ClickHouse/src/Storages/Kafka/ReadBufferFromKafkaConsumer.cpp

#include <Storages/Kafka/ReadBufferFromKafkaConsumer.h>

#include <common/logger_useful.h>

#include <cppkafka/cppkafka.h>
#include <boost/algorithm/string/join.hpp>
#include <algorithm>

namespace DB
{

namespace ErrorCodes
{
    extern const int CANNOT_COMMIT_OFFSET;
}

using namespace std::chrono_literals;
const auto MAX_TIME_TO_WAIT_FOR_ASSIGNMENT_MS = 15000;
const std::size_t POLL_TIMEOUT_WO_ASSIGNMENT_MS = 50;
const auto DRAIN_TIMEOUT_MS = 5000ms;


ReadBufferFromKafkaConsumer::ReadBufferFromKafkaConsumer(
    ConsumerPtr consumer_,
    Poco::Logger * log_,
    size_t max_batch_size,
    size_t poll_timeout_,
    bool intermediate_commit_,
    const std::atomic<bool> & stopped_,
    const Names & _topics)
    : ReadBuffer(nullptr, 0)
    , consumer(consumer_)
    , log(log_)
    , batch_size(max_batch_size)
    , poll_timeout(poll_timeout_)
    , intermediate_commit(intermediate_commit_)
    , stopped(stopped_)
    , current(messages.begin())
    , topics(_topics)
{
    // called (synchronously, during poll) when we enter the consumer group
    consumer->set_assignment_callback([this](const cppkafka::TopicPartitionList & topic_partitions)
    {
        LOG_TRACE(log, "Topics/partitions assigned: {}", topic_partitions);
        assignment = topic_partitions;
    });

    // called (synchronously, during poll) when we leave the consumer group
    consumer->set_revocation_callback([this](const cppkafka::TopicPartitionList & topic_partitions)
    {
        // Rebalance is happening now, and now we have a chance to finish the work
        // with topics/partitions we were working with before rebalance
        LOG_TRACE(log, "Rebalance initiated. Revoking partitions: {}", topic_partitions);

        // we can not flush data to target from that point (it is pulled, not pushed)
        // so the best we can now it to
        // 1) repeat last commit in sync mode (async could be still in queue, we need to be sure is is properly committed before rebalance)
        // 2) stop / brake the current reading:
        //     * clean buffered non-commited messages
        //     * set flag / flush

        cleanUnprocessed();

        stalled_status = REBALANCE_HAPPENED;
        assignment.clear();
        waited_for_assignment = 0;

        // for now we use slower (but reliable) sync commit in main loop, so no need to repeat
        // try
        // {
        //     consumer->commit();
        // }
        // catch (cppkafka::HandleException & e)
        // {
        //     LOG_WARNING(log, "Commit error: {}", e.what());
        // }
    });

    consumer->set_rebalance_error_callback([this](cppkafka::Error err)
    {
        LOG_ERROR(log, "Rebalance error: {}", err);
    });
}

ReadBufferFromKafkaConsumer::~ReadBufferFromKafkaConsumer()
{
    try
    {
        if (!consumer->get_subscription().empty())
        {
            try
            {
                consumer->unsubscribe();
            }
            catch (const cppkafka::HandleException & e)
            {
                LOG_ERROR(log, "Error during unsubscribe: {}", e.what());
            }
            drain();
        }
    }
    catch (const cppkafka::HandleException & e)
    {
        LOG_ERROR(log, "Error while destructing consumer: {}", e.what());
    }

}

// Needed to drain rest of the messages / queued callback calls from the consumer
// after unsubscribe, otherwise consumer will hang on destruction
// see https://github.com/edenhill/librdkafka/issues/2077
//     https://github.com/confluentinc/confluent-kafka-go/issues/189 etc.
void ReadBufferFromKafkaConsumer::drain()
{
    auto start_time = std::chrono::steady_clock::now();
    cppkafka::Error last_error(RD_KAFKA_RESP_ERR_NO_ERROR);

    while (true)
    {
        auto msg = consumer->poll(100ms);
        if (!msg)
            break;

        auto error = msg.get_error();

        if (error)
        {
            if (msg.is_eof() || error == last_error)
            {
                break;
            }
            else
            {
                LOG_ERROR(log, "Error during draining: {}", error);
            }
        }

