|
| 1 | +/* |
| 2 | +Copyright 2023 The Vitess Authors. |
| 3 | +
|
| 4 | +Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | +you may not use this file except in compliance with the License. |
| 6 | +You may obtain a copy of the License at |
| 7 | +
|
| 8 | + http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | +
|
| 10 | +Unless required by applicable law or agreed to in writing, software |
| 11 | +distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | +See the License for the specific language governing permissions and |
| 14 | +limitations under the License. |
| 15 | +*/ |
| 16 | + |
| 17 | +package balancer |
| 18 | + |
| 19 | +import ( |
| 20 | + "encoding/json" |
| 21 | + "fmt" |
| 22 | + "math/rand" |
| 23 | + "net/http" |
| 24 | + "sync" |
| 25 | + |
| 26 | + "vitess.io/vitess/go/vt/discovery" |
| 27 | + querypb "vitess.io/vitess/go/vt/proto/query" |
| 28 | +) |
| 29 | + |
| 30 | +/* |
| 31 | +
|
| 32 | +The tabletBalancer probabalistically orders the list of available tablets into |
| 33 | +a ranked order of preference in order to satisfy two high-level goals: |
| 34 | +
|
| 35 | +1. Balance the load across the available replicas |
| 36 | +2. Prefer a replica in the same cell as the vtgate if possible |
| 37 | +
|
| 38 | +In some topologies this is trivial to accomplish by simply preferring tablets in the |
| 39 | +local cell, assuming there are a proportional number of local tablets in each cell to |
| 40 | +satisfy the inbound traffic to the vtgates in that cell. |
| 41 | +
|
| 42 | +However, for topologies with a relatively small number of tablets in each cell, a simple |
| 43 | +affinity algorithm does not effectively balance the load. |
| 44 | +
|
| 45 | +As a simple example: |
| 46 | +
|
| 47 | + Given three cells with vtgates, four replicas spread into those cells, where each vtgate |
| 48 | + receives an equal query share. If each routes only to its local cell, the tablets will be |
| 49 | + unbalanced since two of them receive 1/3 of the queries, but the two replicas in the same |
| 50 | + cell will only receive 1/6 of the queries. |
| 51 | +
|
| 52 | + Cell A: 1/3 --> vtgate --> 1/3 => vttablet |
| 53 | +
|
| 54 | + Cell B: 1/3 --> vtgate --> 1/3 => vttablet |
| 55 | +
|
| 56 | + Cell C: 1/3 --> vtgate --> 1/6 => vttablet |
| 57 | + \-> 1/6 => vttablet |
| 58 | +
|
| 59 | +Other topologies that can cause similar pathologies include cases where there may be cells |
| 60 | +containing replicas but no local vtgates, and/or cells that have only vtgates but no replicas. |
| 61 | +
|
| 62 | +For these topologies, the tabletBalancer proportionally assigns the output flow to each tablet, |
| 63 | +preferring the local cell where possible, but only as long as the global query balance is |
| 64 | +maintained. |
| 65 | +
|
| 66 | +To accomplish this goal, the balancer is given: |
| 67 | +
|
| 68 | +* The list of cells that receive inbound traffic to vtgates |
| 69 | +* The local cell where the vtgate exists |
| 70 | +* The set of tablets and their cells (learned from discovery) |
| 71 | +
|
| 72 | +The model assumes there is an equal probablility of a query coming from each vtgate cell, i.e. |
| 73 | +traffic is effectively load balanced between the cells with vtgates. |
| 74 | +
|
| 75 | +Given that information, the balancer builds a simple model to determine how much query load |
| 76 | +would go to each tablet if vtgate only routed to its local cell. Then if any tablets are |
| 77 | +unbalanced, it shifts the desired allocation away from the local cell preference in order to |
| 78 | +even out the query load. |
| 79 | +
|
| 80 | +Based on this global model, the vtgate then probabalistically picks a destination for each |
| 81 | +query to be sent and uses these weights to order the available tablets accordingly. |
| 82 | +
|
| 83 | +Assuming each vtgate is configured with and discovers the same information about the topology, |
