1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
|
package app
import (
"sort"
"github.com/cowsql/go-cowsql/client"
)
const minVoters = 3
// RolesConfig can be used to tweak the algorithm implemented by RolesChanges.
type RolesConfig struct {
Voters int // Target number of voters, 3 by default.
StandBys int // Target number of stand-bys, 3 by default.
}
// RolesChanges implements an algorithm to take decisions about which node
// should have which role in a cluster.
//
// You normally don't need to use this data structure since it's already
// transparently wired into the high-level App object. However this is exposed
// for users who don't want to use the high-level App object but still want to
// implement the same roles management algorithm.
type RolesChanges struct {
// Algorithm configuration.
Config RolesConfig
// Current state of the cluster. Each node in the cluster must be
// present as a key in the map, and its value should be its associated
// failure domain and weight metadata or nil if the node is currently
// offline.
State map[client.NodeInfo]*client.NodeMetadata
}
// Assume decides if a node should assume a different role than the one it
// currently has. It should normally be run at node startup, where the
// algorithm might decide that the node should assume the Voter or Stand-By
// role in case there's a shortage of them.
//
// Return -1 in case no role change is needed.
func (c *RolesChanges) Assume(id uint64) client.NodeRole {
// If the cluster is still too small, do nothing.
if c.size() < minVoters {
return -1
}
node := c.get(id)
// If we are not in the cluster, it means we were removed, just do nothing.
if node == nil {
return -1
}
// If we already have the Voter or StandBy role, there's nothing to do.
if node.Role == client.Voter || node.Role == client.StandBy {
return -1
}
onlineVoters := c.list(client.Voter, true)
onlineStandbys := c.list(client.StandBy, true)
// If we have already the desired number of online voters and
// stand-bys, there's nothing to do.
if len(onlineVoters) >= c.Config.Voters && len(onlineStandbys) >= c.Config.StandBys {
return -1
}
// Figure if we need to become stand-by or voter.
role := client.StandBy
if len(onlineVoters) < c.Config.Voters {
role = client.Voter
}
return role
}
// Handover decides if a node should transfer its current role to another
// node. This is typically run when the node is shutting down and is hence going to be offline soon.
//
// Return the role that should be handed over and list of candidates that
// should receive it, in order of preference.
func (c *RolesChanges) Handover(id uint64) (client.NodeRole, []client.NodeInfo) {
node := c.get(id)
// If we are not in the cluster, it means we were removed, just do nothing.
if node == nil {
return -1, nil
}
// If we aren't a voter or a stand-by, there's nothing to do.
if node.Role != client.Voter && node.Role != client.StandBy {
return -1, nil
}
// Make a list of all online nodes with the same role and get their
// failure domains.
peers := c.list(node.Role, true)
for i := range peers {
if peers[i].ID == node.ID {
peers = append(peers[:i], peers[i+1:]...)
break
}
}
domains := c.failureDomains(peers)
// Online spare nodes are always candidates.
candidates := c.list(client.Spare, true)
// Stand-by nodes are candidates if we need to transfer voting
// rights, and they are preferred over spares.
if node.Role == client.Voter {
candidates = append(c.list(client.StandBy, true), candidates...)
}
if len(candidates) == 0 {
// No online node available to be promoted.
return -1, nil
}
c.sortCandidates(candidates, domains)
return node.Role, candidates
}
// Adjust decides if there should be changes in the current roles.
//
// Return the role that should be assigned and a list of candidates that should
// assume it, in order of preference.
func (c *RolesChanges) Adjust(leader uint64) (client.NodeRole, []client.NodeInfo) {
if c.size() == 1 {
return -1, nil
}
// If the cluster is too small, make sure we have just one voter (us).
if c.size() < minVoters {
for node := range c.State {
if node.ID == leader || node.Role != client.Voter {
continue
}
return client.Spare, []client.NodeInfo{node}
}
return -1, nil
}
onlineVoters := c.list(client.Voter, true)
onlineStandbys := c.list(client.StandBy, true)
offlineVoters := c.list(client.Voter, false)
offlineStandbys := c.list(client.StandBy, false)
// If we have exactly the desired number of voters and stand-bys, and they are all
// online, we're good.
