diff options
author | George Wangensteen <george.wangensteen@mongodb.com> | 2022-10-19 18:23:57 +0000 |
---|---|---|
committer | Evergreen Agent <no-reply@evergreen.mongodb.com> | 2022-10-19 19:49:08 +0000 |
commit | f84dd22211da9ed5865e5ea11f7a189e95e07f20 (patch) | |
tree | 61837fc48317d44e23ad9b45745a406aeab59250 /src/mongo/executor/async_rpc.h | |
parent | dc00665b514233ebbe4240da31bddba1dacb5298 (diff) | |
download | mongo-f84dd22211da9ed5865e5ea11f7a189e95e07f20.tar.gz |
SERVER-68650 Improve async RPC API naming conventions and fix use of retry policy
Diffstat (limited to 'src/mongo/executor/async_rpc.h')
-rw-r--r-- | src/mongo/executor/async_rpc.h | 258 |
1 files changed, 258 insertions, 0 deletions
diff --git a/src/mongo/executor/async_rpc.h b/src/mongo/executor/async_rpc.h new file mode 100644 index 00000000000..7eff757578c --- /dev/null +++ b/src/mongo/executor/async_rpc.h @@ -0,0 +1,258 @@ +/** + * Copyright (C) 2022-present MongoDB, Inc. + * + * This program is free software: you can redistribute it and/or modify + * it under the terms of the Server Side Public License, version 1, + * as published by MongoDB, Inc. + * + * This program is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the + * Server Side Public License for more details. + * + * You should have received a copy of the Server Side Public License + * along with this program. If not, see + * <http://www.mongodb.com/licensing/server-side-public-license>. + * + * As a special exception, the copyright holders give permission to link the + * code of portions of this program with the OpenSSL library under certain + * conditions as described in each individual source file and distribute + * linked combinations including the program with the OpenSSL library. You + * must comply with the Server Side Public License in all respects for + * all of the code used other than as permitted herein. If you modify file(s) + * with this exception, you may extend this exception to your version of the + * file(s), but you are not obligated to do so. If you do not wish to do so, + * delete this exception statement from your version. If you delete this + * exception statement from all source files in the program, then also delete + * it in the license file. + */ + +#pragma once + +#include <memory> + +#include "mongo/bson/bsonobj.h" +#include "mongo/db/operation_context.h" +#include "mongo/executor/async_rpc_error_info.h" +#include "mongo/executor/async_rpc_retry_policy.h" +#include "mongo/executor/async_rpc_targeter.h" +#include "mongo/executor/remote_command_response.h" +#include "mongo/executor/task_executor.h" +#include "mongo/logv2/log.h" +#include "mongo/rpc/get_status_from_command_result.h" +#include "mongo/util/assert_util.h" +#include "mongo/util/cancellation.h" +#include "mongo/util/future.h" +#include "mongo/util/future_util.h" +#include "mongo/util/net/hostandport.h" +#include "mongo/util/time_support.h" + +#define MONGO_LOGV2_DEFAULT_COMPONENT mongo::logv2::LogComponent::kExecutor + +/** + * This header provides an API of `sendCommand(...)` functions that can be used to asynchronously + * invoke well-typed commands on a remote node. Each takes an IDL-defined or similarly specified + * command type as argument, and invokes it on a remote node, internally handling targeting + * the remote node, cancellation, and performing retries as needed according to rules provided by + * arguments. Each function returns a future containing the response to the command, parsed into the + * response-type provided. See the function comments below for details. + */ +namespace mongo::async_rpc { +using executor::TaskExecutor; + +/** + * The response type used by `sendCommand(...)` functions, containing the typed response to the + * command as well as the host it was run on + */ +template <typename CommandReplyType> +struct AsyncRPCResponse { + CommandReplyType response; + HostAndPort targetUsed; +}; + +/** + * The response type used by `sendCommand(...)` functions if the return type of the command is + * 'void'. + */ +template <> +struct AsyncRPCResponse<void> { + HostAndPort targetUsed; +}; + +/** + * Details used internally by the API. Users of the API can skip the code in this namespace + * and proceed to the `sendCommand(...)