khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 1 | /* |
| 2 | * |
| 3 | * Copyright 2017 gRPC authors. |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | * |
| 17 | */ |
| 18 | |
| 19 | // Package balancer defines APIs for load balancing in gRPC. |
| 20 | // All APIs in this package are experimental. |
| 21 | package balancer |
| 22 | |
| 23 | import ( |
| 24 | "context" |
| 25 | "encoding/json" |
| 26 | "errors" |
| 27 | "net" |
| 28 | "strings" |
| 29 | |
| 30 | "google.golang.org/grpc/connectivity" |
| 31 | "google.golang.org/grpc/credentials" |
| 32 | "google.golang.org/grpc/internal" |
| 33 | "google.golang.org/grpc/metadata" |
| 34 | "google.golang.org/grpc/resolver" |
| 35 | "google.golang.org/grpc/serviceconfig" |
| 36 | ) |
| 37 | |
| 38 | var ( |
| 39 | // m is a map from name to balancer builder. |
| 40 | m = make(map[string]Builder) |
| 41 | ) |
| 42 | |
| 43 | // Register registers the balancer builder to the balancer map. b.Name |
| 44 | // (lowercased) will be used as the name registered with this builder. If the |
| 45 | // Builder implements ConfigParser, ParseConfig will be called when new service |
| 46 | // configs are received by the resolver, and the result will be provided to the |
| 47 | // Balancer in UpdateClientConnState. |
| 48 | // |
| 49 | // NOTE: this function must only be called during initialization time (i.e. in |
| 50 | // an init() function), and is not thread-safe. If multiple Balancers are |
| 51 | // registered with the same name, the one registered last will take effect. |
| 52 | func Register(b Builder) { |
| 53 | m[strings.ToLower(b.Name())] = b |
| 54 | } |
| 55 | |
| 56 | // unregisterForTesting deletes the balancer with the given name from the |
| 57 | // balancer map. |
| 58 | // |
| 59 | // This function is not thread-safe. |
| 60 | func unregisterForTesting(name string) { |
| 61 | delete(m, name) |
| 62 | } |
| 63 | |
| 64 | func init() { |
| 65 | internal.BalancerUnregister = unregisterForTesting |
| 66 | } |
| 67 | |
| 68 | // Get returns the resolver builder registered with the given name. |
| 69 | // Note that the compare is done in a case-insensitive fashion. |
| 70 | // If no builder is register with the name, nil will be returned. |
| 71 | func Get(name string) Builder { |
| 72 | if b, ok := m[strings.ToLower(name)]; ok { |
| 73 | return b |
| 74 | } |
| 75 | return nil |
| 76 | } |
| 77 | |
| 78 | // A SubConn represents a single connection to a gRPC backend service. |
| 79 | // |
| 80 | // Each SubConn contains a list of addresses. |
| 81 | // |
| 82 | // All SubConns start in IDLE, and will not try to connect. To trigger the |
| 83 | // connecting, Balancers must call Connect. If a connection re-enters IDLE, |
| 84 | // Balancers must call Connect again to trigger a new connection attempt. |
| 85 | // |
| 86 | // gRPC will try to connect to the addresses in sequence, and stop trying the |
| 87 | // remainder once the first connection is successful. If an attempt to connect |
| 88 | // to all addresses encounters an error, the SubConn will enter |
| 89 | // TRANSIENT_FAILURE for a backoff period, and then transition to IDLE. |
| 90 | // |
| 91 | // Once established, if a connection is lost, the SubConn will transition |
| 92 | // directly to IDLE. |
| 93 | // |
| 94 | // This interface is to be implemented by gRPC. Users should not need their own |
| 95 | // implementation of this interface. For situations like testing, any |
| 96 | // implementations should embed this interface. This allows gRPC to add new |
| 97 | // methods to this interface. |
| 98 | type SubConn interface { |
| 99 | // UpdateAddresses updates the addresses used in this SubConn. |
| 100 | // gRPC checks if currently-connected address is still in the new list. |
| 101 | // If it's in the list, the connection will be kept. |
| 102 | // If it's not in the list, the connection will gracefully closed, and |
| 103 | // a new connection will be created. |
| 104 | // |
| 105 | // This will trigger a state transition for the SubConn. |
| 106 | // |
| 107 | // Deprecated: This method is now part of the ClientConn interface and will |
| 108 | // eventually be removed from here. |
