hedged_request.hpp

Go to the documentation of this file.

#pragma once
 
/// @file userver/utils/hedged_request.hpp
/// @brief
/// Classes and functions for performing hedged requests.
///
/// To perform hedged request you need to define RequestStrategy - a class
/// similar to following ExampleStrategy:
///
/// class ExampleStrategy {
///   public:
///     /// Create request future.
///
///     /// Called at least once per hedged request. First time at the beginning
///     /// of a hedged request and then every HedgingSettings::hedging_delay
///     /// milliseconds if none of the previous requests are ready or
///     /// ProcessReply returned non-nullopt result. If ProcessReply returned
///     /// some interval of time, then additional request will be scheduled at
///     /// this interval of time.
///     /// @param attempt - increasing number for each try
///     std::optional<RequestType> Create(std::size_t attempt);
///
///     /// ProcessReply is called when some request has finished. Method should
///     /// evaluate request status and decide whether new attempt is required.
///     /// If new attempt is required, then method must return delay for next
///     /// attempt. If no other attempt is required, then method must return
///     /// std::nullopt. It is expected that successful result will be stored
///     /// somewhere internally and will be available with ExtractReply()
///     /// method.
///     std::optional<std::chrono::milliseconds> ProcessReply(RequestType&&);
///
///     std::optional<ReplyType> ExtractReply();
///
///     /// Method is called when system does not need request any more.
///     /// For example, if one of the requests has been finished successfully,
///     /// the system will finish all related subrequests.
///     /// It is recommended to make this call as fast as possible in order
///     /// not to block overall execution. For example, if you internally have
///     /// some future, it is recommended to call TryCancel() here and wait
///     /// for a cancellation in destructor, rather than call TryCancel() and
///     /// immediately wait.
///     void Finish(RequestType&&);
/// };
///
/// Then call any of these functions:
/// - HedgeRequest
/// - HedgeRequestsBulk
/// - HedgeRequestAsync
/// - HedgeRequestsBulkAsync
///
 
#include <chrono>
#include <functional>
#include <optional>
#include <queue>
#include <tuple>
#include <type_traits>
 
#include <userver/engine/task/cancel.hpp>
#include <userver/engine/wait_any.hpp>
#include <userver/utils/assert.hpp>
#include <userver/utils/async.hpp>
#include <userver/utils/datetime.hpp>
 
USERVER_NAMESPACE_BEGIN
 
namespace utils::hedging {
 
/// Class define hedging settings
struct HedgingSettings {
    /// Maximum requests to do
    std::size_t max_attempts{3};
    /// Delay between attempts
    std::chrono::milliseconds hedging_delay{7};
    /// Max time to wait for all requests
    std::chrono::milliseconds timeout_all{100};
};
 
template <typename RequestStrategy>
struct RequestTraits {
    using RequestType =
        typename std::invoke_result_t<decltype(&RequestStrategy::Create), RequestStrategy, int>::value_type;
    using ReplyType =
        typename std::invoke_result_t<decltype(&RequestStrategy::ExtractReply), RequestStrategy>::value_type;
};
 
namespace impl {
 
using Clock = utils::datetime::SteadyClock;
using TimePoint = Clock::time_point;
 
enum class Action { StartTry, Stop };
 
struct PlanEntry {
public:
    PlanEntry(TimePoint timepoint, std::size_t request_index, std::size_t attempt_id, Action action)
        : timepoint(timepoint), request_index(request_index), attempt_id(attempt_id), action(action) {}
 
    bool operator<(const PlanEntry& other) const noexcept { return tie() < other.tie(); }
    bool operator>(const PlanEntry& other) const noexcept { return tie() > other.tie(); }
    bool operator==(const PlanEntry& other) const noexcept { return tie() == other.tie(); }
    bool operator<=(const PlanEntry& other) const noexcept { return tie() <= other.tie(); }
    bool operator>=(const PlanEntry& other) const noexcept { return tie() >= other.tie(); }
    bool operator!=(const PlanEntry& other) const noexcept { return tie() != other.tie(); }
 
    TimePoint timepoint;
    std::size_t request_index{0};
    std::size_t attempt_id{0};
    Action action;
 
private:
    std::tuple<const TimePoint&, const size_t&, const size_t&, const Action&> tie() const noexcept {
        return std::tie(timepoint, request_index, attempt_id, action);
    }
};
 
