
|
|
use garage_util::data::*;
|
2022-05-24 10:16:39 +00:00
|
|
|
use garage_util::error::*;
|
2022-05-01 07:54:19 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
use crate::graph_algo::*;
|
2021-11-09 11:24:04 +00:00
|
|
|
|
|
|
|
use crate::ring::*;
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
use std::convert::TryInto;
|
|
|
|
|
|
|
|
//The Message type will be used to collect information on the algorithm.
|
|
|
|
type Message = Vec<String>;
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
/// The layout of the cluster, i.e. the list of roles
|
|
|
|
/// which are assigned to each cluster node
|
|
|
|
#[derive(Clone, Debug, Serialize, Deserialize)]
|
|
|
|
pub struct ClusterLayout {
|
|
|
|
pub version: u64,
|
|
|
|
|
|
|
|
pub replication_factor: usize,
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//This attribute is only used to retain the previously computed partition size,
|
|
|
|
//to know to what extent does it change with the layout update.
|
2022-10-05 13:29:48 +00:00
|
|
|
#[serde(default="default_partition_size")]
|
2022-09-21 12:39:59 +00:00
|
|
|
pub partition_size: u32,
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
pub roles: LwwMap<Uuid, NodeRoleV>,
|
|
|
|
|
|
|
|
/// node_id_vec: a vector of node IDs with a role assigned
|
|
|
|
/// in the system (this includes gateway nodes).
|
|
|
|
/// The order here is different than the vec stored by `roles`, because:
|
2022-09-21 12:39:59 +00:00
|
|
|
/// 1. non-gateway nodes are first so that they have lower numbers holding
|
|
|
|
/// in u8 (the number of non-gateway nodes is at most 256).
|
2021-11-09 11:24:04 +00:00
|
|
|
/// 2. nodes that don't have a role are excluded (but they need to
|
|
|
|
/// stay in the CRDT as tombstones)
|
|
|
|
pub node_id_vec: Vec<Uuid>,
|
|
|
|
/// the assignation of data partitions to node, the values
|
|
|
|
/// are indices in node_id_vec
|
|
|
|
#[serde(with = "serde_bytes")]
|
|
|
|
pub ring_assignation_data: Vec<CompactNodeType>,
|
|
|
|
|
|
|
|
/// Role changes which are staged for the next version of the layout
|
2022-10-05 13:29:48 +00:00
|
|
|
#[serde(default="default_layout_parameters")]
|
|
|
|
pub parameters: Lww<LayoutParameters>,
|
2021-11-09 11:24:04 +00:00
|
|
|
pub staging: LwwMap<Uuid, NodeRoleV>,
|
|
|
|
pub staging_hash: Hash,
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
fn default_partition_size() -> u32{
|
2022-09-21 12:39:59 +00:00
|
|
|
return 0;
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
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
|
|
|
|
///algorithm. It is stored as a Crdt.
|
|
|
|
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
|
|
|
|
pub struct LayoutParameters {
|
|
|
|
pub zone_redundancy:usize,
|
|
|
|
}
|
|
|
|
|
|
|
|
impl AutoCrdt for LayoutParameters {
|
|
|
|
const WARN_IF_DIFFERENT: bool = true;
|
|
|
|
}
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
const NB_PARTITIONS : usize = 1usize << PARTITION_BITS;
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
|
|
|
|
pub struct NodeRoleV(pub Option<NodeRole>);
|
|
|
|
|
|
|
|
impl AutoCrdt for NodeRoleV {
|
|
|
|
const WARN_IF_DIFFERENT: bool = true;
|
|
|
|
}
|
|
|
|
|
|
|
|
/// The user-assigned roles of cluster nodes
|
|
|
|
#[derive(PartialEq, Eq, PartialOrd, Ord, Clone, Debug, Serialize, Deserialize)]
|
|
|
|
pub struct NodeRole {
|
|
|
|
/// Datacenter at which this entry belong. This information might be used to perform a better
|
|
|
|
/// geodistribution
|
|
|
|
pub zone: String,
|
|
|
|
/// The (relative) capacity of the node
|
|
|
|
/// If this is set to None, the node does not participate in storing data for the system
|
|
|
|
/// and is only active as an API gateway to other nodes
|
|
|
|
pub capacity: Option<u32>,
|
|
|
|
/// A set of tags to recognize the node
|
|
|
|
pub tags: Vec<String>,
|
|
|
|
}
|
|
|
|
|
|
|
|
impl NodeRole {
|
|
|
|
pub fn capacity_string(&self) -> String {
|
|
|
|
match self.capacity {
|
|
|
|
Some(c) => format!("{}", c),
|
|
|
|
None => "gateway".to_string(),
|
|
|
|
}
|
|
|
|
}
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
pub fn tags_string(&self) -> String {
|
|
|
|
let mut tags = String::new();
|
|
|
|
if self.tags.len() == 0 {
|
|
|
|
return tags
|
|
|
|
}
|
|
|
|
tags.push_str(&self.tags[0].clone());
|
|
|
|
for t in 1..self.tags.len(){
|
|
|
|
tags.push_str(",");
|
|
|
|
tags.push_str(&self.tags[t].clone());
|
|
|
|
}
|
|
|
|
return tags;
|
|
|
|
}
|
2021-11-09 11:24:04 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
impl ClusterLayout {
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn new(replication_factor: usize) -> Self {
|
|
|
|
|
|
|
|
//We set the default zone redundancy to be equal to the replication factor,
|
|
|
|
//i.e. as strict as possible.