| 84 | +and the input flow is balanced across the vtgate cells (as mentioned above), then each vtgate |
| 85 | +should come the the same conclusion about the global flows, and cooperatively should |
| 86 | +converge on the desired balanced query load. |
| 87 | +
|
| 88 | +*/ |
| 89 | + |
| 90 | +type TabletBalancer interface { |
| 91 | + // Pick is the main entry point to the balancer. Returns the best tablet out of the list |
| 92 | + // for a given query to maintain the desired balanced allocation over multiple executions. |
| 93 | + Pick(target *querypb.Target, tablets []*discovery.TabletHealth) *discovery.TabletHealth |
| 94 | + |
| 95 | + // Balancer debug page request |
| 96 | + DebugHandler(w http.ResponseWriter, r *http.Request) |
| 97 | +} |
| 98 | + |
| 99 | +func NewTabletBalancer(localCell string, vtGateCells []string) TabletBalancer { |
| 100 | + return &tabletBalancer{ |
| 101 | + localCell: localCell, |
| 102 | + vtGateCells: vtGateCells, |
| 103 | + allocations: map[discovery.KeyspaceShardTabletType]*targetAllocation{}, |
| 104 | + } |
| 105 | +} |
| 106 | + |
| 107 | +type tabletBalancer struct { |
| 108 | + // |
| 109 | + // Configuration |
| 110 | + // |
| 111 | + |
| 112 | + // The local cell for the vtgate |
| 113 | + localCell string |
| 114 | + |
| 115 | + // The set of cells that have vtgates |
| 116 | + vtGateCells []string |
| 117 | + |
| 118 | + // mu protects the allocation map |
| 119 | + mu sync.Mutex |
| 120 | + |
| 121 | + // |
| 122 | + // Allocations for balanced mode, calculated once per target and invalidated |
| 123 | + // whenever the topology changes. |
| 124 | + // |
| 125 | + allocations map[discovery.KeyspaceShardTabletType]*targetAllocation |
| 126 | +} |
| 127 | + |
| 128 | +type targetAllocation struct { |
| 129 | + // Target flow per cell based on the number of tablets discovered in the cell |
| 130 | + Target map[string]int // json:target |
| 131 | + |
| 132 | + // Input flows allocated for each cell |
| 133 | + Inflows map[string]int |
| 134 | + |
| 135 | + // Output flows from each vtgate cell to each target cell |
| 136 | + Outflows map[string]map[string]int |
| 137 | + |
| 138 | + // Allocation routed to each tablet from the local cell used for ranking |
| 139 | + Allocation map[uint32]int |
| 140 | + |
| 141 | + // Tablets that local cell does not route to |
| 142 | + Unallocated map[uint32]struct{} |
| 143 | + |
| 144 | + // Total allocation which is basically 1,000,000 / len(vtgatecells) |
| 145 | + TotalAllocation int |
| 146 | +} |
| 147 | + |
| 148 | +func (b *tabletBalancer) print() string { |
| 149 | + allocations, _ := json.Marshal(&b.allocations) |
| 150 | + return fmt.Sprintf("LocalCell: %s, VtGateCells: %s, allocations: %s", |
| 151 | + b.localCell, b.vtGateCells, string(allocations)) |
| 152 | +} |
| 153 | + |
| 154 | +func (b *tabletBalancer) DebugHandler(w http.ResponseWriter, _ *http.Request) { |
| 155 | + w.Header().Set("Content-Type", "text/plain") |
| 156 | + fmt.Fprintf(w, "Local Cell: %v\r\n", b.localCell) |
| 157 | + fmt.Fprintf(w, "Vtgate Cells: %v\r\n", b.vtGateCells) |
| 158 | + |
| 159 | + b.mu.Lock() |
| 160 | + defer b.mu.Unlock() |
| 161 | + allocations, _ := json.MarshalIndent(b.allocations, "", " ") |
| 162 | + fmt.Fprintf(w, "Allocations: %v\r\n", string(allocations)) |
| 163 | +} |
| 164 | + |
| 165 | +// Pick is the main entry point to the balancer. |
| 166 | +// |
| 167 | +// Given the total allocation for the set of tablets, choose the best target |
| 168 | +// by a weighted random sample so that over time the system will achieve the |
| 169 | +// desired balanced allocation. |
| 170 | +func (b *tabletBalancer) Pick(target *querypb.Target, tablets []*discovery.TabletHealth) *discovery.TabletHealth { |
| 171 | + |
| 172 | + numTablets := len(tablets) |
| 173 | + if numTablets == 0 { |
| 174 | + return nil |
| 175 | + } |
| 176 | + |
| 177 | + allocationMap, totalAllocation := b.getAllocation(target, tablets) |
| 178 | + |
| 179 | + r := rand.Intn(totalAllocation) |