| 109 | UpdateAddresses([]resolver.Address) |
| 110 | // Connect starts the connecting for this SubConn. |
| 111 | Connect() |
| 112 | } |
| 113 | |
| 114 | // NewSubConnOptions contains options to create new SubConn. |
| 115 | type NewSubConnOptions struct { |
| 116 | // CredsBundle is the credentials bundle that will be used in the created |
| 117 | // SubConn. If it's nil, the original creds from grpc DialOptions will be |
| 118 | // used. |
| 119 | // |
| 120 | // Deprecated: Use the Attributes field in resolver.Address to pass |
| 121 | // arbitrary data to the credential handshaker. |
| 122 | CredsBundle credentials.Bundle |
| 123 | // HealthCheckEnabled indicates whether health check service should be |
| 124 | // enabled on this SubConn |
| 125 | HealthCheckEnabled bool |
| 126 | } |
| 127 | |
| 128 | // State contains the balancer's state relevant to the gRPC ClientConn. |
| 129 | type State struct { |
| 130 | // State contains the connectivity state of the balancer, which is used to |
| 131 | // determine the state of the ClientConn. |
| 132 | ConnectivityState connectivity.State |
| 133 | // Picker is used to choose connections (SubConns) for RPCs. |
| 134 | Picker Picker |
| 135 | } |
| 136 | |
| 137 | // ClientConn represents a gRPC ClientConn. |
| 138 | // |
| 139 | // This interface is to be implemented by gRPC. Users should not need a |
| 140 | // brand new implementation of this interface. For the situations like |
| 141 | // testing, the new implementation should embed this interface. This allows |
| 142 | // gRPC to add new methods to this interface. |
| 143 | type ClientConn interface { |
| 144 | // NewSubConn is called by balancer to create a new SubConn. |
| 145 | // It doesn't block and wait for the connections to be established. |
| 146 | // Behaviors of the SubConn can be controlled by options. |
| 147 | NewSubConn([]resolver.Address, NewSubConnOptions) (SubConn, error) |
| 148 | // RemoveSubConn removes the SubConn from ClientConn. |
| 149 | // The SubConn will be shutdown. |
| 150 | RemoveSubConn(SubConn) |
| 151 | // UpdateAddresses updates the addresses used in the passed in SubConn. |
| 152 | // gRPC checks if the currently connected address is still in the new list. |
| 153 | // If so, the connection will be kept. Else, the connection will be |
| 154 | // gracefully closed, and a new connection will be created. |
| 155 | // |
| 156 | // This will trigger a state transition for the SubConn. |
| 157 | UpdateAddresses(SubConn, []resolver.Address) |
| 158 | |
| 159 | // UpdateState notifies gRPC that the balancer's internal state has |
| 160 | // changed. |
| 161 | // |
| 162 | // gRPC will update the connectivity state of the ClientConn, and will call |
| 163 | // Pick on the new Picker to pick new SubConns. |
| 164 | UpdateState(State) |
| 165 | |
| 166 | // ResolveNow is called by balancer to notify gRPC to do a name resolving. |
| 167 | ResolveNow(resolver.ResolveNowOptions) |
| 168 | |
| 169 | // Target returns the dial target for this ClientConn. |
| 170 | // |
| 171 | // Deprecated: Use the Target field in the BuildOptions instead. |
| 172 | Target() string |
| 173 | } |
| 174 | |
| 175 | // BuildOptions contains additional information for Build. |
| 176 | type BuildOptions struct { |
khenaidoo | 257f319 | 2021-12-15 16:46:37 -0500 | [diff] [blame] | 177 | // DialCreds is the transport credentials to use when communicating with a |
| 178 | // remote load balancer server. Balancer implementations which do not |
| 179 | // communicate with a remote load balancer server can ignore this field. |
khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 180 | DialCreds credentials.TransportCredentials |
khenaidoo | 257f319 | 2021-12-15 16:46:37 -0500 | [diff] [blame] | 181 | // CredsBundle is the credentials bundle to use when communicating with a |
| 182 | // remote load balancer server. Balancer implementations which do not |
| 183 | // communicate with a remote load balancer server can ignore this field. |
khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 184 | CredsBundle credentials.Bundle |