|
|
|
|
let default_parameters = Lww::<LayoutParameters>::new(
|
|
|
|
LayoutParameters{ zone_redundancy: replication_factor});
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
let empty_lwwmap = LwwMap::new();
|
|
|
|
let empty_lwwmap_hash = blake2sum(&rmp_to_vec_all_named(&empty_lwwmap).unwrap()[..]);
|
|
|
|
|
|
|
|
ClusterLayout {
|
|
|
|
version: 0,
|
|
|
|
replication_factor,
|
2022-09-21 12:39:59 +00:00
|
|
|
partition_size: 0,
|
2021-11-09 11:24:04 +00:00
|
|
|
roles: LwwMap::new(),
|
|
|
|
node_id_vec: Vec::new(),
|
|
|
|
ring_assignation_data: Vec::new(),
|
2022-10-05 13:29:48 +00:00
|
|
|
parameters: default_parameters,
|
2021-11-09 11:24:04 +00:00
|
|
|
staging: empty_lwwmap,
|
|
|
|
staging_hash: empty_lwwmap_hash,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn merge(&mut self, other: &ClusterLayout) -> bool {
|
|
|
|
match other.version.cmp(&self.version) {
|
|
|
|
Ordering::Greater => {
|
|
|
|
*self = other.clone();
|
|
|
|
true
|
|
|
|
}
|
|
|
|
Ordering::Equal => {
|
2022-10-06 10:54:51 +00:00
|
|
|
let param_changed = self.parameters.get() != other.parameters.get();
|
2022-10-05 13:29:48 +00:00
|
|
|
self.parameters.merge(&other.parameters);
|
2021-11-09 11:24:04 +00:00
|
|
|
self.staging.merge(&other.staging);
|
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
let new_staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
|
2022-10-06 10:54:51 +00:00
|
|
|
let stage_changed = new_staging_hash != self.staging_hash;
|
2021-11-09 11:24:04 +00:00
|
|
|
|
|
|
|
self.staging_hash = new_staging_hash;
|
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
stage_changed || param_changed
|
2021-11-09 11:24:04 +00:00
|
|
|
}
|
|
|
|
Ordering::Less => false,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn apply_staged_changes(mut self, version: Option<u64>) -> Result<(Self,Message), Error> {
|
2022-05-24 10:16:39 +00:00
|
|
|
match version {
|
|
|
|
None => {
|
|
|
|
let error = r#"
|
|
|
|
Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
|
|
|
|
To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
|
|
|
|
"#;
|
|
|
|
return Err(Error::Message(error.into()));
|
|
|
|
}
|
|
|
|
Some(v) => {
|
|
|
|
if v != self.version + 1 {
|
|
|
|
return Err(Error::Message("Invalid new layout version".into()));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
self.roles.merge(&self.staging);
|
|
|
|
self.roles.retain(|(_, _, v)| v.0.is_some());
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
let msg = self.calculate_partition_assignation()?;
|
2022-05-24 10:16:39 +00:00
|
|
|
|
|
|
|
self.staging.clear();
|
|
|
|
self.staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
|
|
|
|
|
|
|
|
self.version += 1;
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
Ok((self,msg))
|
2022-05-24 10:16:39 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
pub fn revert_staged_changes(mut self, version: Option<u64>) -> Result<Self, Error> {
|
|
|
|
match version {
|
|
|
|
None => {
|
|
|
|
let error = r#"
|
|
|
|
Please pass the new layout version number to ensure that you are writing the correct version of the cluster layout.
|
|
|
|
To know the correct value of the new layout version, invoke `garage layout show` and review the proposed changes.
|
|
|
|
"#;
|
|
|
|
return Err(Error::Message(error.into()));
|
|
|
|
}
|
|
|
|
Some(v) => {
|
|
|
|
if v != self.version + 1 {
|
|
|
|
return Err(Error::Message("Invalid new layout version".into()));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
self.staging.clear();
|
|
|
|
self.staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
|
|
|
|
|
|
|
|
self.version += 1;
|
|
|
|
|
|
|
|
Ok(self)
|
|
|
|
}
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
/// Returns a list of IDs of nodes that currently have
|
|
|
|
/// a role in the cluster
|
|
|
|
pub fn node_ids(&self) -> &[Uuid] {
|
|
|
|
&self.node_id_vec[..]
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn num_nodes(&self) -> usize {
|
|
|
|
self.node_id_vec.len()
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Returns the role of a node in the layout
|
|
|
|
pub fn node_role(&self, node: &Uuid) -> Option<&NodeRole> {
|
|
|
|
match self.roles.get(node) {
|
|
|
|
Some(NodeRoleV(Some(v))) => Some(v),
|
|
|
|
_ => None,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
///Returns the uuids of the non_gateway nodes in self.node_id_vec.
|
|
|
|
pub fn useful_nodes(&self) -> Vec<Uuid> {
|
|
|
|
let mut result = Vec::<Uuid>::new();
|
|
|
|
for uuid in self.node_id_vec.iter() {
|
|
|
|
match self.node_role(uuid) {
|
|
|
|
Some(role) if role.capacity != None => result.push(*uuid),
|
|
|
|
_ => ()
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return result;
|
|
|
|
}
|
|
|
|
|
|
|
|
///Given a node uuids, this function returns the label of its zone
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn get_node_zone(&self, uuid : &Uuid) -> Result<String,Error> {
|
2022-09-21 12:39:59 +00:00
|
|
|
match self.node_role(uuid) {
|
|
|
|
Some(role) => return Ok(role.zone.clone()),
|
2022-10-05 13:29:48 +00:00
|
|
|
_ => return Err(Error::Message("The Uuid does not correspond to a node present in the cluster.".into()))
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
///Given a node uuids, this function returns its capacity or fails if it does not have any
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn get_node_capacity(&self, uuid : &Uuid) -> Result<u32,Error> {
|
2022-09-21 12:39:59 +00:00
|
|
|
match self.node_role(uuid) {
|
|
|
|
Some(NodeRole{capacity : Some(cap), zone: _, tags: _}) => return Ok(*cap),
|
2022-10-05 13:29:48 +00:00
|
|
|
_ => return Err(Error::Message("The Uuid does not correspond to a node present in the \
|
|
|
|
cluster or this node does not have a positive capacity.".into()))
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
