YugabyteDB (2.13.0.0-b42, bfc6a6643e7399ac8a0e81d06a3ee6d6571b33ab)

// Copyright (c) YugaByte, Inc.

//

// Licensed 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 "yb/cdc/cdc_producer.h"

#include "yb/cdc/cdc_common_util.h"

#include "yb/cdc/cdc_service.pb.h"

#include "yb/common/schema.h"

#include "yb/common/transaction.h"

#include "yb/common/wire_protocol.h"

#include "yb/consensus/raft_consensus.h"

#include "yb/consensus/replicate_msgs_holder.h"

#include "yb/docdb/doc_key.h"

#include "yb/docdb/docdb.pb.h"

#include "yb/docdb/primitive_value.h"

#include "yb/docdb/value.h"

#include "yb/docdb/value_type.h"

#include "yb/tablet/tablet.h"

#include "yb/tablet/tablet_metadata.h"

#include "yb/tablet/tablet_peer.h"

#include "yb/tablet/transaction_participant.h"

#include "yb/tserver/tablet_server.h"

#include "yb/tserver/ts_tablet_manager.h"

DEFINE_int32(cdc_transaction_timeout_ms, 0,

  "Don't check for an aborted transaction unless its original write is lagging by this duration.");

// Todo(Rahul): Enable this by default (https://github.com/yugabyte/yugabyte-db/issues/6128)

DEFINE_bool(cdc_enable_replicate_intents, true,

            "Enable replication of intents before they've been committed.");

namespace yb {

namespace cdc {

using consensus::ReplicateMsgPtr;

using consensus::ReplicateMsgs;

using docdb::PrimitiveValue;

using tablet::TransactionParticipant;

void AddColumnToMap(const ColumnSchema& col_schema,

                    const docdb::PrimitiveValue& col,

                    cdc::KeyValuePairPB* kv_pair) {

  kv_pair->set_key(col_schema.name());

  PrimitiveValue::ToQLValuePB(col, col_schema.type(), kv_pair->mutable_value());

}

void AddPrimaryKey(const docdb::SubDocKey& decoded_key,

                   const Schema& tablet_schema,

                   CDCRecordPB* record) {

  size_t i = 0;

  for (const auto& col : decoded_key.doc_key().hashed_group()) {

    AddColumnToMap(tablet_schema.column(i), col, record->add_key());

    i++;

  }

  for (const auto& col : decoded_key.doc_key().range_group()) {

    AddColumnToMap(tablet_schema.column(i), col, record->add_key());

    i++;

  }

}

// Set committed record information including commit time for record.

// This will look at transaction status to determine commit time to be used for CDC record.

// Returns true if we need to stop processing WAL records beyond this, false otherwise.

Result<bool> SetCommittedRecordIndexForReplicateMsg(

    const ReplicateMsgPtr& msg, size_t index, const TxnStatusMap& txn_map,

    ReplicateIntents replicate_intents, std::vector<RecordTimeIndex>* records) {

  if (replicate_intents) {

    // If we're replicating intents, we have no stop condition, so add the record and continue.

    records->emplace_back(msg->hybrid_time(), index);

    return false;

  }

  switch (msg->op_type()) {

    case consensus::OperationType::UPDATE_TRANSACTION_OP: {

      if (msg->transaction_state().status() == TransactionStatus::APPLYING) {

        records->emplace_back(msg->transaction_state().commit_hybrid_time(), index);

      }

      // Ignore other transaction statuses since we only care about APPLYING

      // while sending CDC records.

      return false;

    }

    case consensus::OperationType::WRITE_OP: {

      if (msg->write().write_batch().has_transaction()) {

        auto txn_id = VERIFY_RESULT(FullyDecodeTransactionId(

            msg->write().write_batch().transaction().transaction_id()));

        const auto txn_status = txn_map.find(txn_id);

        if (txn_status == txn_map.end()) {

          return STATUS(IllegalState, "Unexpected transaction ID", txn_id.ToString());

        }

        if (txn_status->second.status == PENDING || txn_status->second.status == CREATED) {

          // Ignore all records beyond this because we don't know whether those records

          // were committed before or after this record without the transaction commit time.

