src/kudu/tablet/rowset_info.cc

// 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 "kudu/tablet/rowset_info.h"

#include <algorithm>
#include <cstdint>
#include <cstring>
#include <memory>
#include <ostream>
#include <string>
#include <unordered_map>
#include <utility>

#include <gflags/gflags.h>
#include <glog/logging.h>

#include "kudu/common/key_range.h"
#include "kudu/gutil/casts.h"
#include "kudu/gutil/endian.h"
#include "kudu/gutil/macros.h"
#include "kudu/gutil/map-util.h"
#include "kudu/gutil/stringprintf.h"
#include "kudu/gutil/strings/substitute.h"
#include "kudu/tablet/rowset.h"
#include "kudu/tablet/rowset_tree.h"
#include "kudu/util/flag_tags.h"
#include "kudu/util/flag_validators.h"
#include "kudu/util/logging.h"
#include "kudu/util/slice.h"
#include "kudu/util/status.h"

using std::shared_ptr;
using std::string;
using std::unordered_map;
using std::vector;

DECLARE_double(compaction_minimum_improvement);
DECLARE_int64(budgeted_compaction_target_rowset_size);

DEFINE_bool(compaction_force_small_rowset_tradeoff, false,
            "Whether to allow the compaction small rowset tradeoff factor to "
            "be larger than the compaction minimum improvement score. Doing so "
            "will have harmful effects on the performance of the tablet "
            "server. Do not set this unless you know what you are doing.");
TAG_FLAG(compaction_force_small_rowset_tradeoff, advanced);
TAG_FLAG(compaction_force_small_rowset_tradeoff, experimental);
TAG_FLAG(compaction_force_small_rowset_tradeoff, runtime);
TAG_FLAG(compaction_force_small_rowset_tradeoff, unsafe);

DEFINE_double(compaction_small_rowset_tradeoff, 0.009,
              "The weight of small rowset compaction compared to "
              "height-based compaction. This value must be less than "
              "-compaction_minimum_improvement to prevent compaction loops. "
              "Only change this if you know what you are doing.");
TAG_FLAG(compaction_small_rowset_tradeoff, advanced);
TAG_FLAG(compaction_small_rowset_tradeoff, experimental);
TAG_FLAG(compaction_small_rowset_tradeoff, runtime);

// Enforce a minimum size of 1MB, since otherwise the knapsack algorithm
// will always pick up small rowsets no matter what.
static const int kMinSizeMb = 1;

namespace kudu {
namespace tablet {

namespace {

bool ValidateSmallRowSetTradeoffVsMinScore() {
  if (FLAGS_compaction_force_small_rowset_tradeoff) {
    return true;
  }
  const auto tradeoff = FLAGS_compaction_small_rowset_tradeoff;
  const auto min_score = FLAGS_compaction_minimum_improvement;
  if (tradeoff >= min_score) {
    LOG(ERROR) << strings::Substitute(
        "-compaction_small_rowset_tradeoff=$0 must be less than "
        "-compaction_minimum_improvement=$1 in order to prevent pointless "
        "compactions; if you know what you are doing, pass "
        "-compaction_force_small_rowset_tradeoff to permit this",
        tradeoff, min_score);
    return false;
  }
  return true;
}
GROUP_FLAG_VALIDATOR(compaction_small_rowset_tradeoff_and_min_score,
                     &ValidateSmallRowSetTradeoffVsMinScore);

// Less-than comparison by minimum key (both by actual encoded key and cdf)
bool LessCDFAndRSMin(const RowSetInfo& a, const RowSetInfo& b) {
  return a.cdf_min_key() < b.cdf_min_key() && a.min_key().compare(b.min_key()) < 0;
}

// Less-than comparison by maximum key (both by actual key slice and cdf)
bool LessCDFAndRSMax(const RowSetInfo& a, const RowSetInfo& b) {
  return a.cdf_max_key() < b.cdf_max_key() && a.max_key().compare(b.max_key()) < 0;
}

