doris/common/cpp/token_bucket_rate_limiter.cpp

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

#include "token_bucket_rate_limiter.h"

#include <bthread/bthread.h>
#include <glog/logging.h> // IWYU pragma: export

#include <atomic>
#include <chrono>
#include <mutex>
#include <thread>

#if defined(__APPLE__)
#include <ctime>
#endif

namespace doris {
// Just 10^6.
static constexpr auto NS = 1000000000UL;

bvar::Adder<int64_t> s3_get_rate_limit_sleep_ns("s3_get_rate_limit_sleep_ns");
bvar::Adder<int64_t> s3_get_rate_limit_sleep_count("s3_get_rate_limit_sleep_count");
bvar::Adder<int64_t> s3_get_rate_limit_rejected_count("s3_get_rate_limit_rejected_count");
bvar::Adder<int64_t> s3_put_rate_limit_sleep_ns("s3_put_rate_limit_sleep_ns");
bvar::Adder<int64_t> s3_put_rate_limit_sleep_count("s3_put_rate_limit_sleep_count");
bvar::Adder<int64_t> s3_put_rate_limit_rejected_count("s3_put_rate_limit_rejected_count");

static std::atomic<int64_t> s3_get_rate_limit_sleep_log_count {0};
static std::atomic<int64_t> s3_get_rate_limit_rejected_log_count {0};
static std::atomic<int64_t> s3_put_rate_limit_sleep_log_count {0};
static std::atomic<int64_t> s3_put_rate_limit_rejected_log_count {0};

class TokenBucketRateLimiter::SimpleSpinLock {
public:
    SimpleSpinLock() = default;
    ~SimpleSpinLock() = default;

    void lock() {
        int spin_count = 0;
        static constexpr int MAX_SPIN_COUNT = 50;
        while (_flag.test_and_set(std::memory_order_acq_rel)) {
            spin_count++;
            if (spin_count >= MAX_SPIN_COUNT) {
                LOG(WARNING) << "Warning: Excessive spinning detected while acquiring lock. Spin "
                                "count: "
                             << spin_count;
                spin_count = 0;
            }
            // Spin until we acquire the lock
        }
    }

    void unlock() { _flag.clear(std::memory_order_release); }

private:
    std::atomic_flag _flag = ATOMIC_FLAG_INIT;
};

TokenBucketRateLimiter::TokenBucketRateLimiter(size_t max_speed, size_t max_burst, size_t limit)
        : _max_speed(max_speed),
          _max_burst(max_burst),
          _limit(limit),
          _mutex(std::make_unique<TokenBucketRateLimiter::SimpleSpinLock>()),
          _remain_tokens(max_burst) {}

TokenBucketRateLimiter::~TokenBucketRateLimiter() = default;

TokenBucketRateLimiterHolder::~TokenBucketRateLimiterHolder() = default;

std::pair<size_t, double> TokenBucketRateLimiter::_update_remain_token(long now, size_t amount) {
    // Values obtained under lock to be checked after release
    size_t count_value;
    double tokens_value;
    {
        std::lock_guard<SimpleSpinLock> lock(*_mutex);
        now = (now < _prev_ns_count) ? _prev_ns_count : now;
        if (_max_speed) {
            double delta_seconds =
                    _prev_ns_count ? static_cast<double>(now - _prev_ns_count) / NS : 0;
            _remain_tokens = std::min<double>(_remain_tokens + _max_speed * delta_seconds - amount,
                                              _max_burst);
        }
        _count += amount;
        count_value = _count;
        if (_limit && count_value > _limit) {
            // Keep rejection side-effect free. Roll back before releasing the lock so
            // concurrent callers cannot observe debt from a request that will not run.
            _count -= amount;
            if (_max_speed) {
                _remain_tokens = std::min<double>(_remain_tokens + amount, _max_burst);
            }
        }
        tokens_value = _remain_tokens;
        _prev_ns_count = now;
    }
    return {count_value, tokens_value};
}

int64_t TokenBucketRateLimiter::add(size_t amount) {
    int64_t sleep_time_ns = reserve(amount);
    if (sleep_time_ns > 0) {
        bthread_usleep(sleep_time_ns / 1000);
    }
    return sleep_time_ns;
}

int64_t TokenBucketRateLimiter::reserve(size_t amount) {
    // Values obtained under lock to be checked after release
    auto duration = std::chrono::steady_clock::now().time_since_epoch();
    auto time_nano_count = std::chrono::duration_cast<std::chrono::nanoseconds>(duration).count();
    auto [count_value, tokens_value] = _update_remain_token(time_nano_count, amount);

    if (_limit && count_value > _limit) {
        // CK would throw exception
        return -1;
    }

    // Wait unless there is positive amount of remain_tokens - throttling
    int64_t sleep_time_ns = 0;
    if (_max_speed && tokens_value < 0) {
        sleep_time_ns = static_cast<int64_t>(-tokens_value / _max_speed * NS);
    }

    return sleep_time_ns;
}

void TokenBucketRateLimiter::refund(size_t amount) {
    std::lock_guard<SimpleSpinLock> lock(*_mutex);
    if (_max_speed) {
        _remain_tokens = std::min<double>(_remain_tokens + amount, _max_burst);
    }
    _count = (_count >= amount) ? _count - amount : 0;
}

TokenBucketRateLimiterHolder::TokenBucketRateLimiterHolder(size_t max_speed, size_t max_burst,
                                                           size_t limit,
                                                           std::function<void(int64_t)> metric_func)
        : rate_limiter(std::make_shared<TokenBucketRateLimiter>(max_speed, max_burst, limit)),
          _enabled(max_speed > 0 || limit > 0),
          metric_func(std::move(metric_func)) {}

int64_t TokenBucketRateLimiterHolder::add(size_t amount) {
    return add_with_config(amount).sleep_duration;
}