` functions below. + */ +namespace detail { +struct AsyncRPCInternalResponse { + BSONObj response; + HostAndPort targetUsed; +}; + +/** + * The AsyncRPCRunner class stores the implementation details used by the free function + * async_rpc::sendCommand defined below. It takes a command and runs it on the provided + * HostAndPort/database name asynchronously, using the given executor. It keeps the executor alive + * for the duration of the command's execution; to cancel the command's execution, use the provided + * cancellation token. + * + * This is *not* part of the public API, and is deliberately in the detail namespace. Use the + * async_rpc::sendCommand free-function/public API below instead, which contains + * additional functionality and type checking. + */ +class AsyncRPCRunner { +public: + virtual ~AsyncRPCRunner() = default; + virtual ExecutorFuture<AsyncRPCInternalResponse> _sendCommand( + StringData dbName, + BSONObj cmdBSON, + Targeter* targeter, + OperationContext* opCtx, + std::shared_ptr<TaskExecutor> exec, + CancellationToken token) = 0; + static AsyncRPCRunner* get(ServiceContext* serviceContext); + static void set(ServiceContext* serviceContext, std::unique_ptr<AsyncRPCRunner> theRunner); +}; + +/** + * Returns a RemoteCommandExecutionError with ErrorExtraInfo populated to contain + * details about any error, local or remote, contained in `r`. + */ +inline Status makeErrorIfNeeded(TaskExecutor::ResponseOnAnyStatus r) { + if (r.status.isOK() && getStatusFromCommandResult(r.data).isOK() && + getWriteConcernStatusFromCommandResult(r.data).isOK() && + getFirstWriteErrorStatusFromCommandResult(r.data).isOK()) { + return Status::OK(); + } + return {AsyncRPCErrorInfo(r), "Remote command execution failed"}; +} + +/** + * Adaptor that allows a RetryPolicy to be used with AsyncTry::withBackoffBetweenIterations. + */ +struct RetryDelayAsBackoff { + RetryDelayAsBackoff(RetryPolicy* policy) : _policy{policy} {} + Milliseconds nextSleep() { + return _policy->getNextRetryDelay(); + } + RetryPolicy* _policy; +}; + +template <typename CommandType> +ExecutorFuture<AsyncRPCResponse<typename CommandType::Reply>> sendCommandWithRunner( + CommandType cmd, + detail::AsyncRPCRunner* runner, + OperationContext* opCtx, + std::unique_ptr<Targeter> targeter, + std::shared_ptr<executor::TaskExecutor> exec, + CancellationToken token, + std::shared_ptr<RetryPolicy> retryPolicy = std::make_shared<NeverRetryPolicy>()) { + using ReplyType = AsyncRPCResponse<typename CommandType::Reply>; + auto tryBody = [=, targeter = std::move(targeter)] { + // Execute the command after extracting the db name and bson from the CommandType. + // Wrapping this function allows us to separate the CommandType parsing logic from the + // implementation details of executing the remote command asynchronously. + return runner->_sendCommand( + cmd.getDbName().db(), cmd.toBSON({}), targeter.get(), opCtx, exec, token); + }; + auto resFuture = AsyncTry<decltype(tryBody)>(std::move(tryBody)) + .until([token, retryPolicy, cmd]( + StatusWith<detail::AsyncRPCInternalResponse> swResponse) { + bool shouldStopRetry = token.isCanceled() || + !retryPolicy->recordAndEvaluateRetry(swResponse.getStatus()); + return shouldStopRetry; + }) + .withBackoffBetweenIterations(RetryDelayAsBackoff(retryPolicy.get())) + .on(exec, CancellationToken::uncancelable()); + + return std::move(resFuture) + .then([](detail::AsyncRPCInternalResponse r) -> ReplyType { + // TODO SERVER-67661: Make IDL reply types have string representation for logging + auto res = CommandType::Reply::parseSharingOwnership(IDLParserContext("AsyncRPCRunner"), + r.response); + + return {res, r.targetUsed}; + }) + .unsafeToInlineFuture() + .onError( + // We go inline here to intercept executor-shutdown errors and re-write them + // so that the API always returns RemoteCommandExecutionError. + [](Status s) -> StatusWith<ReplyType> { + if (s.code() == ErrorCodes::RemoteCommandExecutionError) { + return s; + } + // The