/// This wrapper allows us to cancel subrequests without removing elements
/// from vector of requests
template <typename RequestStrategy>
struct SubrequestWrapper {
    using RequestType = typename RequestTraits<RequestStrategy>::RequestType;
 
    SubrequestWrapper() = default;
    SubrequestWrapper(SubrequestWrapper&&) noexcept = default;
    explicit SubrequestWrapper(std::optional<RequestType>&& request) : request(std::move(request)) {}
 
    engine::impl::ContextAccessor* TryGetContextAccessor() {
        if (!request) return nullptr;
        return request->TryGetContextAccessor();
    }
 
    std::optional<RequestType> request;
};
 
struct RequestState {
    std::vector<std::size_t> subrequest_indices;
    std::size_t attempts_made = 0;
    bool finished = false;
};
 
template <typename RequestStrategy>
struct Context {
    using RequestType = typename RequestTraits<RequestStrategy>::RequestType;
    using ReplyType = typename RequestTraits<RequestStrategy>::ReplyType;
 
    Context(std::vector<RequestStrategy> inputs, HedgingSettings settings)
        : inputs_(std::move(inputs)), settings(std::move(settings)) {
        const std::size_t size = this->inputs_.size();
        request_states_.resize(size);
    }
    Context(Context&&) noexcept = default;
 
    void Prepare(TimePoint start_time) {
        const auto request_count = GetRequestsCount();
        for (std::size_t request_id = 0; request_id < request_count; ++request_id) {
            plan_.emplace(start_time, request_id, 0, Action::StartTry);
        }
        plan_.emplace(start_time + settings.timeout_all, 0, 0, Action::Stop);
        subrequests_.reserve(settings.max_attempts * request_count);
    }
 
    std::optional<TimePoint> NextEventTime() const {
        if (plan_.empty()) return std::nullopt;
        return plan_.top().timepoint;
    }
    std::optional<PlanEntry> PopPlan() {
        if (plan_.empty()) return std::nullopt;
        auto ret = plan_.top();
        plan_.pop();
        return ret;
    }
    bool IsStop() const { return stop_; }
 
    void FinishAllSubrequests(std::size_t request_index) {
        auto& request_state = request_states_[request_index];
        request_state.finished = true;
        const auto& subrequest_indices = request_state.subrequest_indices;
        auto& strategy = inputs_[request_index];
        for (auto i : subrequest_indices) {
            auto& request = subrequests_[i].request;
            if (request) {
                strategy.Finish(std::move(*request));
                request.reset();
            }
        }
    }
 
    const HedgingSettings& GetSettings() const { return settings; }
 
    size_t GetRequestsCount() const { return inputs_.size(); }
 
    size_t GetRequestIdxBySubrequestIdx(size_t subrequest_idx) const {
        return input_by_subrequests_.at(subrequest_idx);
    }
 
    RequestStrategy& GetStrategy(size_t index) { return inputs_[index]; }
 
    auto& GetSubRequests() { return subrequests_; }
 
    std::vector<std::optional<ReplyType>> ExtractAllReplies() {
        std::vector<std::optional<ReplyType>> ret;
        ret.reserve(GetRequestsCount());
        for (auto&& strategy : inputs_) {
            ret.emplace_back(strategy.ExtractReply());
        }
        return ret;
    }
 
    /// @name Reactions on events with WaitAny*
    /// @{
    /// Called on elapsed timeout of WaitAny when next event is Stop some
    /// request
    void OnActionStop() {
        for (std::size_t i = 0; i < inputs_.size(); ++i) FinishAllSubrequests(i);
        stop_ = true;
    }
 
    /// Called on elapsed timeout of WaitAny when next event is Start retry of
    /// request with id equal @param request_index
    void OnActionStartTry(std::size_t request_index, std::size_t attempt_id, TimePoint now) {
        auto& request_state = request_states_[request_index];
        if (request_state.finished) {
            return;
        }
        auto& attempts_made = request_state.attempts_made;
        // We could have already launched attempt with this number, for example if
        // attempt number 2 failed with retryable code, we will add it to plan with
        // number 3. This way, there are now two planned events, both with number=3
        // - one from retry, one from hedging. We will execute the earliest one and
        // skip the second one.
        if (attempt_id < attempts_made) {
            return;
        }
 
        if (attempts_made >= settings.max_attempts) {
            return;
        }
        auto& strategy = inputs_[request_index];
        auto request_opt = strategy.Create(attempts_made);
        if (!request_opt) {
            request_state.finished = true;
            // User do not want to make another request (maybe some retry budget is
            // used)
            return;
        }
        const auto idx = subrequests_.size();
        subrequests_.emplace_back(std::move(request_opt));
        request_state.subrequest_indices.push_back(idx);
        input_by_subrequests_[idx] = request_index;
        attempts_made++;
        plan_.emplace(now + settings.hedging_delay, request_index, attempts_made, Action::StartTry);
    }
 