TokenBucketRateLimiterResult TokenBucketRateLimiterHolder::add_with_config(size_t amount) {
    return _consume_with_config(amount, true);
}

TokenBucketRateLimiterResult TokenBucketRateLimiterHolder::reserve_with_config(size_t amount) {
    return _consume_with_config(amount, false);
}

TokenBucketRateLimiterResult TokenBucketRateLimiterHolder::_consume_with_config(size_t amount,
                                                                                bool wait) {
    // Snapshot the current limiter and consume outside the read lock. The waiting
    // path may sleep for a long time when throttled, and holding the read lock across
    // the sleep would block reset() (dynamic config update) for the whole duration.
    std::shared_ptr<TokenBucketRateLimiter> limiter;
    {
        std::shared_lock read {rate_limiter_rw_lock};
        limiter = rate_limiter;
    }
    TokenBucketRateLimiterResult result = {.sleep_duration = wait ? limiter->add(amount)
                                                                  : limiter->reserve(amount),
                                           .max_speed = limiter->get_max_speed(),
                                           .max_burst = limiter->get_max_burst(),
                                           .limit = limiter->get_limit()};
    metric_func(result.sleep_duration);
    return result;
}

std::shared_ptr<TokenBucketRateLimiter> TokenBucketRateLimiterHolder::charge(size_t amount) {
    std::shared_ptr<TokenBucketRateLimiter> limiter;
    {
        std::shared_lock read {rate_limiter_rw_lock};
        limiter = rate_limiter;
    }
    int64_t sleep_duration = limiter->add(amount);
    metric_func(sleep_duration);
    return sleep_duration < 0 ? nullptr : limiter;
}

int TokenBucketRateLimiterHolder::reset(size_t max_speed, size_t max_burst, size_t limit) {
    auto new_rate_limiter = std::make_shared<TokenBucketRateLimiter>(max_speed, max_burst, limit);
    {
        std::unique_lock write {rate_limiter_rw_lock};
        rate_limiter = std::move(new_rate_limiter);
        _enabled.store(max_speed > 0 || limit > 0, std::memory_order_release);
    }
    return 0;
}

size_t TokenBucketRateLimiterHolder::get_max_speed() const {
    std::shared_lock read {rate_limiter_rw_lock};
    return rate_limiter->get_max_speed();
}

size_t TokenBucketRateLimiterHolder::get_max_burst() const {
    std::shared_lock read {rate_limiter_rw_lock};
    return rate_limiter->get_max_burst();
}

size_t TokenBucketRateLimiterHolder::get_limit() const {
    std::shared_lock read {rate_limiter_rw_lock};
    return rate_limiter->get_limit();
}

std::string to_string(S3RateLimitType type) {
    switch (type) {
    case S3RateLimitType::GET:
        return "get";
    case S3RateLimitType::PUT:
        return "put";
    default:
        return std::to_string(static_cast<size_t>(type));
    }
}

S3RateLimitType string_to_s3_rate_limit_type(std::string_view value) {
    if (value == "get") {
        return S3RateLimitType::GET;
    } else if (value == "put") {
        return S3RateLimitType::PUT;
    }
    return S3RateLimitType::UNKNOWN;
}

std::function<void(int64_t)> s3_rate_limiter_metric_func(S3RateLimitType type) {
    switch (type) {
    case S3RateLimitType::GET:
        return metric_func_factory(s3_get_rate_limit_sleep_ns, s3_get_rate_limit_sleep_count,
                                   &s3_get_rate_limit_rejected_count);
    case S3RateLimitType::PUT:
        return metric_func_factory(s3_put_rate_limit_sleep_ns, s3_put_rate_limit_sleep_count,
                                   &s3_put_rate_limit_rejected_count);
    default:
        return [](int64_t) {};
    }
}

int64_t apply_s3_rate_limit(S3RateLimitType type, S3RateLimiterHolder* rate_limiter,
                            int64_t log_interval) {
    auto result = rate_limiter->add_with_config(1);
    auto sleep_duration = result.sleep_duration;
    if (log_interval <= 0 || sleep_duration == 0) {
        return sleep_duration;
    }

    auto is_get = type == S3RateLimitType::GET;
    auto* sleep_log_count =
            is_get ? &s3_get_rate_limit_sleep_log_count : &s3_put_rate_limit_sleep_log_count;
    auto* rejected_log_count =
            is_get ? &s3_get_rate_limit_rejected_log_count : &s3_put_rate_limit_rejected_log_count;

    if (sleep_duration > 0) {
        int64_t count = sleep_log_count->fetch_add(1, std::memory_order_relaxed) + 1;
        if (count == 1 || count % log_interval == 0) {
            LOG(INFO) << "S3 " << to_string(type) << " request is throttled by local rate limiter"
                      << ", sleep_ms=" << sleep_duration / 1000000 << ", sleep_count=" << count
                      << ", token_per_second=" << result.max_speed
                      << ", bucket_tokens=" << result.max_burst << ", token_limit=" << result.limit;
        }
    } else {
        int64_t count = rejected_log_count->fetch_add(1, std::memory_order_relaxed) + 1;
        if (count == 1 || count % log_interval == 0) {
            LOG(WARNING) << "S3 " << to_string(type) << " request is rejected by local rate limiter"
                         << ", rejected_count=" << count
                         << ", token_per_second=" << result.max_speed
                         << ", bucket_tokens=" << result.max_burst
                         << ", token_limit=" << result.limit;
        }
    }
    return sleep_duration;
}
} // namespace doris