khenaidoo | 257f319 | 2021-12-15 16:46:37 -0500 | [diff] [blame] | 185 | // Dialer is the custom dialer to use when communicating with a remote load |
| 186 | // balancer server. Balancer implementations which do not communicate with a |
| 187 | // remote load balancer server can ignore this field. |
khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 188 | Dialer func(context.Context, string) (net.Conn, error) |
khenaidoo | 257f319 | 2021-12-15 16:46:37 -0500 | [diff] [blame] | 189 | // Authority is the server name to use as part of the authentication |
| 190 | // handshake when communicating with a remote load balancer server. Balancer |
| 191 | // implementations which do not communicate with a remote load balancer |
| 192 | // server can ignore this field. |
| 193 | Authority string |
| 194 | // ChannelzParentID is the parent ClientConn's channelz ID. |
khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 195 | ChannelzParentID int64 |
| 196 | // CustomUserAgent is the custom user agent set on the parent ClientConn. |
| 197 | // The balancer should set the same custom user agent if it creates a |
| 198 | // ClientConn. |
| 199 | CustomUserAgent string |
khenaidoo | 257f319 | 2021-12-15 16:46:37 -0500 | [diff] [blame] | 200 | // Target contains the parsed address info of the dial target. It is the |
| 201 | // same resolver.Target as passed to the resolver. See the documentation for |
| 202 | // the resolver.Target type for details about what it contains. |
khenaidoo | 5fc5cea | 2021-08-11 17:39:16 -0400 | [diff] [blame] | 203 | Target resolver.Target |
| 204 | } |
| 205 | |
| 206 | // Builder creates a balancer. |
| 207 | type Builder interface { |
| 208 | // Build creates a new balancer with the ClientConn. |
| 209 | Build(cc ClientConn, opts BuildOptions) Balancer |
| 210 | // Name returns the name of balancers built by this builder. |
| 211 | // It will be used to pick balancers (for example in service config). |
| 212 | Name() string |
| 213 | } |
| 214 | |
| 215 | // ConfigParser parses load balancer configs. |
| 216 | type ConfigParser interface { |
| 217 | // ParseConfig parses the JSON load balancer config provided into an |
| 218 | // internal form or returns an error if the config is invalid. For future |
| 219 | // compatibility reasons, unknown fields in the config should be ignored. |
| 220 | ParseConfig(LoadBalancingConfigJSON json.RawMessage) (serviceconfig.LoadBalancingConfig, error) |
| 221 | } |
| 222 | |
| 223 | // PickInfo contains additional information for the Pick operation. |
| 224 | type PickInfo struct { |
| 225 | // FullMethodName is the method name that NewClientStream() is called |
| 226 | // with. The canonical format is /service/Method. |
| 227 | FullMethodName string |
| 228 | // Ctx is the RPC's context, and may contain relevant RPC-level information |
| 229 | // like the outgoing header metadata. |
| 230 | Ctx context.Context |
| 231 | } |
| 232 | |
| 233 | // DoneInfo contains additional information for done. |
| 234 | type DoneInfo struct { |
| 235 | // Err is the rpc error the RPC finished with. It could be nil. |
| 236 | Err error |
| 237 | // Trailer contains the metadata from the RPC's trailer, if present. |
| 238 | Trailer metadata.MD |
| 239 | // BytesSent indicates if any bytes have been sent to the server. |
| 240 | BytesSent bool |
| 241 | // BytesReceived indicates if any byte has been received from the server. |
| 242 | BytesReceived bool |
| 243 | // ServerLoad is the load received from server. It's usually sent as part of |
| 244 | // trailing metadata. |
| 245 | // |
| 246 | // The only supported type now is *orca_v1.LoadReport. |
| 247 | ServerLoad interface{} |
| 248 | } |
| 249 | |
| 250 | var ( |
| 251 | // ErrNoSubConnAvailable indicates no SubConn is available for pick(). |
| 252 | // gRPC will block the RPC until a new picker is available via UpdateState(). |
| 253 | ErrNoSubConnAvailable = errors.New("no SubConn is available") |
| 254 | // ErrTransientFailure indicates all SubConns are in TransientFailure. |
| 255 | // WaitForReady RPCs will block, non-WaitForReady RPCs will fail. |
| 256 | // |
| 257 | // Deprecated: return an appropriate error based on the last resolution or |