// Debug-checks that min <= imin <= imax <= max
void DCheckInside(const Slice& min, const Slice& max,
                  const Slice& imin, const Slice& imax) {
  DCHECK_LE(min, max);
  DCHECK_LE(imin, imax);
  DCHECK_LE(min, imin);
  DCHECK_LE(imax, max);
}

// Return the number of bytes of common prefix shared by 'min' and 'max'
int CommonPrefix(const Slice& min, const Slice& max) {
  int min_len = std::min(min.size(), max.size());
  int common_prefix = 0;
  while (common_prefix < min_len &&
         min[common_prefix] == max[common_prefix]) {
    ++common_prefix;
  }
  return common_prefix;
}

void DCheckCommonPrefix(const Slice& min, const Slice& imin,
                        const Slice& imax, int common_prefix) {
  DCHECK_EQ(memcmp(min.data(), imin.data(), common_prefix), 0)
    << "slices should share common prefix:\n"
    << "\t" << KUDU_REDACT(min.ToDebugString()) << "\n"
    << "\t" << KUDU_REDACT(imin.ToDebugString());
  DCHECK_EQ(memcmp(min.data(), imax.data(), common_prefix), 0)
    << "slices should share common prefix:\n"
    << "\t" << KUDU_REDACT(min.ToDebugString()) << "\n"
    << "\t" << KUDU_REDACT(imin.ToDebugString());
}

uint64_t SliceTailToInt(const Slice& slice, int start) {
  uint64_t ret = 0;
  DCHECK_GE(start, 0);
  DCHECK_LE(start, slice.size());
  memcpy(&ret, &slice.data()[start], std::min(slice.size() - start, sizeof(ret)));
  ret = BigEndian::ToHost64(ret);
  return ret;
}

// Finds fraction (imin, imax) takes up of rs->GetBounds().
// Requires that (imin, imax) is contained in rs->GetBounds().
double StringFractionInRange(const RowSetInfo* rsi,
                             const Slice& imin,
                             const Slice& imax) {
  Slice min(rsi->min_key());
  Slice max(rsi->max_key());
  if (!rsi->has_bounds()) {
    VLOG(2) << "Ignoring " << rsi->rowset()->ToString() << " in CDF calculation";
    return 0;
  }
  DCheckInside(min, max, imin, imax);

  int common_prefix = CommonPrefix(min, max);
  DCheckCommonPrefix(min, imin, imax, common_prefix);

  // Convert the remaining portion of each string to an integer.
  uint64_t min_int = SliceTailToInt(min, common_prefix);
  uint64_t max_int = SliceTailToInt(max, common_prefix);
  uint64_t imin_int = SliceTailToInt(imin, common_prefix);
  uint64_t imax_int = SliceTailToInt(imax, common_prefix);

  // Compute how far between min and max the query point falls.
  if (min_int == max_int) return 0;
  return static_cast<double>(imax_int - imin_int) / (max_int - min_int);
}

// Computes the "width" of an interval [prev, next] according to the amount
// of data estimated to be inside the interval, where this is calculated by
// multiplying the fraction that the interval takes up in the keyspace of
// each rowset by the rowset's size (assumes distribution of rows is somewhat
// uniform).
// Requires: [prev, next] contained in each rowset in "active"
double WidthByDataSize(const Slice& prev, const Slice& next,
                       const unordered_map<RowSet*, RowSetInfo*>& active) {
  double weight = 0;

  for (const auto& rs_rsi : active) {
    double fraction = StringFractionInRange(rs_rsi.second, prev, next);
    weight += rs_rsi.second->base_and_redos_size_bytes() * fraction;
  }

  return weight;
}

// Computes the "width" of an interval as above, for the provided columns in the rowsets.
double WidthByDataSize(const Slice& prev, const Slice& next,
                       const unordered_map<RowSet*, RowSetInfo*>& active,
                       const vector<ColumnId>& col_ids) {
  double weight = 0;

  for (const auto& rs_rsi : active) {
    double fraction = StringFractionInRange(rs_rsi.second, prev, next);
    for (const auto& col_id : col_ids) {
      weight += rs_rsi.second->size_bytes(col_id) * fraction;
    }
  }