| 180 | + for i := 0; i < numTablets; i++ { |
| 181 | + flow := allocationMap[tablets[i].Tablet.Alias.Uid] |
| 182 | + if r < flow { |
| 183 | + return tablets[i] |
| 184 | + } |
| 185 | + r -= flow |
| 186 | + } |
| 187 | + |
| 188 | + return tablets[0] |
| 189 | +} |
| 190 | + |
| 191 | +// To stick with integer arithmetic, use 1,000,000 as the full load |
| 192 | +const ALLOCATION = 1000000 |
| 193 | + |
| 194 | +func (b *tabletBalancer) allocateFlows(allTablets []*discovery.TabletHealth) *targetAllocation { |
| 195 | + // Initialization: Set up some data structures and derived values |
| 196 | + a := targetAllocation{ |
| 197 | + Target: map[string]int{}, |
| 198 | + Inflows: map[string]int{}, |
| 199 | + Outflows: map[string]map[string]int{}, |
| 200 | + Allocation: map[uint32]int{}, |
| 201 | + Unallocated: map[uint32]struct{}{}, |
| 202 | + } |
| 203 | + flowPerVtgateCell := ALLOCATION / len(b.vtGateCells) |
| 204 | + flowPerTablet := ALLOCATION / len(allTablets) |
| 205 | + cellExistsWithNoTablets := false |
| 206 | + |
| 207 | + for _, th := range allTablets { |
| 208 | + a.Target[th.Tablet.Alias.Cell] += flowPerTablet |
| 209 | + } |
| 210 | + |
| 211 | + // |
| 212 | + // First pass: Allocate vtgate flow to the local cell where the vtgate exists |
| 213 | + // and along the way figure out if there are any vtgates with no local tablets. |
| 214 | + // |
| 215 | + for _, cell := range b.vtGateCells { |
| 216 | + outflow := map[string]int{} |
| 217 | + target := a.Target[cell] |
| 218 | + |
| 219 | + if target > 0 { |
| 220 | + a.Inflows[cell] += flowPerVtgateCell |
| 221 | + outflow[cell] = flowPerVtgateCell |
| 222 | + } else { |
| 223 | + cellExistsWithNoTablets = true |
| 224 | + } |
| 225 | + |
| 226 | + a.Outflows[cell] = outflow |
| 227 | + } |
| 228 | + |
| 229 | + // |
| 230 | + // Figure out if there is a shortfall |
| 231 | + // |
| 232 | + underAllocated := make(map[string]int) |
| 233 | + unbalancedFlow := 0 |
| 234 | + for cell, allocation := range a.Target { |
| 235 | + if a.Inflows[cell] < allocation { |
| 236 | + underAllocated[cell] = allocation - a.Inflows[cell] |
| 237 | + unbalancedFlow += underAllocated[cell] |
| 238 | + } |
| 239 | + } |
| 240 | + |
| 241 | + // |
| 242 | + // Second pass: if there are any vtgates with no local tablets, allocate the underallocated amount |
| 243 | + // proportionally to all cells that may need it |
| 244 | + // |
| 245 | + if cellExistsWithNoTablets { |
| 246 | + for _, vtgateCell := range b.vtGateCells { |
| 247 | + target := a.Target[vtgateCell] |
| 248 | + if target != 0 { |
| 249 | + continue |
| 250 | + } |
| 251 | + |
| 252 | + for underAllocatedCell, underAllocatedFlow := range underAllocated { |
| 253 | + allocation := flowPerVtgateCell * underAllocatedFlow / unbalancedFlow |
| 254 | + a.Inflows[underAllocatedCell] += allocation |
| 255 | + a.Outflows[vtgateCell][underAllocatedCell] += allocation |
| 256 | + } |
| 257 | + } |
| 258 | + |
| 259 | + // Recompute underallocated after these flows were assigned |
| 260 | + unbalancedFlow = 0 |
| 261 | + underAllocated = make(map[string]int) |
| 262 | + for cell, allocation := range a.Target { |
| 263 | + if a.Inflows[cell] < allocation { |
| 264 | + underAllocated[cell] = allocation - a.Inflows[cell] |
| 265 | + unbalancedFlow += underAllocated[cell] |
| 266 | + } |
| 267 | + } |
| 268 | + } |
| 269 | + |
| 270 | + // |
| 271 | + // Third pass: Shift remaining imbalance if any cell is over/under allocated after |
| 272 | + // assigning local cell traffic and distributing load from cells without tablets. |
| 273 | + // |
| 274 | + if /* fudge for integer arithmetic */ unbalancedFlow > 10 { |
| 275 | + |
| 276 | + // cells which are overallocated |
| 277 | + overAllocated := make(map[string]int) |
| 278 | + for cell, allocation := range a.Target { |
| 279 | + if a.Inflows[cell] > allocation { |
| 280 | + overAllocated[cell] = a.Inflows[cell] - allocation |