| 258 | // connection attempt instead. The behavior is the same for any non-gRPC |
| 259 | // status error. |
| 260 | ErrTransientFailure = errors.New("all SubConns are in TransientFailure") |
| 261 | ) |
| 262 | |
| 263 | // PickResult contains information related to a connection chosen for an RPC. |
| 264 | type PickResult struct { |
| 265 | // SubConn is the connection to use for this pick, if its state is Ready. |
| 266 | // If the state is not Ready, gRPC will block the RPC until a new Picker is |
| 267 | // provided by the balancer (using ClientConn.UpdateState). The SubConn |
| 268 | // must be one returned by ClientConn.NewSubConn. |
| 269 | SubConn SubConn |
| 270 | |
| 271 | // Done is called when the RPC is completed. If the SubConn is not ready, |
| 272 | // this will be called with a nil parameter. If the SubConn is not a valid |
| 273 | // type, Done may not be called. May be nil if the balancer does not wish |
| 274 | // to be notified when the RPC completes. |
| 275 | Done func(DoneInfo) |
| 276 | } |
| 277 | |
| 278 | // TransientFailureError returns e. It exists for backward compatibility and |
| 279 | // will be deleted soon. |
| 280 | // |
| 281 | // Deprecated: no longer necessary, picker errors are treated this way by |
| 282 | // default. |
| 283 | func TransientFailureError(e error) error { return e } |
| 284 | |
| 285 | // Picker is used by gRPC to pick a SubConn to send an RPC. |
| 286 | // Balancer is expected to generate a new picker from its snapshot every time its |
| 287 | // internal state has changed. |
| 288 | // |
| 289 | // The pickers used by gRPC can be updated by ClientConn.UpdateState(). |
| 290 | type Picker interface { |
| 291 | // Pick returns the connection to use for this RPC and related information. |
| 292 | // |
| 293 | // Pick should not block. If the balancer needs to do I/O or any blocking |
| 294 | // or time-consuming work to service this call, it should return |
| 295 | // ErrNoSubConnAvailable, and the Pick call will be repeated by gRPC when |
| 296 | // the Picker is updated (using ClientConn.UpdateState). |
| 297 | // |
| 298 | // If an error is returned: |
| 299 | // |
| 300 | // - If the error is ErrNoSubConnAvailable, gRPC will block until a new |
| 301 | // Picker is provided by the balancer (using ClientConn.UpdateState). |
| 302 | // |
| 303 | // - If the error is a status error (implemented by the grpc/status |
| 304 | // package), gRPC will terminate the RPC with the code and message |
| 305 | // provided. |
| 306 | // |
| 307 | // - For all other errors, wait for ready RPCs will wait, but non-wait for |
| 308 | // ready RPCs will be terminated with this error's Error() string and |
| 309 | // status code Unavailable. |
| 310 | Pick(info PickInfo) (PickResult, error) |
| 311 | } |
| 312 | |
| 313 | // Balancer takes input from gRPC, manages SubConns, and collects and aggregates |
| 314 | // the connectivity states. |
| 315 | // |
| 316 | // It also generates and updates the Picker used by gRPC to pick SubConns for RPCs. |
| 317 | // |
| 318 | // UpdateClientConnState, ResolverError, UpdateSubConnState, and Close are |
| 319 | // guaranteed to be called synchronously from the same goroutine. There's no |
| 320 | // guarantee on picker.Pick, it may be called anytime. |
| 321 | type Balancer interface { |
| 322 | // UpdateClientConnState is called by gRPC when the state of the ClientConn |
| 323 | // changes. If the error returned is ErrBadResolverState, the ClientConn |
| 324 | // will begin calling ResolveNow on the active name resolver with |
| 325 | // exponential backoff until a subsequent call to UpdateClientConnState |
| 326 | // returns a nil error. Any other errors are currently ignored. |
| 327 | UpdateClientConnState(ClientConnState) error |
| 328 | // ResolverError is called by gRPC when the name resolver reports an error. |
| 329 | ResolverError(error) |
| 330 | // UpdateSubConnState is called by gRPC when the state of a SubConn |
| 331 | // changes. |
| 332 | UpdateSubConnState(SubConn, SubConnState) |
| 333 | // Close closes the balancer. The balancer is not required to call |
| 334 | // ClientConn.RemoveSubConn for its existing SubConns. |