if len(offlineVoters) == 0 && len(onlineVoters) == c.Config.Voters && len(offlineStandbys) == 0 && len(onlineStandbys) == c.Config.StandBys {
return -1, nil
}
// If we have less online voters than desired, let's try to promote
// some other node.
if n := len(onlineVoters); n < c.Config.Voters {
candidates := c.list(client.StandBy, true)
candidates = append(candidates, c.list(client.Spare, true)...)
if len(candidates) == 0 {
return -1, nil
}
domains := c.failureDomains(onlineVoters)
c.sortCandidates(candidates, domains)
return client.Voter, candidates
}
// If we have more online voters than desired, let's demote one of
// them.
if n := len(onlineVoters); n > c.Config.Voters {
nodes := []client.NodeInfo{}
for _, node := range onlineVoters {
// Don't demote the leader.
if node.ID == leader {
continue
}
nodes = append(nodes, node)
}
return client.Spare, nodes
}
// If we have offline voters, let's demote one of them.
if n := len(offlineVoters); n > 0 {
return client.Spare, offlineVoters
}
// If we have less online stand-bys than desired, let's try to promote
// some other node.
if n := len(onlineStandbys); n < c.Config.StandBys {
candidates := c.list(client.Spare, true)
if len(candidates) == 0 {
return -1, nil
}
domains := c.failureDomains(onlineStandbys)
c.sortCandidates(candidates, domains)
return client.StandBy, candidates
}
// If we have more online stand-bys than desired, let's demote one of
// them.
if n := len(onlineStandbys); n > c.Config.StandBys {
nodes := []client.NodeInfo{}
for _, node := range onlineStandbys {
// Don't demote the leader.
if node.ID == leader {
continue
}
nodes = append(nodes, node)
}
return client.Spare, nodes
}
// If we have offline stand-bys, let's demote one of them.
if n := len(offlineStandbys); n > 0 {
return client.Spare, offlineStandbys
}
return -1, nil
}
// Return the number of nodes il the cluster.
func (c *RolesChanges) size() int {
return len(c.State)
}
// Return information about the node with the given ID, or nil if no node
// matches.
func (c *RolesChanges) get(id uint64) *client.NodeInfo {
for node := range c.State {
if node.ID == id {
return &node
}
}
return nil
}
// Return the online or offline nodes with the given role.
func (c *RolesChanges) list(role client.NodeRole, online bool) []client.NodeInfo {
nodes := []client.NodeInfo{}
for node, metadata := range c.State {
if node.Role == role && metadata != nil == online {
nodes = append(nodes, node)
}
}
return nodes
}
// Return the number of online or offline nodes with the given role.
func (c *RolesChanges) count(role client.NodeRole, online bool) int {
return len(c.list(role, online))
}
// Return a map of the failure domains associated with the
// given nodes.
func (c *RolesChanges) failureDomains(nodes []client.NodeInfo) map[uint64]bool {
domains := map[uint64]bool{}
for _, node := range nodes {
metadata := c.State[node]
if metadata == nil {
continue
}
domains[metadata.FailureDomain] = true
}
return domains
}
// Sort the given candidates according to their failure domain and
// weight. Candidates belonging to a failure domain different from the given
// domains take precedence.
func (c *RolesChanges) sortCandidates(candidates []client.NodeInfo, domains map[uint64]bool) {
less := func(i, j int) bool {
metadata1 := c.metadata(candidates[i])
metadata2 := c.metadata(candidates[j])
// If i's failure domain is not in the given list, but j's is,
// then i takes precedence.
if !domains[metadata1.FailureDomain] && domains[metadata2.FailureDomain] {
return true
}
// If j's failure domain is not in the given list, but i's is,
// then j takes precedence.
if !domains[metadata2.FailureDomain] && domains[metadata1.FailureDomain] {
return false
}
return metadata1.Weight < metadata2.Weight
}
sort.Slice(candidates, less)
}
// Return the metadata of the given node, if any.
func (c *RolesChanges) metadata(node client.NodeInfo) *client.NodeMetadata {
return c.State[node]
}
|