///Returns the sum of capacities of non gateway nodes in the cluster
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn get_total_capacity(&self) -> Result<u32,Error> {
|
2022-09-21 12:39:59 +00:00
|
|
|
let mut total_capacity = 0;
|
|
|
|
for uuid in self.useful_nodes().iter() {
|
|
|
|
total_capacity += self.get_node_capacity(uuid)?;
|
|
|
|
}
|
|
|
|
return Ok(total_capacity);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
/// Check a cluster layout for internal consistency
|
|
|
|
/// returns true if consistent, false if error
|
|
|
|
pub fn check(&self) -> bool {
|
|
|
|
// Check that the hash of the staging data is correct
|
|
|
|
let staging_hash = blake2sum(&rmp_to_vec_all_named(&self.staging).unwrap()[..]);
|
|
|
|
if staging_hash != self.staging_hash {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check that node_id_vec contains the correct list of nodes
|
|
|
|
let mut expected_nodes = self
|
|
|
|
.roles
|
|
|
|
.items()
|
|
|
|
.iter()
|
|
|
|
.filter(|(_, _, v)| v.0.is_some())
|
|
|
|
.map(|(id, _, _)| *id)
|
|
|
|
.collect::<Vec<_>>();
|
|
|
|
expected_nodes.sort();
|
|
|
|
let mut node_id_vec = self.node_id_vec.clone();
|
|
|
|
node_id_vec.sort();
|
|
|
|
if expected_nodes != node_id_vec {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check that the assignation data has the correct length
|
|
|
|
if self.ring_assignation_data.len() != (1 << PARTITION_BITS) * self.replication_factor {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
// Check that the assigned nodes are correct identifiers
|
|
|
|
// of nodes that are assigned a role
|
|
|
|
// and that role is not the role of a gateway nodes
|
|
|
|
for x in self.ring_assignation_data.iter() {
|
|
|
|
if *x as usize >= self.node_id_vec.len() {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
let node = self.node_id_vec[*x as usize];
|
|
|
|
match self.roles.get(&node) {
|
|
|
|
Some(NodeRoleV(Some(x))) if x.capacity.is_some() => (),
|
|
|
|
_ => return false,
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-09-22 17:30:01 +00:00
|
|
|
//Check that every partition is associated to distinct nodes
|
|
|
|
let rf = self.replication_factor;
|
|
|
|
for p in 0..(1 << PARTITION_BITS) {
|
|
|
|
let nodes_of_p = self.ring_assignation_data[rf*p..rf*(p+1)].to_vec();
|
|
|
|
if nodes_of_p.iter().unique().count() != rf {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
//Check that every partition is spread over at least zone_redundancy zones.
|
|
|
|
let zones_of_p = nodes_of_p.iter()
|
|
|
|
.map(|n| self.get_node_zone(&self.node_id_vec[*n as usize])
|
|
|
|
.expect("Zone not found."));
|
2022-10-05 13:29:48 +00:00
|
|
|
let redundancy = self.parameters.get().zone_redundancy;
|
|
|
|
if zones_of_p.unique().count() < redundancy {
|
2022-09-22 17:30:01 +00:00
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//Check that the nodes capacities is consistent with the stored partitions
|
|
|
|
let mut node_usage = vec![0; MAX_NODE_NUMBER];
|
|
|
|
for n in self.ring_assignation_data.iter() {
|
|
|
|
node_usage[*n as usize] += 1;
|
|
|
|
}
|
|
|
|
for n in 0..MAX_NODE_NUMBER {
|
|
|
|
if node_usage[n] > 0 {
|
|
|
|
let uuid = self.node_id_vec[n];
|
|
|
|
if node_usage[n]*self.partition_size > self.get_node_capacity(&uuid)
|
|
|
|
.expect("Critical Error"){
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//Check that the partition size stored is the one computed by the asignation
|
|
|
|
//algorithm.
|
|
|
|
let cl2 = self.clone();
|
2022-10-06 10:54:51 +00:00
|
|
|
let (_ , zone_to_id) = cl2.generate_useful_zone_ids().expect("Critical Error");
|
2022-09-22 17:30:01 +00:00
|
|
|
let partition_size = cl2.compute_optimal_partition_size(&zone_to_id).expect("Critical Error");
|
|
|
|
if partition_size != self.partition_size {
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2021-11-09 11:24:04 +00:00
|
|
|
true
|
|
|
|
}
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
impl ClusterLayout {
|
2022-05-01 07:57:05 +00:00
|
|
|
/// This function calculates a new partition-to-node assignation.
|
2022-09-21 12:39:59 +00:00
|
|
|
/// The computed assignation respects the node replication factor
|
|
|
|
/// and the zone redundancy parameter It maximizes the capacity of a
|
2022-05-01 07:57:05 +00:00
|
|
|
/// partition (assuming all partitions have the same size).
|
|
|
|
/// Among such optimal assignation, it minimizes the distance to
|
|
|
|
/// the former assignation (if any) to minimize the amount of
|
2022-10-06 10:54:51 +00:00
|
|
|
/// data to be moved.
|
|
|
|
/// Staged changes must be merged with nodes roles before calling this function.
|
2022-10-05 13:29:48 +00:00
|
|
|
pub fn calculate_partition_assignation(&mut self) -> Result<Message,Error> {
|
2022-05-01 07:57:05 +00:00
|
|
|
//The nodes might have been updated, some might have been deleted.
|
|
|
|
//So we need to first update the list of nodes and retrieve the
|
|
|
|
//assignation.
|
2022-10-06 10:54:51 +00:00
|
|
|
|
|
|
|
//We update the node ids, since the node role list might have changed with the
|
2022-09-21 12:39:59 +00:00
|
|
|
//changes in the layout. We retrieve the old_assignation reframed with the new ids
|
|
|
|
let old_assignation_opt = self.update_node_id_vec()?;
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
let redundancy = self.parameters.get().zone_redundancy;
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
let mut msg = Message::new();
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!("Computation of a new cluster layout where partitions are \
|
2022-10-05 13:29:48 +00:00
|
|
|
replicated {} times on at least {} distinct zones.", self.replication_factor, redundancy));
|
2022-09-21 12:39:59 +00:00
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
//We generate for once numerical ids for the zones of non gateway nodes,
|
|
|
|
//to use them as indices in the flow graphs.