    /// Called on getting error in request with @param request_idx
    void OnRetriableReply(std::size_t request_idx, std::chrono::milliseconds retry_delay, TimePoint now) {
        const auto& request_state = request_states_[request_idx];
        if (request_state.finished) return;
        if (request_state.attempts_made >= settings.max_attempts) return;
 
        plan_.emplace(now + retry_delay, request_idx, request_state.attempts_made, Action::StartTry);
    }
 
    void OnNonRetriableReply(std::size_t request_idx) { FinishAllSubrequests(request_idx); }
    /// @}
 
private:
    /// user provided request strategies bulk
    std::vector<RequestStrategy> inputs_;
    HedgingSettings settings;
 
    /// Our plan of what we will do at what time
    std::priority_queue<PlanEntry, std::vector<PlanEntry>, std::greater<>> plan_{};
    std::vector<SubrequestWrapper<RequestStrategy>> subrequests_{};
    /// Store index of input by subrequest index
    std::unordered_map<std::size_t, std::size_t> input_by_subrequests_{};
    std::vector<RequestState> request_states_{};
    bool stop_{false};
};
 
}  // namespace impl
 
/// Future of hedged bulk request
template <typename RequestStrategy>
struct HedgedRequestBulkFuture {
    using RequestType = typename RequestTraits<RequestStrategy>::RequestType;
    using ReplyType = typename RequestTraits<RequestStrategy>::ReplyType;
 
    HedgedRequestBulkFuture(HedgedRequestBulkFuture&&) noexcept = default;
    ~HedgedRequestBulkFuture() { task_.SyncCancel(); }
 
    void Wait() { task_.Wait(); }
    std::vector<std::optional<ReplyType>> Get() { return task_.Get(); }
    engine::impl::ContextAccessor* TryGetContextAccessor() { return task_.TryGetContextAccessor(); }
 
private:
    template <typename RequestStrategy_>
    friend auto HedgeRequestsBulkAsync(std::vector<RequestStrategy_> inputs, HedgingSettings settings);
    using Task = engine::TaskWithResult<std::vector<std::optional<ReplyType>>>;
    HedgedRequestBulkFuture(Task&& task) : task_(std::move(task)) {}
    Task task_;
};
 
/// Future of hedged request
template <typename RequestStrategy>
struct HedgedRequestFuture {
    using RequestType = typename RequestTraits<RequestStrategy>::RequestType;
    using ReplyType = typename RequestTraits<RequestStrategy>::ReplyType;
 
    HedgedRequestFuture(HedgedRequestFuture&&) noexcept = default;
    ~HedgedRequestFuture() { task_.SyncCancel(); }
 
    void Wait() { task_.Wait(); }
    std::optional<ReplyType> Get() { return task_.Get(); }
    void IgnoreResult() {}
    engine::impl::ContextAccessor* TryGetContextAccessor() { return task_.TryGetContextAccessor(); }
 
private:
    template <typename RequestStrategy_>
    friend auto HedgeRequestAsync(RequestStrategy_ input, HedgingSettings settings);
    using Task = engine::TaskWithResult<std::optional<ReplyType>>;
    HedgedRequestFuture(Task&& task) : task_(std::move(task)) {}
    Task task_;
};
 