          return true;

        } else if (txn_status->second.status == COMMITTED) {

          // Add record to txn_msgs because there may be records appearing after this in WAL

          // but committed before this one. Example:

          // T0: WRITE K1 [TXN1]

          // T1: WRITE K2

          // T2: APPLYING TXN1

          // Here, WRITE K2 appears after WRITE K1 but is committed before K1.

          records->emplace_back(txn_status->second.status_time.ToUint64(), index);

        }

      } else {

        // Msg is not part of transaction. Use write hybrid time from msg itself.

        records->emplace_back(msg->hybrid_time(), index);

      }

      return false;

    }

    case consensus::OperationType::SPLIT_OP: {

      records->emplace_back(msg->hybrid_time(), index);

      return true;  // Don't need to process any records after a SPLIT_OP.

    }

    case consensus::OperationType::CHANGE_CONFIG_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::CHANGE_METADATA_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::HISTORY_CUTOFF_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::NO_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::SNAPSHOT_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::TRUNCATE_OP:

      FALLTHROUGH_INTENDED;

    case consensus::OperationType::UNKNOWN_OP:

      return false;

  }

  FATAL_INVALID_ENUM_VALUE(consensus::OperationType, msg->op_type());

}

Result<std::vector<RecordTimeIndex>> GetCommittedRecordIndexes(

    const ReplicateMsgs& msgs, const TxnStatusMap& txn_map, ReplicateIntents replicate_intents,

    OpId* checkpoint) {

  size_t index = 0;

  std::vector<RecordTimeIndex> records;

  // Order ReplicateMsgs based on commit time.

  for (const auto &msg : msgs) {

    if (!msg->write().has_external_hybrid_time()) {

      // If the message came from an external source, ignore it when producing change list.

      // Note that checkpoint, however, will be updated and will account for external message too.

      bool stop = VERIFY_RESULT(SetCommittedRecordIndexForReplicateMsg(

          msg, index, txn_map, replicate_intents, &records));

      if (stop) {

        return records;

      }

    }

    *checkpoint = OpId::FromPB(msg->id());

    index++;

  }

  return records;

}

// Filter out WAL records that are external and order records based on transaction commit time.

// Records in WAL don't represent the exact order in which records are written in DB due to delay

// in writing txn APPLYING record.

// Consider the following WAL entries:

// TO: WRITE K0

// T1: WRITE K1 (TXN1)

// T2: WRITE K2 (TXN2)

// T3: WRITE K3

// T4: APPLYING TXN2

// T5: APPLYING TXN1

// T6: WRITE K4

// The order in which keys are written to DB in this example is K0, K3, K2, K1, K4.

// This method will also set checkpoint to the op id of last processed record.

Result<ReplicateMsgs> FilterAndSortWrites(const ReplicateMsgs& msgs,

                                          const TxnStatusMap& txn_map,

                                          ReplicateIntents replicate_intents,

                                          OpId* checkpoint) {

  std::vector<RecordTimeIndex> records = VERIFY_RESULT(GetCommittedRecordIndexes(

      msgs, txn_map, replicate_intents, checkpoint));

  if (!replicate_intents) {

    std::sort(records.begin(), records.end());

  }

  ReplicateMsgs ordered_msgs;

  ordered_msgs.reserve(records.size());

  for (const auto& record : records) {

    ordered_msgs.emplace_back(msgs[record.second]);

  }

  return ordered_msgs;

}

Result<TransactionStatusResult> GetTransactionStatus(

    const TransactionId& txn_id,

    const HybridTime& hybrid_time,

    TransactionParticipant* txn_participant) {

  static const std::string reason = "cdc";

  std::promise<Result<TransactionStatusResult>> txn_status_promise;

  auto future = txn_status_promise.get_future();

  auto callback = [&txn_status_promise](Result<TransactionStatusResult> result) {

    txn_status_promise.set_value(std::move(result));

  };

  txn_participant->RequestStatusAt(

      {&txn_id, hybrid_time, hybrid_time, 0, &reason, TransactionLoadFlags{}, callback});

  future.wait();

  return future.get();

}

// Build transaction status as of hybrid_time.