|
|
|
|
let (id_to_zone , zone_to_id) = self.generate_useful_zone_ids()?;
|
2022-09-21 12:39:59 +00:00
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
let nb_useful_nodes = self.useful_nodes().len();
|
2022-09-21 12:39:59 +00:00
|
|
|
msg.push(format!("The cluster contains {} nodes spread over {} zones.",
|
2022-10-06 10:54:51 +00:00
|
|
|
nb_useful_nodes, id_to_zone.len()));
|
|
|
|
if nb_useful_nodes < self.replication_factor{
|
|
|
|
return Err(Error::Message(format!("The number of nodes with positive \
|
|
|
|
capacity ({}) is smaller than the replication factor ({}).",
|
|
|
|
nb_useful_nodes, self.replication_factor)));
|
|
|
|
}
|
|
|
|
if id_to_zone.len() < redundancy {
|
|
|
|
return Err(Error::Message(format!("The number of zones with non-gateway \
|
|
|
|
nodes ({}) is smaller than the redundancy parameter ({})",
|
|
|
|
id_to_zone.len() , redundancy)));
|
|
|
|
}
|
2022-09-22 17:30:01 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
//We compute the optimal partition size
|
2022-09-22 17:30:01 +00:00
|
|
|
//Capacities should be given in a unit so that partition size is at least 100.
|
|
|
|
//In this case, integer rounding plays a marginal role in the percentages of
|
|
|
|
//optimality.
|
2022-09-21 12:39:59 +00:00
|
|
|
let partition_size = self.compute_optimal_partition_size(&zone_to_id)?;
|
2022-10-05 13:29:48 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
if old_assignation_opt != None {
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!("Given the replication and redundancy constraint, the \
|
|
|
|
optimal size of a partition is {}. In the previous layout, it used to \
|
2022-09-21 12:39:59 +00:00
|
|
|
be {}.", partition_size, self.partition_size));
|
|
|
|
}
|
|
|
|
else {
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!("Given the replication and redundancy constraints, the \
|
2022-09-21 12:39:59 +00:00
|
|
|
optimal size of a partition is {}.", partition_size));
|
|
|
|
}
|
|
|
|
self.partition_size = partition_size;
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
if partition_size < 100 {
|
|
|
|
msg.push("WARNING: The partition size is low (< 100), you might consider to \
|
2022-10-06 10:54:51 +00:00
|
|
|
provide the nodes capacities in a smaller unit (e.g. Mb instead of Gb).".into());
|
2022-10-05 13:29:48 +00:00
|
|
|
}
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
//We compute a first flow/assignment that is heuristically close to the previous
|
|
|
|
//assignment
|
|
|
|
let mut gflow = self.compute_candidate_assignment( &zone_to_id, &old_assignation_opt)?;
|
|
|
|
if let Some(assoc) = &old_assignation_opt {
|
|
|
|
//We minimize the distance to the previous assignment.
|
|
|
|
self.minimize_rebalance_load(&mut gflow, &zone_to_id, &assoc)?;
|
|
|
|
}
|
|
|
|
|
|
|
|
msg.append(&mut self.output_stat(&gflow, &old_assignation_opt, &zone_to_id,&id_to_zone)?);
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push("".to_string());
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//We update the layout structure
|
|
|
|
self.update_ring_from_flow(id_to_zone.len() , &gflow)?;
|
|
|
|
return Ok(msg);
|
|
|
|
}
|
2022-05-01 07:57:05 +00:00
|
|
|
|
|
|
|
/// The LwwMap of node roles might have changed. This function updates the node_id_vec
|
|
|
|
/// and returns the assignation given by ring, with the new indices of the nodes, and
|
2022-09-21 12:39:59 +00:00
|
|
|
/// None if the node is not present anymore.
|
2022-05-01 07:57:05 +00:00
|
|
|
/// We work with the assumption that only this function and calculate_new_assignation
|
|
|
|
/// do modify assignation_ring and node_id_vec.
|
2022-10-05 13:29:48 +00:00
|
|
|
fn update_node_id_vec(&mut self) -> Result< Option< Vec<Vec<usize> > > ,Error> {
|
2022-09-21 12:39:59 +00:00
|
|
|
// (1) We compute the new node list
|
|
|
|
//Non gateway nodes should be coded on 8bits, hence they must be first in the list
|
|
|
|
//We build the new node ids
|
|
|
|
let mut new_non_gateway_nodes: Vec<Uuid> = self.roles.items().iter()
|
|
|
|
.filter(|(_, _, v)|
|
|
|
|
match &v.0 {Some(r) if r.capacity != None => true, _=> false })
|
|
|
|
.map(|(k, _, _)| *k).collect();
|
|
|
|
|
|
|
|
if new_non_gateway_nodes.len() > MAX_NODE_NUMBER {
|
2022-10-05 13:29:48 +00:00
|
|
|
return Err(Error::Message(format!("There are more than {} non-gateway nodes in the new \
|
|
|
|
layout. This is not allowed.", MAX_NODE_NUMBER).into() ));
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
let mut new_gateway_nodes: Vec<Uuid> = self.roles.items().iter()
|
|
|
|
.filter(|(_, _, v)|
|
|
|
|
match v {NodeRoleV(Some(r)) if r.capacity == None => true, _=> false })
|
|
|
|
.map(|(k, _, _)| *k).collect();
|
2022-10-06 10:54:51 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
let nb_useful_nodes = new_non_gateway_nodes.len();
|
|
|
|
let mut new_node_id_vec = Vec::<Uuid>::new();
|
|
|
|
new_node_id_vec.append(&mut new_non_gateway_nodes);
|
|
|
|
new_node_id_vec.append(&mut new_gateway_nodes);
|
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
let old_node_id_vec = self.node_id_vec.clone();
|
|
|
|
self.node_id_vec = new_node_id_vec.clone();
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
// (2) We retrieve the old association
|
|
|
|
//We rewrite the old association with the new indices. We only consider partition
|
|
|
|
//to node assignations where the node is still in use.