/// Synchronously perform bulk hedged requests described by RequestStrategy and
/// return result of type std::vector<std::optional<ResultType>>. Result
/// contains replies for each element in @param inputs or std::nullopt in case
/// either timeouts or bad replies (RequestStrategy::ProcessReply(RequestType&&)
/// returned std::nullopt and RequestStrategy::ExtractReply() returned
/// std::nullopt)
template <typename RequestStrategy>
auto HedgeRequestsBulk(std::vector<RequestStrategy> inputs, HedgingSettings hedging_settings) {
    {
        using Action = impl::Action;
        using Clock = impl::Clock;
        auto context = impl::Context(std::move(inputs), std::move(hedging_settings));
 
        auto& sub_requests = context.GetSubRequests();
 
        auto wakeup_time = Clock::now();
        context.Prepare(wakeup_time);
 
        while (!context.IsStop()) {
            auto wait_result = engine::WaitAnyUntil(wakeup_time, sub_requests);
            if (!wait_result.has_value()) {
                if (engine::current_task::ShouldCancel()) {
                    context.OnActionStop();
                    break;
                }
                /// timeout - need to process plan
                auto plan_entry = context.PopPlan();
                if (!plan_entry.has_value()) {
                    /// Timeout but we don't have planed actions any more
                    break;
                }
                const auto [timestamp, request_index, attempt_id, action] = *plan_entry;
                switch (action) {
                    case Action::StartTry:
                        context.OnActionStartTry(request_index, attempt_id, timestamp);
                        break;
                    case Action::Stop:
                        context.OnActionStop();
                        break;
                }
                auto next_wakeup_time = context.NextEventTime();
                if (!next_wakeup_time.has_value()) {
                    break;
                }
                wakeup_time = *next_wakeup_time;
                continue;
            }
            const auto result_idx = *wait_result;
            UASSERT(result_idx < sub_requests.size());
            const auto request_idx = context.GetRequestIdxBySubrequestIdx(result_idx);
            auto& strategy = context.GetStrategy(request_idx);
 
            auto& request = sub_requests[result_idx].request;
            UASSERT_MSG(request, "Finished requests must not be empty");
            auto reply = strategy.ProcessReply(std::move(*request));
            if (reply.has_value()) {
                /// Got reply but it's not OK and user wants to retry over
                /// some delay
                context.OnRetriableReply(request_idx, *reply, Clock::now());
                /// No need to check. we just added one entry
                wakeup_time = *context.NextEventTime();
            } else {
                context.OnNonRetriableReply(request_idx);
            }
        }
        return context.ExtractAllReplies();
    }
}
 
/// Asynchronously perform bulk hedged requests described by RequestStrategy and
/// return future which returns Result of type
/// std::vector<std::optional<ResultType>>. Result contains replies for each
/// element in @param inputs or std::nullopt in case either timeouts or bad
/// replies (RequestStrategy::ProcessReply(RequestType&&) returned std::nullopt
/// and RequestStrategy::ExtractReply() returned std::nullopt)
template <typename RequestStrategy>
auto HedgeRequestsBulkAsync(std::vector<RequestStrategy> inputs, HedgingSettings settings) {
    return HedgedRequestBulkFuture<RequestStrategy>(utils::Async(
        "hedged-bulk-request",
        [inputs{std::move(inputs)}, settings{std::move(settings)}]() mutable {
            return HedgeRequestsBulk(std::move(inputs), std::move(settings));
        }
    ));
}
 
/// Synchronously Perform hedged request described by RequestStrategy and return
/// result or throw runtime_error. Exception can be thrown in case of timeout or
/// if request was denied by strategy e.g. ProcessReply always returned
/// std::nullopt or ExtractReply returned std::nullopt
template <typename RequestStrategy>
std::optional<typename RequestTraits<RequestStrategy>::ReplyType>
HedgeRequest(RequestStrategy input, HedgingSettings settings) {
    std::vector<RequestStrategy> inputs;
    inputs.emplace_back(std::move(input));
    auto bulk_ret = HedgeRequestsBulk(std::move(inputs), std::move(settings));
    if (bulk_ret.size() != 1) {
        return std::nullopt;
    }
    return bulk_ret[0];
}
 
/// Create future which perform hedged request described by RequestStrategy and
/// return result or throw runtime_error. Exception can be thrown in case of
/// timeout or if request was denied by strategy e.g. ProcessReply always
/// returned std::nullopt or ExtractReply returned std::nullopt
template <typename RequestStrategy>
auto HedgeRequestAsync(RequestStrategy input, HedgingSettings settings) {
    return HedgedRequestFuture<RequestStrategy>(utils::Async(
        "hedged-request",
        [input{std::move(input)}, settings{std::move(settings)}]() mutable {
            return HedgeRequest(std::move(input), std::move(settings));
        }
    ));
}
 
}  // namespace utils::hedging
 
USERVER_NAMESPACE_END