Result<TxnStatusMap> BuildTxnStatusMap(const ReplicateMsgs& messages,

                                       bool more_replicate_msgs,

                                       const HybridTime& cdc_read_hybrid_time,

                                       TransactionParticipant* txn_participant) {

  TxnStatusMap txn_map;

  // First go through all APPLYING records and mark transaction as committed.

  for (const auto& msg : messages) {

    if (msg->op_type() == consensus::OperationType::UPDATE_TRANSACTION_OP

        && msg->transaction_state().status() == TransactionStatus::APPLYING) {

      auto txn_id = VERIFY_RESULT(FullyDecodeTransactionId(

          msg->transaction_state().transaction_id()));

      txn_map.emplace(txn_id,

                      TransactionStatusResult(

                          TransactionStatus::COMMITTED,

                          HybridTime(msg->transaction_state().commit_hybrid_time())));

    }

  }

  // Now go through all WRITE_OP records and get transaction status of records for which

  // corresponding APPLYING record does not exist in WAL as yet.

  for (const auto& msg : messages) {

    if (msg->op_type() == consensus::OperationType::WRITE_OP

        && msg->write().write_batch().has_transaction()) {

      auto txn_id = VERIFY_RESULT(FullyDecodeTransactionId(

          msg->write().write_batch().transaction().transaction_id()));

      if (!txn_map.count(txn_id)) {

        TransactionStatusResult txn_status(TransactionStatus::PENDING, HybridTime::kMin);

        auto result = GetTransactionStatus(txn_id, cdc_read_hybrid_time, txn_participant);

        if (!result.ok()) {

          if (result.status().IsNotFound()) {

            // Naive heuristic for handling whether a transaction is aborted or still pending:

            // 1. If the normal transaction timeout is not reached, assume good operation.

            // 2. If more_replicate_messages, assume a race between reading

            //    TransactionParticipant & LogCache.

            // TODO (#2405) : Handle long running or very large transactions correctly.

            if (!more_replicate_msgs) {

              auto timeout = HybridTime::FromPB(msg->hybrid_time())

                  .AddMilliseconds(FLAGS_cdc_transaction_timeout_ms);

              if (timeout < cdc_read_hybrid_time) {

                LOG(INFO) << "Transaction not found, considering it aborted: " << txn_id;

                txn_status = TransactionStatusResult::Aborted();

              }

            }

          } else {

            return result.status();

          }

        } else {

          txn_status = *result;

        }

        txn_map.emplace(txn_id, txn_status);

      }

    }

  }

  return txn_map;

}

CHECKED_STATUS SetRecordTime(const TransactionId& txn_id,

                             const TxnStatusMap& txn_map,

                             CDCRecordPB* record) {

  auto txn_status = txn_map.find(txn_id);

  if (txn_status == txn_map.end()) {

    return STATUS(IllegalState, "Unexpected transaction ID", txn_id.ToString());

  }

  record->set_time(txn_status->second.status_time.ToUint64());

  return Status::OK();

}

// Populate CDC record corresponding to WAL batch in ReplicateMsg.

CHECKED_STATUS PopulateWriteRecord(const ReplicateMsgPtr& msg,

                                   const TxnStatusMap& txn_map,

                                   const StreamMetadata& metadata,

                                   const std::shared_ptr<tablet::TabletPeer>& tablet_peer,

                                   ReplicateIntents replicate_intents,

                                   GetChangesResponsePB* resp) {

  const auto& batch = msg->write().write_batch();

  const auto& schema = *tablet_peer->tablet()->schema();

  // Write batch may contain records from different rows.

  // For CDC, we need to split the batch into 1 CDC record per row of the table.

  // We'll use DocDB key hash to identify the records that belong to the same row.

  Slice prev_key;

  CDCRecordPB* record = nullptr;

  for (const auto& write_pair : batch.write_pairs()) {

    Slice key = write_pair.key();

    const auto key_size = VERIFY_RESULT(

        docdb::DocKey::EncodedSize(key, docdb::DocKeyPart::kWholeDocKey));

    Slice value = write_pair.value();

    docdb::Value decoded_value;

    RETURN_NOT_OK(decoded_value.Decode(value));

    // Compare key hash with previously seen key hash to determine whether the write pair

    // is part of the same row or not.