|
|
|
|
let nb_partitions = 1usize << PARTITION_BITS;
|
|
|
|
let mut old_assignation = vec![ Vec::<usize>::new() ; nb_partitions];
|
|
|
|
|
|
|
|
if self.ring_assignation_data.len() == 0 {
|
|
|
|
//This is a new association
|
|
|
|
return Ok(None);
|
|
|
|
}
|
|
|
|
if self.ring_assignation_data.len() != nb_partitions * self.replication_factor {
|
2022-10-05 13:29:48 +00:00
|
|
|
return Err(Error::Message("The old assignation does not have a size corresponding to \
|
|
|
|
the old replication factor or the number of partitions.".into()));
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
//We build a translation table between the uuid and new ids
|
|
|
|
let mut uuid_to_new_id = HashMap::<Uuid, usize>::new();
|
|
|
|
|
|
|
|
//We add the indices of only the new non-gateway nodes that can be used in the
|
|
|
|
//association ring
|
|
|
|
for i in 0..nb_useful_nodes {
|
|
|
|
uuid_to_new_id.insert(new_node_id_vec[i], i );
|
|
|
|
}
|
|
|
|
|
|
|
|
let rf= self.replication_factor;
|
|
|
|
for p in 0..nb_partitions {
|
|
|
|
for old_id in &self.ring_assignation_data[p*rf..(p+1)*rf] {
|
2022-10-06 10:54:51 +00:00
|
|
|
let uuid = old_node_id_vec[*old_id as usize];
|
2022-09-21 12:39:59 +00:00
|
|
|
if uuid_to_new_id.contains_key(&uuid) {
|
|
|
|
old_assignation[p].push(uuid_to_new_id[&uuid]);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2022-10-06 10:54:51 +00:00
|
|
|
//We write the ring
|
2022-09-21 12:39:59 +00:00
|
|
|
self.ring_assignation_data = Vec::<CompactNodeType>::new();
|
|
|
|
|
|
|
|
return Ok(Some(old_assignation));
|
2022-05-01 07:57:05 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
///This function generates ids for the zone of the nodes appearing in
|
|
|
|
///self.node_id_vec.
|
2022-10-06 10:54:51 +00:00
|
|
|
fn generate_useful_zone_ids(&self) -> Result<(Vec<String>, HashMap<String, usize>),Error>{
|
2022-09-21 12:39:59 +00:00
|
|
|
let mut id_to_zone = Vec::<String>::new();
|
|
|
|
let mut zone_to_id = HashMap::<String,usize>::new();
|
2022-10-06 10:54:51 +00:00
|
|
|
|
|
|
|
for uuid in self.useful_nodes().iter() {
|
2022-09-21 12:39:59 +00:00
|
|
|
if self.roles.get(uuid) == None {
|
2022-10-05 13:29:48 +00:00
|
|
|
return Err(Error::Message("The uuid was not found in the node roles (this should \
|
|
|
|
not happen, it might be a critical error).".into()));
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
match self.node_role(&uuid) {
|
|
|
|
Some(r) => if !zone_to_id.contains_key(&r.zone) && r.capacity != None {
|
|
|
|
zone_to_id.insert(r.zone.clone() , id_to_zone.len());
|
|
|
|
id_to_zone.push(r.zone.clone());
|
|
|
|
}
|
|
|
|
_ => ()
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return Ok((id_to_zone, zone_to_id));
|
|
|
|
}
|
|
|
|
|
|
|
|
///This function computes by dichotomy the largest realizable partition size, given
|
|
|
|
///the layout.
|
2022-10-05 13:29:48 +00:00
|
|
|
fn compute_optimal_partition_size(&self, zone_to_id: &HashMap<String, usize>) -> Result<u32,Error>{
|
2022-09-21 12:39:59 +00:00
|
|
|
let nb_partitions = 1usize << PARTITION_BITS;
|
|
|
|
let empty_set = HashSet::<(usize,usize)>::new();
|
|
|
|
let mut g = self.generate_flow_graph(1, zone_to_id, &empty_set)?;
|
|
|
|
g.compute_maximal_flow()?;
|
|
|
|
if g.get_flow_value()? < (nb_partitions*self.replication_factor).try_into().unwrap() {
|
2022-10-05 13:29:48 +00:00
|
|
|
return Err(Error::Message("The storage capacity of he cluster is to small. It is \
|
|
|
|
impossible to store partitions of size 1.".into()));
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
|
|
|
|
let mut s_down = 1;
|
|
|
|
let mut s_up = self.get_total_capacity()?;
|
|
|
|
while s_down +1 < s_up {
|
|
|
|
g = self.generate_flow_graph((s_down+s_up)/2, zone_to_id, &empty_set)?;
|
|
|
|
g.compute_maximal_flow()?;
|
|
|
|
if g.get_flow_value()? < (nb_partitions*self.replication_factor).try_into().unwrap() {
|
|
|
|
s_up = (s_down+s_up)/2;
|
|
|
|
}
|
|
|
|
else {
|
|
|
|
s_down = (s_down+s_up)/2;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return Ok(s_down);
|
|
|
|
}
|
|
|
|
|
|
|
|
fn generate_graph_vertices(nb_zones : usize, nb_nodes : usize) -> Vec<Vertex> {
|
|
|
|
let mut vertices = vec![Vertex::Source, Vertex::Sink];
|
|
|
|
for p in 0..NB_PARTITIONS {
|
|
|
|
vertices.push(Vertex::Pup(p));
|
|
|
|
vertices.push(Vertex::Pdown(p));
|
|
|
|
for z in 0..nb_zones {
|
|
|
|
vertices.push(Vertex::PZ(p, z));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for n in 0..nb_nodes {
|
|
|
|
vertices.push(Vertex::N(n));