API implementation guarantees that all errors are provided as + // RemoteCommandExecutionError, so if we've reached this code, it means that the API + // internals were unable to run due to executor shutdown. Today, the only guarantee + // we can make about an executor-shutdown error is that it is in the cancellation + // category. We dassert that this is the case to make it easy to find errors in the + // API implementation's error-handling while still ensuring that we always return + // the correct error code in production. + dassert(ErrorCodes::isA<ErrorCategory::CancellationError>(s.code())); + return Status{AsyncRPCErrorInfo(s), + "Remote command execution failed due to executor shutdown"}; + }) + .thenRunOn(exec); +} +} // namespace detail + +/** + * Execute the command asynchronously on the given target with the provided executor. + * + * The command type specified must have a `toBSON(BSONObj)` member function that transforms the + * command into BSON suitable for sending over-the-wire. It also must have a nested `Reply` type + * with a static `Reply parseSharingOwnership(BSONObj)` member function that parses BSON recieved in + * response to the command into the `Reply` type. Note all IDL-defined command types meet these + * requirements. Returns an ExecutorFuture with the reply from the IDL command. + * + * If there is any error, local or remote, while executing the command, the future is set with + * ErrorCodes::RemoteCommandExecutionError. This is the only error returned by the API. Additional + * information about the error, such as its provenance, code, whether it was a command-error or + * write{concern}error, etc, is available in the ExtraInfo object attached to the error. See + * async_rpc_error_info.h for details. + * + * Cancelling the source associated with the provided token will cancel any outstanding RPC work. + * The `targeter` and optional `retryPolicy` arguments allow you to specify how to target the + * command and when to retry it; see the class comments for those arguments for details. The default + * retry policy is to not do any retries. + * + * The `opCtx` argument is used by NetworkEgressMetadataHooks to append operation-specific metadata + * (i.e. potential cluster-time ticking). (TODO: SERVER-70191) Additionally, if the `opCtx` has an + * attached baton, the baton may be used to run portions of the commands targeting logic and/or + * retry logic, as well as process the network response. + */ +template <typename CommandType> +ExecutorFuture<AsyncRPCResponse<typename CommandType::Reply>> sendCommand( + CommandType cmd, + OperationContext* opCtx, + std::unique_ptr<Targeter> targeter, + std::shared_ptr<executor::TaskExecutor> exec, + CancellationToken token, + std::shared_ptr<RetryPolicy> retryPolicy = std::make_shared<NeverRetryPolicy>()) { + auto runner = detail::AsyncRPCRunner::get(opCtx->getServiceContext()); + return detail::sendCommandWithRunner( + cmd, runner, opCtx, std::move(targeter), exec, token, retryPolicy); +} + +/** + * This function operates the same to `sendCommand` above, but without taking an operation context. + * It therefore does not append operation/client specific metadata via NetworkEgressMetadataHooks, + * and all work runs on the provided executor. + */ +template <typename CommandType> +ExecutorFuture<AsyncRPCResponse<typename CommandType::Reply>> sendCommand( + CommandType cmd, + ServiceContext* const svcCtx, + std::unique_ptr<Targeter> targeter, + std::shared_ptr<executor::TaskExecutor> exec, + CancellationToken token, + std::shared_ptr<RetryPolicy> retryPolicy = std::make_shared<NeverRetryPolicy>()) { + // Execute the command after extracting the db name and bson from the CommandType. + // Wrapping this function allows us to separate the CommandType parsing logic from the + // implementation details of executing the remote command asynchronously. + auto runner = detail::AsyncRPCRunner::get(svcCtx); + return detail::sendCommandWithRunner( + cmd, runner, nullptr, std::move(targeter), exec, token, retryPolicy); +} +} // namespace mongo::async_rpc +#undef MONGO_LOGV2_DEFAULT_COMPONENT |