        // i don't stop draining on first error,
        // only if it repeats once again sequentially
        last_error = error;

        auto ts = std::chrono::steady_clock::now();
        if (std::chrono::duration_cast<std::chrono::milliseconds>(ts-start_time) > DRAIN_TIMEOUT_MS)
        {
            LOG_ERROR(log, "Timeout during draining.");
            break;
        }
    }
}


void ReadBufferFromKafkaConsumer::commit()
{
    auto print_offsets = [this] (const char * prefix, const cppkafka::TopicPartitionList & offsets)
    {
        for (const auto & topic_part : offsets)
        {
            auto print_special_offset = [&topic_part]
            {
                switch (topic_part.get_offset())
                {
                    case cppkafka::TopicPartition::OFFSET_BEGINNING: return "BEGINNING";
                    case cppkafka::TopicPartition::OFFSET_END: return "END";
                    case cppkafka::TopicPartition::OFFSET_STORED: return "STORED";
                    case cppkafka::TopicPartition::OFFSET_INVALID: return "INVALID";
                    default: return "";
                }
            };

            if (topic_part.get_offset() < 0)
            {
                LOG_TRACE(log, "{} {} (topic: {}, partition: {})", prefix, print_special_offset(), topic_part.get_topic(), topic_part.get_partition());
            }
            else
            {
                LOG_TRACE(log, "{} {} (topic: {}, partition: {})", prefix, topic_part.get_offset(), topic_part.get_topic(), topic_part.get_partition());
            }
        }
    };

    print_offsets("Polled offset", consumer->get_offsets_position(consumer->get_assignment()));

    if (hasMorePolledMessages())
    {
        LOG_WARNING(log, "Logical error. Non all polled messages were processed.");
    }

    if (offsets_stored > 0)
    {
        // if we will do async commit here (which is faster)
        // we may need to repeat commit in sync mode in revocation callback,
        // but it seems like existing API doesn't allow us to to that
        // in a controlled manner (i.e. we don't know the offsets to commit then)

        size_t max_retries = 5;
        bool committed = false;

        while (!committed && max_retries > 0)
        {
            try
            {
                // See https://github.com/edenhill/librdkafka/issues/1470
                // broker may reject commit if during offsets.commit.timeout.ms (5000 by default),
                // there were not enough replicas available for the __consumer_offsets topic.
                // also some other temporary issues like client-server connectivity problems are possible
                consumer->commit();
                committed = true;
                print_offsets("Committed offset", consumer->get_offsets_committed(consumer->get_assignment()));
            }
            catch (const cppkafka::HandleException & e)
            {
                LOG_ERROR(log, "Exception during commit attempt: {}", e.what());
            }
            --max_retries;
        }

        if (!committed)
        {
            // TODO: insert atomicity / transactions is needed here (possibility to rollback, on 2 phase commits)
            throw Exception("All commit attempts failed. Last block was already written to target table(s), but was not committed to Kafka.", ErrorCodes::CANNOT_COMMIT_OFFSET);
        }
    }
    else
    {
        LOG_TRACE(log, "Nothing to commit.");
    }

    offsets_stored = 0;
}

void ReadBufferFromKafkaConsumer::subscribe()
{
    LOG_TRACE(log, "Already subscribed to topics: [{}]", boost::algorithm::join(consumer->get_subscription(), ", "));
    LOG_TRACE(log, "Already assigned to: {}", assignment);

    size_t max_retries = 5;

    while (consumer->get_subscription().empty())
    {
        --max_retries;
        try
        {
            consumer->subscribe(topics);
            // FIXME: if we failed to receive "subscribe" response while polling and destroy consumer now, then we may hang up.
            //        see https://github.com/edenhill/librdkafka/issues/2077
        }
        catch (cppkafka::HandleException & e)
        {
            if (max_retries > 0 && e.get_error() == RD_KAFKA_RESP_ERR__TIMED_OUT)
                continue;
            throw;
        }
    }

    cleanUnprocessed();
    allowed = false;

    // we can reset any flags (except of CONSUMER_STOPPED) before attempt of reading new block of data
    if (stalled_status != CONSUMER_STOPPED)
        stalled_status = NO_MESSAGES_RETURNED;
}

void ReadBufferFromKafkaConsumer::cleanUnprocessed()
{
    messages.clear();
    current = messages.begin();
    BufferBase::set(nullptr, 0, 0);
    offsets_stored = 0;
}

void ReadBufferFromKafkaConsumer::unsubscribe()
{
    LOG_TRACE(log, "Re-joining claimed consumer after failure");
    cleanUnprocessed();

    // it should not raise exception as used in destructor
    try
    {
        // From docs: Any previous subscription will be unassigned and unsubscribed first.
        consumer->subscribe(topics);

        // I wanted to avoid explicit unsubscribe as it requires draining the messages
        // to close the consumer safely after unsubscribe
        // see https://github.com/edenhill/librdkafka/issues/2077
        //     https://github.com/confluentinc/confluent-kafka-go/issues/189 etc.
    }
    catch (const cppkafka::HandleException & e)
    {
        LOG_ERROR(log, "Exception from ReadBufferFromKafkaConsumer::unsubscribe: {}", e.what());
    }