|
|
|
|
}
|
|
|
|
return vertices;
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
fn generate_flow_graph(&self, size: u32, zone_to_id: &HashMap<String, usize>, exclude_assoc : &HashSet<(usize,usize)>) -> Result<Graph<FlowEdge>, Error> {
|
2022-09-21 12:39:59 +00:00
|
|
|
let vertices = ClusterLayout::generate_graph_vertices(zone_to_id.len(),
|
|
|
|
self.useful_nodes().len());
|
|
|
|
let mut g= Graph::<FlowEdge>::new(&vertices);
|
|
|
|
let nb_zones = zone_to_id.len();
|
2022-10-05 13:29:48 +00:00
|
|
|
let redundancy = self.parameters.get().zone_redundancy;
|
2022-09-21 12:39:59 +00:00
|
|
|
for p in 0..NB_PARTITIONS {
|
2022-10-05 13:29:48 +00:00
|
|
|
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)?;
|
2022-09-21 12:39:59 +00:00
|
|
|
for z in 0..nb_zones {
|
|
|
|
g.add_edge(Vertex::Pup(p) , Vertex::PZ(p,z) , 1)?;
|
|
|
|
g.add_edge(Vertex::Pdown(p) , Vertex::PZ(p,z) ,
|
|
|
|
self.replication_factor as u32)?;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
for n in 0..self.useful_nodes().len() {
|
|
|
|
let node_capacity = self.get_node_capacity(&self.node_id_vec[n])?;
|
|
|
|
let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[n])?];
|
|
|
|
g.add_edge(Vertex::N(n), Vertex::Sink, node_capacity/size)?;
|
|
|
|
for p in 0..NB_PARTITIONS {
|
|
|
|
if !exclude_assoc.contains(&(p,n)) {
|
|
|
|
g.add_edge(Vertex::PZ(p, node_zone), Vertex::N(n), 1)?;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
return Ok(g);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
fn compute_candidate_assignment(&self, zone_to_id: &HashMap<String, usize>,
|
2022-10-05 13:29:48 +00:00
|
|
|
old_assoc_opt : &Option<Vec< Vec<usize> >>) -> Result<Graph<FlowEdge>, Error > {
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//We list the edges that are not used in the old association
|
|
|
|
let mut exclude_edge = HashSet::<(usize,usize)>::new();
|
|
|
|
if let Some(old_assoc) = old_assoc_opt {
|
|
|
|
let nb_nodes = self.useful_nodes().len();
|
|
|
|
for p in 0..NB_PARTITIONS {
|
|
|
|
for n in 0..nb_nodes {
|
|
|
|
exclude_edge.insert((p,n));
|
|
|
|
}
|
|
|
|
for n in old_assoc[p].iter() {
|
|
|
|
exclude_edge.remove(&(p,*n));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//We compute the best flow using only the edges used in the old assoc
|
|
|
|
let mut g = self.generate_flow_graph(self.partition_size, zone_to_id, &exclude_edge )?;
|
|
|
|
g.compute_maximal_flow()?;
|
|
|
|
for (p,n) in exclude_edge.iter() {
|
|
|
|
let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?];
|
|
|
|
g.add_edge(Vertex::PZ(*p,node_zone), Vertex::N(*n), 1)?;
|
|
|
|
}
|
|
|
|
g.compute_maximal_flow()?;
|
|
|
|
return Ok(g);
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
fn minimize_rebalance_load(&self, gflow: &mut Graph<FlowEdge>, zone_to_id: &HashMap<String, usize>, old_assoc : &Vec< Vec<usize> >) -> Result<(), Error > {
|
2022-09-21 12:39:59 +00:00
|
|
|
let mut cost = CostFunction::new();
|
|
|
|
for p in 0..NB_PARTITIONS {
|
|
|
|
for n in old_assoc[p].iter() {
|
|
|
|
let node_zone = zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?];
|
|
|
|
cost.insert((Vertex::PZ(p,node_zone), Vertex::N(*n)), -1);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
let nb_nodes = self.useful_nodes().len();
|
|
|
|
let path_length = 4*nb_nodes;
|
|
|
|
gflow.optimize_flow_with_cost(&cost, path_length)?;
|
|
|
|
|
|
|
|
return Ok(());
|
|
|
|
}
|
|
|
|
|
2022-10-05 13:29:48 +00:00
|
|
|
fn update_ring_from_flow(&mut self, nb_zones : usize, gflow: &Graph<FlowEdge> ) -> Result<(), Error>{
|
2022-09-21 12:39:59 +00:00
|
|
|
self.ring_assignation_data = Vec::<CompactNodeType>::new();
|
|
|
|
for p in 0..NB_PARTITIONS {
|
|
|
|
for z in 0..nb_zones {
|
|
|
|
let assoc_vertex = gflow.get_positive_flow_from(Vertex::PZ(p,z))?;
|
|
|
|
for vertex in assoc_vertex.iter() {
|
|
|
|
match vertex{
|
|
|
|
Vertex::N(n) => self.ring_assignation_data.push((*n).try_into().unwrap()),
|
|
|
|
_ => ()
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
if self.ring_assignation_data.len() != NB_PARTITIONS*self.replication_factor {
|
2022-10-05 13:29:48 +00:00
|
|
|
return Err(Error::Message("Critical Error : the association ring we produced does not \
|
|
|
|
have the right size.".into()));
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
return Ok(());
|
|
|
|
}
|
2022-09-22 17:30:01 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//This function returns a message summing up the partition repartition of the new
|
|
|
|
//layout.