| 335 | Close() |
| 336 | } |
| 337 | |
| 338 | // ExitIdler is an optional interface for balancers to implement. If |
| 339 | // implemented, ExitIdle will be called when ClientConn.Connect is called, if |
| 340 | // the ClientConn is idle. If unimplemented, ClientConn.Connect will cause |
| 341 | // all SubConns to connect. |
| 342 | // |
| 343 | // Notice: it will be required for all balancers to implement this in a future |
| 344 | // release. |
| 345 | type ExitIdler interface { |
| 346 | // ExitIdle instructs the LB policy to reconnect to backends / exit the |
| 347 | // IDLE state, if appropriate and possible. Note that SubConns that enter |
| 348 | // the IDLE state will not reconnect until SubConn.Connect is called. |
| 349 | ExitIdle() |
| 350 | } |
| 351 | |
| 352 | // SubConnState describes the state of a SubConn. |
| 353 | type SubConnState struct { |
| 354 | // ConnectivityState is the connectivity state of the SubConn. |
| 355 | ConnectivityState connectivity.State |
| 356 | // ConnectionError is set if the ConnectivityState is TransientFailure, |
| 357 | // describing the reason the SubConn failed. Otherwise, it is nil. |
| 358 | ConnectionError error |
| 359 | } |
| 360 | |
| 361 | // ClientConnState describes the state of a ClientConn relevant to the |
| 362 | // balancer. |
| 363 | type ClientConnState struct { |
| 364 | ResolverState resolver.State |
| 365 | // The parsed load balancing configuration returned by the builder's |
| 366 | // ParseConfig method, if implemented. |
| 367 | BalancerConfig serviceconfig.LoadBalancingConfig |
| 368 | } |
| 369 | |
| 370 | // ErrBadResolverState may be returned by UpdateClientConnState to indicate a |
| 371 | // problem with the provided name resolver data. |
| 372 | var ErrBadResolverState = errors.New("bad resolver state") |
| 373 | |
| 374 | // ConnectivityStateEvaluator takes the connectivity states of multiple SubConns |
| 375 | // and returns one aggregated connectivity state. |
| 376 | // |
| 377 | // It's not thread safe. |
| 378 | type ConnectivityStateEvaluator struct { |
| 379 | numReady uint64 // Number of addrConns in ready state. |
| 380 | numConnecting uint64 // Number of addrConns in connecting state. |
| 381 | numTransientFailure uint64 // Number of addrConns in transient failure state. |
| 382 | numIdle uint64 // Number of addrConns in idle state. |
| 383 | } |
| 384 | |
| 385 | // RecordTransition records state change happening in subConn and based on that |
| 386 | // it evaluates what aggregated state should be. |
| 387 | // |
| 388 | // - If at least one SubConn in Ready, the aggregated state is Ready; |
| 389 | // - Else if at least one SubConn in Connecting, the aggregated state is Connecting; |
| 390 | // - Else if at least one SubConn is TransientFailure, the aggregated state is Transient Failure; |
| 391 | // - Else if at least one SubConn is Idle, the aggregated state is Idle; |
| 392 | // - Else there are no subconns and the aggregated state is Transient Failure |
| 393 | // |
| 394 | // Shutdown is not considered. |
| 395 | func (cse *ConnectivityStateEvaluator) RecordTransition(oldState, newState connectivity.State) connectivity.State { |
| 396 | // Update counters. |
| 397 | for idx, state := range []connectivity.State{oldState, newState} { |
| 398 | updateVal := 2*uint64(idx) - 1 // -1 for oldState and +1 for new. |
| 399 | switch state { |
| 400 | case connectivity.Ready: |
| 401 | cse.numReady += updateVal |
| 402 | case connectivity.Connecting: |
| 403 | cse.numConnecting += updateVal |
| 404 | case connectivity.TransientFailure: |
| 405 | cse.numTransientFailure += updateVal |
| 406 | case connectivity.Idle: |
| 407 | cse.numIdle += updateVal |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | // Evaluate. |
| 412 | if cse.numReady > 0 { |
| 413 | return connectivity.Ready |
| 414 | } |
| 415 | if cse.numConnecting > 0 { |
| 416 | return connectivity.Connecting |
| 417 | } |
| 418 | if cse.numTransientFailure > 0 { |
| 419 | return connectivity.TransientFailure |
| 420 | } |
| 421 | if cse.numIdle > 0 { |
| 422 | return connectivity.Idle |
| 423 | } |
| 424 | return connectivity.TransientFailure |
| 425 | } |