| 281 | + } |
| 282 | + } |
| 283 | + |
| 284 | + // fmt.Printf("outflows %v over %v under %v\n", a.Outflows, overAllocated, underAllocated) |
| 285 | + |
| 286 | + // |
| 287 | + // For each overallocated cell, proportionally shift flow from targets that are overallocated |
| 288 | + // to targets that are underallocated. |
| 289 | + // |
| 290 | + // Note this is an O(N^3) loop, but only over the cells which need adjustment. |
| 291 | + // |
| 292 | + for _, vtgateCell := range b.vtGateCells { |
| 293 | + for underAllocatedCell, underAllocatedFlow := range underAllocated { |
| 294 | + for overAllocatedCell, overAllocatedFlow := range overAllocated { |
| 295 | + |
| 296 | + currentFlow := a.Outflows[vtgateCell][overAllocatedCell] |
| 297 | + if currentFlow == 0 { |
| 298 | + continue |
| 299 | + } |
| 300 | + |
| 301 | + // Shift a proportional fraction of the amount that the cell is currently allocated weighted |
| 302 | + // by the fraction that this vtgate cell is already sending to the overallocated cell, and the |
| 303 | + // fraction that the new target is underallocated |
| 304 | + // |
| 305 | + // Note that the operator order matters -- multiplications need to occur before divisions |
| 306 | + // to avoid truncating the integer values. |
| 307 | + shiftFlow := overAllocatedFlow * currentFlow * underAllocatedFlow / a.Inflows[overAllocatedCell] / unbalancedFlow |
| 308 | + |
| 309 | + //fmt.Printf("shift %d %s %s -> %s (over %d current %d in %d under %d unbalanced %d) \n", shiftFlow, vtgateCell, overAllocatedCell, underAllocatedCell, |
| 310 | + // overAllocatedFlow, currentFlow, a.Inflows[overAllocatedCell], underAllocatedFlow, unbalancedFlow) |
| 311 | + |
| 312 | + a.Outflows[vtgateCell][overAllocatedCell] -= shiftFlow |
| 313 | + a.Inflows[overAllocatedCell] -= shiftFlow |
| 314 | + |
| 315 | + a.Inflows[underAllocatedCell] += shiftFlow |
| 316 | + a.Outflows[vtgateCell][underAllocatedCell] += shiftFlow |
| 317 | + } |
| 318 | + } |
| 319 | + } |
| 320 | + } |
| 321 | + |
| 322 | + // |
| 323 | + // Finally, once the cell flows are all adjusted, figure out the local allocation to each |
| 324 | + // tablet in the target cells |
| 325 | + // |
| 326 | + outflow := a.Outflows[b.localCell] |
| 327 | + for _, tablet := range allTablets { |
| 328 | + cell := tablet.Tablet.Alias.Cell |
| 329 | + flow := outflow[cell] |
| 330 | + if flow > 0 { |
| 331 | + a.Allocation[tablet.Tablet.Alias.Uid] = flow * flowPerTablet / a.Target[cell] |
| 332 | + a.TotalAllocation += flow * flowPerTablet / a.Target[cell] |
| 333 | + } else { |
| 334 | + a.Unallocated[tablet.Tablet.Alias.Uid] = struct{}{} |
| 335 | + } |
| 336 | + } |
| 337 | + |
| 338 | + return &a |
| 339 | +} |
| 340 | + |
| 341 | +// getAllocation builds the allocation map if needed and returns a copy of the map |
| 342 | +func (b *tabletBalancer) getAllocation(target *querypb.Target, tablets []*discovery.TabletHealth) (map[uint32]int, int) { |
| 343 | + b.mu.Lock() |
| 344 | + defer b.mu.Unlock() |
| 345 | + |
| 346 | + allocation, exists := b.allocations[discovery.KeyFromTarget(target)] |
| 347 | + if exists && (len(allocation.Allocation)+len(allocation.Unallocated)) == len(tablets) { |
| 348 | + mismatch := false |
| 349 | + for _, tablet := range tablets { |
| 350 | + if _, ok := allocation.Allocation[tablet.Tablet.Alias.Uid]; !ok { |
| 351 | + if _, ok := allocation.Unallocated[tablet.Tablet.Alias.Uid]; !ok { |
| 352 | + mismatch = true |
| 353 | + break |
| 354 | + } |
| 355 | + } |
| 356 | + } |
| 357 | + if !mismatch { |
| 358 | + // No change in tablets for this target. Return computed allocation |
| 359 | + return allocation.Allocation, allocation.TotalAllocation |
| 360 | + } |
| 361 | + } |
| 362 | + |
| 363 | + allocation = b.allocateFlows(tablets) |
| 364 | + b.allocations[discovery.KeyFromTarget(target)] = allocation |
| 365 | + |
| 366 | + return allocation.Allocation, allocation.TotalAllocation |
| 367 | +} |
0 commit comments