|
|
|
|
fn output_stat(&self , gflow : &Graph<FlowEdge>,
|
|
|
|
old_assoc_opt : &Option< Vec<Vec<usize>> >,
|
|
|
|
zone_to_id: &HashMap<String, usize>,
|
2022-10-05 13:29:48 +00:00
|
|
|
id_to_zone : &Vec<String>) -> Result<Message, Error>{
|
2022-09-21 12:39:59 +00:00
|
|
|
let mut msg = Message::new();
|
|
|
|
|
|
|
|
let nb_partitions = 1usize << PARTITION_BITS;
|
|
|
|
let used_cap = self.partition_size * nb_partitions as u32 *
|
|
|
|
self.replication_factor as u32;
|
|
|
|
let total_cap = self.get_total_capacity()?;
|
|
|
|
let percent_cap = 100.0*(used_cap as f32)/(total_cap as f32);
|
|
|
|
msg.push(format!("Available capacity / Total cluster capacity: {} / {} ({:.1} %)",
|
|
|
|
used_cap , total_cap , percent_cap ));
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!(""));
|
|
|
|
msg.push(format!("If the percentage is to low, it might be that the \
|
|
|
|
replication/redundancy constraints force the use of nodes/zones with small \
|
|
|
|
storage capacities. \
|
|
|
|
You might want to rebalance the storage capacities or relax the constraints. \
|
|
|
|
See the detailed statistics below and look for saturated nodes/zones."));
|
2022-10-06 10:54:51 +00:00
|
|
|
msg.push(format!("Recall that because of the replication factor, the actual available \
|
2022-09-22 17:30:01 +00:00
|
|
|
storage capacity is {} / {} = {}.",
|
|
|
|
used_cap , self.replication_factor ,
|
|
|
|
used_cap/self.replication_factor as u32));
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//We define and fill in the following tables
|
|
|
|
let storing_nodes = self.useful_nodes();
|
|
|
|
let mut new_partitions = vec![0; storing_nodes.len()];
|
|
|
|
let mut stored_partitions = vec![0; storing_nodes.len()];
|
|
|
|
|
|
|
|
let mut new_partitions_zone = vec![0; id_to_zone.len()];
|
|
|
|
let mut stored_partitions_zone = vec![0; id_to_zone.len()];
|
|
|
|
|
|
|
|
for p in 0..nb_partitions {
|
|
|
|
for z in 0..id_to_zone.len() {
|
|
|
|
let pz_nodes = gflow.get_positive_flow_from(Vertex::PZ(p,z))?;
|
|
|
|
if pz_nodes.len() > 0 {
|
|
|
|
stored_partitions_zone[z] += 1;
|
2022-09-22 17:30:01 +00:00
|
|
|
if let Some(old_assoc) = old_assoc_opt {
|
|
|
|
let mut old_zones_of_p = Vec::<usize>::new();
|
|
|
|
for n in old_assoc[p].iter() {
|
|
|
|
old_zones_of_p.push(
|
|
|
|
zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?]);
|
|
|
|
}
|
|
|
|
if !old_zones_of_p.contains(&z) {
|
|
|
|
new_partitions_zone[z] += 1;
|
|
|
|
}
|
|
|
|
}
|
2022-09-21 12:39:59 +00:00
|
|
|
}
|
|
|
|
for vert in pz_nodes.iter() {
|
|
|
|
if let Vertex::N(n) = *vert {
|
|
|
|
stored_partitions[n] += 1;
|
|
|
|
if let Some(old_assoc) = old_assoc_opt {
|
|
|
|
if !old_assoc[p].contains(&n) {
|
|
|
|
new_partitions[n] += 1;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2022-09-22 17:30:01 +00:00
|
|
|
|
|
|
|
if *old_assoc_opt == None {
|
|
|
|
new_partitions = stored_partitions.clone();
|
|
|
|
new_partitions_zone = stored_partitions_zone.clone();
|
|
|
|
}
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
//We display the statistics
|
|
|
|
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!(""));
|
2022-09-21 12:39:59 +00:00
|
|
|
if *old_assoc_opt != None {
|
|
|
|
let total_new_partitions : usize = new_partitions.iter().sum();
|
|
|
|
msg.push(format!("A total of {} new copies of partitions need to be \
|
|
|
|
transferred.", total_new_partitions));
|
|
|
|
}
|
|
|
|
msg.push(format!(""));
|
2022-10-06 10:54:51 +00:00
|
|
|
msg.push(format!("==== DETAILED STATISTICS BY ZONES AND NODES ===="));
|
2022-09-21 12:39:59 +00:00
|
|
|
|
|
|
|
for z in 0..id_to_zone.len(){
|
|
|
|
let mut nodes_of_z = Vec::<usize>::new();
|
|
|
|
for n in 0..storing_nodes.len(){
|
|
|
|
if self.get_node_zone(&self.node_id_vec[n])? == id_to_zone[z] {
|
|
|
|
nodes_of_z.push(n);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
let replicated_partitions : usize = nodes_of_z.iter()
|
|
|
|
.map(|n| stored_partitions[*n]).sum();
|
|
|
|
msg.push(format!(""));
|
|
|
|
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!("Zone {}: {} distinct partitions stored ({} new, \
|
2022-09-21 12:39:59 +00:00
|
|
|
{} partition copies) ", id_to_zone[z], stored_partitions_zone[z],
|
|
|
|
new_partitions_zone[z], replicated_partitions));
|
|
|
|
|
|
|
|
let available_cap_z : u32 = self.partition_size*replicated_partitions as u32;
|
|
|
|
let mut total_cap_z = 0;
|
|
|
|
for n in nodes_of_z.iter() {
|
|
|
|
total_cap_z += self.get_node_capacity(&self.node_id_vec[*n])?;
|
|
|
|
}
|
|
|
|
let percent_cap_z = 100.0*(available_cap_z as f32)/(total_cap_z as f32);
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!(" Available capacity / Total capacity: {}/{} ({:.1}%).",
|
2022-09-21 12:39:59 +00:00
|
|
|
available_cap_z, total_cap_z, percent_cap_z));
|
|
|
|
|
|
|
|
for n in nodes_of_z.iter() {
|
|
|
|
let available_cap_n = stored_partitions[*n] as u32 *self.partition_size;
|
|
|
|
let total_cap_n =self.get_node_capacity(&self.node_id_vec[*n])?;
|
|
|
|
let tags_n = (self.node_role(&self.node_id_vec[*n])
|
|
|
|
.ok_or("Node not found."))?.tags_string();
|
2022-09-22 17:30:01 +00:00
|
|
|
msg.push(format!(" Node {}: {} partitions ({} new) ; \
|
2022-09-21 12:39:59 +00:00
|
|
|
available/total capacity: {} / {} ({:.1}%) ; tags:{}",
|
2022-09-22 17:30:01 +00:00
|
|
|