    Slice primary_key(key.data(), key_size);

    if (prev_key != primary_key) {

      // Write pair contains record for different row. Create a new CDCRecord in this case.

      record = resp->add_records();

      Slice sub_doc_key = key;

      docdb::SubDocKey decoded_key;

      RETURN_NOT_OK(decoded_key.DecodeFrom(&sub_doc_key, docdb::HybridTimeRequired::kFalse));

      if (metadata.record_format == CDCRecordFormat::WAL) {

        // For 2DC, populate serialized data from WAL, to avoid unnecessary deserializing on

        // producer and re-serializing on consumer.

        auto kv_pair = record->add_key();

        kv_pair->set_key(std::to_string(decoded_key.doc_key().hash()));

        kv_pair->mutable_value()->set_binary_value(write_pair.key());

      } else {

        AddPrimaryKey(decoded_key, schema, record);

      }

      // Check whether operation is WRITE or DELETE.

      if (decoded_value.value_type() == docdb::ValueType::kTombstone &&

          decoded_key.num_subkeys() == 0) {

        record->set_operation(CDCRecordPB::DELETE);

      } else {

        record->set_operation(CDCRecordPB::WRITE);

      }

      // Process intent records.

      record->set_time(msg->hybrid_time());

      if (batch.has_transaction()) {

        if (!replicate_intents) {

          auto txn_id = VERIFY_RESULT(FullyDecodeTransactionId(

              batch.transaction().transaction_id()));

          // If we're not replicating intents, set record time using the transaction map.

          RETURN_NOT_OK(SetRecordTime(txn_id, txn_map, record));

        } else {

          record->mutable_transaction_state()->set_transaction_id(

              batch.transaction().transaction_id());

          record->mutable_transaction_state()->add_tablets(tablet_peer->tablet_id());

        }

      }

    }

    prev_key = primary_key;

    DCHECK(record);

    if (metadata.record_format == CDCRecordFormat::WAL) {

      auto kv_pair = record->add_changes();

      kv_pair->set_key(write_pair.key());

      kv_pair->mutable_value()->set_binary_value(write_pair.value());

    } else if (record->operation() == CDCRecordPB_OperationType_WRITE) {

      PrimitiveValue column_id;

      Slice key_column = write_pair.key().data() + key_size;

      RETURN_NOT_OK(PrimitiveValue::DecodeKey(&key_column, &column_id));

      if (column_id.value_type() == docdb::ValueType::kColumnId) {

        const ColumnSchema& col = VERIFY_RESULT(schema.column_by_id(column_id.GetColumnId()));

        AddColumnToMap(col, decoded_value.primitive_value(), record->add_changes());

      } else if (column_id.value_type() != docdb::ValueType::kSystemColumnId) {

        LOG(DFATAL) << "Unexpected value type in key: " << column_id.value_type();

      }

    }

  }

  return Status::OK();

}

// Populate CDC record corresponding to WAL UPDATE_TRANSACTION_OP entry.

CHECKED_STATUS PopulateTransactionRecord(const ReplicateMsgPtr& msg,

                                         const std::shared_ptr<tablet::TabletPeer>& tablet_peer,

                                         ReplicateIntents replicate_intents,

                                         CDCRecordPB* record) {

  SCHECK(msg->has_transaction_state(), InvalidArgument,

         Format("Update transaction message requires transaction_state: $0",

                msg->ShortDebugString()));

  record->set_operation(CDCRecordPB_OperationType_WRITE);

  record->set_time(replicate_intents ?

      msg->hybrid_time() : msg->transaction_state().commit_hybrid_time());

  record->mutable_transaction_state()->CopyFrom(msg->transaction_state());

  if (replicate_intents && msg->transaction_state().status() == TransactionStatus::APPLYING) {

    // Add the partition metadata so the consumer knows which tablets to apply the transaction

    // to.