}


void ReadBufferFromKafkaConsumer::resetToLastCommitted(const char * msg)
{
    if (assignment.empty())
    {
        LOG_TRACE(log, "Not assignned. Can't reset to last committed position.");
        return;
    }
    auto committed_offset = consumer->get_offsets_committed(consumer->get_assignment());
    consumer->assign(committed_offset);
    LOG_TRACE(log, "{} Returned to committed position: {}", msg, committed_offset);
}

// it do the poll when needed
bool ReadBufferFromKafkaConsumer::poll()
{
    resetIfStopped();

    if (polledDataUnusable())
        return false;

    if (hasMorePolledMessages())
    {
        allowed = true;
        return true;
    }

    if (intermediate_commit)
        commit();

    while (true)
    {
        stalled_status = NO_MESSAGES_RETURNED;

        // we already wait enough for assignment in the past,
        // let's make polls shorter and not block other consumer
        // which can work successfully in parallel
        // POLL_TIMEOUT_WO_ASSIGNMENT_MS (50ms) is 100% enough just to check if we got assignment
        //  (see https://github.com/ClickHouse/ClickHouse/issues/11218)
        auto actual_poll_timeout_ms = (waited_for_assignment >= MAX_TIME_TO_WAIT_FOR_ASSIGNMENT_MS)
                        ? std::min(POLL_TIMEOUT_WO_ASSIGNMENT_MS,poll_timeout)
                        : poll_timeout;

        /// Don't drop old messages immediately, since we may need them for virtual columns.
        auto new_messages = consumer->poll_batch(batch_size,
                            std::chrono::milliseconds(actual_poll_timeout_ms));

        resetIfStopped();
        if (stalled_status == CONSUMER_STOPPED)
        {
            return false;
        }
        else if (stalled_status == REBALANCE_HAPPENED)
        {
            if (!new_messages.empty())
            {
                // we have polled something just after rebalance.
                // we will not use current batch, so we need to return to last committed position
                // otherwise we will continue polling from that position
                resetToLastCommitted("Rewind last poll after rebalance.");
            }
            return false;
        }

        if (new_messages.empty())
        {
            // While we wait for an assignment after subscription, we'll poll zero messages anyway.
            // If we're doing a manual select then it's better to get something after a wait, then immediate nothing.
            if (assignment.empty())
            {
                waited_for_assignment += poll_timeout; // slightly innaccurate, but rough calculation is ok.
                if (waited_for_assignment < MAX_TIME_TO_WAIT_FOR_ASSIGNMENT_MS)
                {
                    continue;
                }
                else
                {
                    LOG_WARNING(log, "Can't get assignment. It can be caused by some issue with consumer group (not enough partitions?). Will keep trying.");
                    stalled_status = NO_ASSIGNMENT;
                    return false;
                }

            }
            else
            {
                LOG_TRACE(log, "Stalled");
                return false;
            }
        }
        else
        {
            messages = std::move(new_messages);
            current = messages.begin();
            LOG_TRACE(log, "Polled batch of {} messages. Offsets position: {}",
                messages.size(), consumer->get_offsets_position(consumer->get_assignment()));
            break;
        }
    }

    filterMessageErrors();
    if (current == messages.end())
    {
        LOG_ERROR(log, "Only errors left");
        stalled_status = ERRORS_RETURNED;
        return false;
    }

    stalled_status = NOT_STALLED;
    allowed = true;
    return true;
}

size_t ReadBufferFromKafkaConsumer::filterMessageErrors()
{
    assert(current == messages.begin());

    auto new_end = std::remove_if(messages.begin(), messages.end(), [this](auto & message)
    {
        if (auto error = message.get_error())
        {
            LOG_ERROR(log, "Consumer error: {}", error);
            return true;
        }
        return false;
    });

    size_t skipped = std::distance(new_end, messages.end());
    if (skipped)
    {
        LOG_ERROR(log, "There were {} messages with an error", skipped);
        messages.erase(new_end, messages.end());
    }

    return skipped;
}

void ReadBufferFromKafkaConsumer::resetIfStopped()
{
    // we can react on stop only during fetching data
    // after block is formed (i.e. during copying data to MV / committing)  we ignore stop attempts
    if (stopped)
    {
        stalled_status = CONSUMER_STOPPED;
        cleanUnprocessed();
    }
}

/// Do commit messages implicitly after we processed the previous batch.
bool ReadBufferFromKafkaConsumer::nextImpl()
{
    if (!allowed || !hasMorePolledMessages())
        return false;

    // XXX: very fishy place with const casting.
    auto * new_position = reinterpret_cast<char *>(const_cast<unsigned char *>(current->get_payload().get_data()));
    BufferBase::set(new_position, current->get_payload().get_size(), 0);
    allowed = false;

    ++current;

    return true;
}

void ReadBufferFromKafkaConsumer::storeLastReadMessageOffset()
{
    if (!isStalled())
    {
        consumer->store_offset(*(current - 1));
        ++offsets_stored;
    }
}

}