&self.node_id_vec[*n].to_vec()[0..2].to_vec().encode_hex::<String>(),
|
2022-09-21 12:39:59 +00:00
|
|
|
stored_partitions[*n],
|
|
|
|
new_partitions[*n], available_cap_n, total_cap_n,
|
|
|
|
(available_cap_n as f32)/(total_cap_n as f32)*100.0 ,
|
|
|
|
tags_n));
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
return Ok(msg);
|
|
|
|
}
|
|
|
|
|
2022-05-01 07:57:05 +00:00
|
|
|
}
|
2021-11-09 11:24:04 +00:00
|
|
|
|
2022-09-21 12:39:59 +00:00
|
|
|
//====================================================================================
|
|
|
|
|
2022-05-01 07:54:19 +00:00
|
|
|
#[cfg(test)]
|
|
|
|
mod tests {
|
2022-05-01 07:57:05 +00:00
|
|
|
use super::*;
|
2022-09-22 17:30:01 +00:00
|
|
|
use std::io::*;
|
|
|
|
// use itertools::Itertools;
|
|
|
|
/*
|
2022-05-01 07:57:05 +00:00
|
|
|
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();
|
|
|
|
//Check optimality
|
|
|
|
let node_nb_part = (0..nb_nodes)
|
|
|
|
.map(|i| {
|
|
|
|
cl.ring_assignation_data
|
|
|
|
.iter()
|
|
|
|
.filter(|x| **x == i as u8)
|
|
|
|
.count()
|
|
|
|
})
|
|
|
|
.collect::<Vec<_>>();
|
|
|
|
|
|
|
|
let zone_vec = node_zone.iter().unique().collect::<Vec<_>>();
|
|
|
|
let zone_nb_part = zone_vec
|
|
|
|
.iter()
|
|
|
|
.map(|z| {
|
|
|
|
cl.ring_assignation_data
|
|
|
|
.iter()
|
|
|
|
.filter(|x| node_zone[**x as usize] == **z)
|
|
|
|
.count()
|
|
|
|
})
|
|
|
|
.collect::<Vec<_>>();
|
|
|
|
|
|
|
|
//Check optimality of the zone assignation : would it be better for the
|
|
|
|
//node_capacity/node_partitions ratio to change the assignation of a partition
|
|
|
|
|
|
|
|
if let Some(idmin) = (0..nb_nodes).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) = (0..nb_nodes)
|
|
|
|
.filter(|i| {
|
|
|
|
if let Some(p) = zone_vec.iter().position(|z| **z == node_zone[*i]) {
|
|
|
|
zone_nb_part[p] < nb_partitions
|
|
|
|
} else {
|
|
|
|
false
|
|
|
|
}
|
|
|
|
})
|
|
|
|
.max_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
|
|
|
|
);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
//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
|
|
|
|
);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2022-09-22 17:30:01 +00:00
|
|
|
*/
|
|
|
|
|
|
|
|
fn show_msg(msg : &Message) {
|
|
|
|
for s in msg.iter(){
|
|
|
|
println!("{}",s);
|
|
|
|
}
|
|
|
|
}
|
2022-05-01 07:57:05 +00:00
|
|
|
|
|
|
|
fn update_layout(
|
|
|
|
cl: &mut ClusterLayout,
|
|
|
|
node_id_vec: &Vec<u8>,
|
|
|
|
node_capacity_vec: &Vec<u32>,
|
|
|
|
node_zone_vec: &Vec<String>,
|
|
|
|
) {
|
|
|
|
for i in 0..node_id_vec.len() {
|
|
|
|
if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) {
|
|
|
|
cl.node_id_vec.push(x);
|
|
|
|
}
|
2021-11-09 11:24:04 +00:00
|
|
|
|
2022-05-01 07:57:05 +00:00
|
|
|
let update = cl.roles.update_mutator(
|
|
|
|
cl.node_id_vec[i],
|
|
|
|
NodeRoleV(Some(NodeRole {
|
|
|
|
zone: (node_zone_vec[i].to_string()),
|
|
|
|
capacity: (Some(node_capacity_vec[i])),
|
|
|
|
tags: (vec![]),
|
|
|
|
})),
|
|
|
|
);
|
|
|
|
cl.roles.merge(&update);
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
#[test]
|
|
|
|
fn test_assignation() {
|
2022-09-22 17:30:01 +00:00
|
|
|
std::io::stdout().flush().ok().expect("Could not flush stdout");
|
|
|
|
let mut node_id_vec = vec![1, 2, 3];
|
2022-05-01 07:57:05 +00:00
|
|
|
let mut node_capacity_vec = vec![4000, 1000, 2000];
|
|
|
|
let mut node_zone_vec = vec!["A", "B", "C"]
|
|
|
|
.into_iter()
|
|
|
|
.map(|x| x.to_string())
|
|
|
|
.collect();
|
|
|
|
|
|
|
|
let mut cl = ClusterLayout {
|
|
|
|
node_id_vec: vec![],
|
2021-11-09 11:24:04 +00:00
|
|
|
|
2022-05-01 07:57:05 +00:00
|
|
|
roles: LwwMap::new(),
|
2021-11-09 11:24:04 +00:00
|
|
|
|
2022-05-01 07:57:05 +00:00
|
|
|
replication_factor: 3,
|
2022-09-22 17:30:01 +00:00
|
|
|
zone_redundancy: 1,
|
|
|
|
partition_size: 0,
|
2022-05-01 07:57:05 +00:00
|
|
|
ring_assignation_data: vec![],
|
|
|
|
version: 0,
|
|
|
|
staging: LwwMap::new(),
|
2022-09-22 17:30:01 +00:00
|
|
|
staging_hash: blake2sum(&rmp_to_vec_all_named(&LwwMap::<Uuid, NodeRoleV>::new()).unwrap()[..]),
|
2022-05-01 07:57:05 +00:00
|
|
|
};
|
|
|
|
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
|
2022-09-22 17:30:01 +00:00
|
|
|
show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
|
|
|
|
assert!(cl.check());
|
2022-05-01 07:57:05 +00:00
|
|
|
|
|
|
|
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_zone_vec = vec!["A", "B", "C", "C", "C", "B", "G", "H", "I"]
|
|
|
|
.into_iter()
|
|
|
|
.map(|x| x.to_string())
|
|
|
|
.collect();
|
|
|
|
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
|
2022-09-22 17:30:01 +00:00
|
|
|
show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
|
|
|
|
assert!(cl.check());
|
2022-05-01 07:57:05 +00:00
|
|
|
|
|
|
|
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);
|
2022-09-22 17:30:01 +00:00
|
|
|
show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
|
|
|
|
assert!(cl.check());
|
2022-05-01 07:57:05 +00:00
|
|
|
|
2022-09-22 17:30:01 +00:00
|
|
|
node_capacity_vec = vec![4000000, 4000000, 2000000, 7000000, 1000000, 9000000, 2000000, 10000, 2000000];
|
2022-05-01 07:57:05 +00:00
|
|
|
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
|
2022-09-22 17:30:01 +00:00
|
|
|
show_msg(&cl.calculate_partition_assignation(3,1).unwrap());
|
|
|
|
assert!(cl.check());
|
|
|
|
|
2022-05-01 07:57:05 +00:00
|
|
|
}
|
|
|
|
}
|