    tablet_peer->tablet()->metadata()->partition()->ToPB(record->mutable_partition());

  }

  return Status::OK();

}

CHECKED_STATUS PopulateSplitOpRecord(const ReplicateMsgPtr& msg, CDCRecordPB* record) {

  SCHECK(msg->has_split_request(), InvalidArgument,

         Format("Split op message requires split_request: $0", msg->ShortDebugString()));

  record->set_operation(CDCRecordPB::SPLIT_OP);

  record->set_time(msg->hybrid_time());

  record->mutable_split_tablet_request()->CopyFrom(msg->split_request());

  return Status::OK();

}

Status GetChangesForXCluster(const std::string& stream_id,

                             const std::string& tablet_id,

                             const OpId& from_op_id,

                             const StreamMetadata& stream_metadata,

                             const std::shared_ptr<tablet::TabletPeer>& tablet_peer,

                             const MemTrackerPtr& mem_tracker,

                             consensus::ReplicateMsgsHolder* msgs_holder,

                             GetChangesResponsePB* resp,

                             int64_t* last_readable_opid_index,

                             const CoarseTimePoint deadline) {

  auto replicate_intents = ReplicateIntents(GetAtomicFlag(&FLAGS_cdc_enable_replicate_intents));

  // Request scope on transaction participant so that transactions are not removed from participant

  // while RequestScope is active.

  RequestScope request_scope;

  auto read_ops = VERIFY_RESULT(tablet_peer->consensus()->

    ReadReplicatedMessagesForCDC(from_op_id, last_readable_opid_index, deadline));

  ScopedTrackedConsumption consumption;

  if (read_ops.read_from_disk_size && mem_tracker) {

    consumption = ScopedTrackedConsumption(mem_tracker, read_ops.read_from_disk_size);

  }

  OpId checkpoint;

  TxnStatusMap txn_map;

  if (!replicate_intents) {

    auto txn_participant = tablet_peer->tablet()->transaction_participant();

    if (txn_participant) {

      request_scope = RequestScope(txn_participant);

    }

    txn_map = TxnStatusMap(VERIFY_RESULT(BuildTxnStatusMap(

      read_ops.messages, read_ops.have_more_messages, tablet_peer->Now(), txn_participant)));

  }

  ReplicateMsgs messages = VERIFY_RESULT(FilterAndSortWrites(

      read_ops.messages, txn_map, replicate_intents, &checkpoint));

  for (const auto& msg : messages) {

    switch (msg->op_type()) {

      case consensus::OperationType::UPDATE_TRANSACTION_OP:

        if (!replicate_intents) {

          RETURN_NOT_OK(PopulateTransactionRecord(

              msg, tablet_peer, replicate_intents, resp->add_records()));

        } else if (msg->transaction_state().status() == TransactionStatus::APPLYING) {

          auto record = resp->add_records();

          record->set_operation(CDCRecordPB::APPLY);

          record->set_time(msg->hybrid_time());

          auto* txn_state = record->mutable_transaction_state();

          txn_state->set_transaction_id(msg->transaction_state().transaction_id());

          txn_state->set_commit_hybrid_time(msg->transaction_state().commit_hybrid_time());

          tablet_peer->tablet()->metadata()->partition()->ToPB(record->mutable_partition());

        }

        break;

      case consensus::OperationType::WRITE_OP:

        RETURN_NOT_OK(PopulateWriteRecord(msg, txn_map, stream_metadata, tablet_peer,

                                          replicate_intents, resp));

        break;

      case consensus::OperationType::SPLIT_OP:

        RETURN_NOT_OK(PopulateSplitOpRecord(msg, resp->add_records()));

        break;

      default:

        // Nothing to do for other operation types.

        break;

    }

  }

  if (consumption) {

    consumption.Add(resp->SpaceUsedLong());

  }

  *msgs_holder = consensus::ReplicateMsgsHolder(

      nullptr, std::move(messages), std::move(consumption));

  (checkpoint.index > 0 ? checkpoint : from_op_id).ToPB(

      resp->mutable_checkpoint()->mutable_op_id());

  return Status::OK();

}

}  // namespace cdc

}  // namespace yb