  return weight;
}

void CheckCollectOrderedCorrectness(const vector<RowSetInfo>& min_key,
                                    const vector<RowSetInfo>& max_key,
                                    double total_width) {
  CHECK_GE(total_width, 0);
  CHECK_EQ(min_key.size(), max_key.size());
  if (!min_key.empty()) {
    CHECK_EQ(min_key.front().cdf_min_key(), 0.0f);
    CHECK_EQ(max_key.back().cdf_max_key(), total_width);
  }
  DCHECK(std::is_sorted(min_key.begin(), min_key.end(), LessCDFAndRSMin));
  DCHECK(std::is_sorted(max_key.begin(), max_key.end(), LessCDFAndRSMax));
}

double ComputeRowsetValue(double width, uint64_t size_bytes) {
  const auto gamma = FLAGS_compaction_small_rowset_tradeoff;
  const auto target_size_bytes = FLAGS_budgeted_compaction_target_rowset_size;

  // This is an approximation to the expected reduction in rowset count per
  // input rowset. See the compaction policy design doc for more details.
  // The score is floored at 0 to prevent rowsets that are bigger than the
  // target size from adding a negative score that discourages height-based
  // compactions. In extreme circumstances, the overall value could be negative,
  // which violates a knapsack problem invariant.
  const auto size_score =
      std::max(0.0, 1 - static_cast<double>(size_bytes) / target_size_bytes);
  return width + gamma * size_score;
}

} // anonymous namespace

// RowSetInfo class ---------------------------------------------------

void RowSetInfo::Collect(const RowSetTree& tree, vector<RowSetInfo>* rsvec) {
  rsvec->reserve(tree.all_rowsets().size());
  for (const shared_ptr<RowSet>& ptr : tree.all_rowsets()) {
    rsvec->push_back(RowSetInfo(ptr.get(), 0));
  }
}

void RowSetInfo::ComputeCdfAndCollectOrdered(const RowSetTree& tree,
                                             double* rowset_total_height,
                                             double* rowset_total_width,
                                             vector<RowSetInfo>* info_by_min_key,
                                             vector<RowSetInfo>* info_by_max_key) {
  DCHECK((info_by_min_key && info_by_max_key) ||
         (!info_by_min_key && !info_by_max_key))
      << "'info_by_min_key' and 'info_by_max_key' must both be non-null or both be null";

  // The collection process works as follows:
  // For each sorted endpoint, first we identify whether it is a
  // start or stop endpoint.
  //
  // At a start point, the associated rowset is added to the
  // 'active' rowset mapping, allowing us to keep track of the index
  // of the rowset's RowSetInfo in the 'info_by_min_key_tmp' vector.
  //
  // At a stop point, the rowset is removed from the 'active' map.
  // Note that the map allows access to the incomplete RowSetInfo that the
  // RowSet maps to.
  //
  // The height of the tablet replica at the keys in between each successive
  // pair of endpoints is active.size().
  //
  // The algorithm keeps track of its state - a "sliding window"
  // across the keyspace - by maintaining the previous key and current
  // value of the total width traversed over the intervals.
  Slice prev_key;
  unordered_map<RowSet*, RowSetInfo*> active;
  double total_width = 0.0;
  double weighted_height_sum = 0.0;

  // We need to filter out the rowsets that aren't available before we process
  // the endpoints, else there's a race since we see endpoints twice and a delta
  // compaction might finish in between.
  RowSetVector available_rowsets;
  for (const auto& rs : tree.all_rowsets()) {
    if (rs->IsAvailableForCompaction()) {
      available_rowsets.push_back(rs);
    }
  }

  size_t len = available_rowsets.size();
  vector<RowSetInfo> info_by_min_key_tmp;
  vector<RowSetInfo> info_by_max_key_tmp;

  // NB: Since the algorithm will store pointers to elements in these vectors
  // while they grow, the reserve calls are necessary for correctness.
  info_by_min_key_tmp.reserve(len);
  info_by_max_key_tmp.reserve(len);


  RowSetTree available_rs_tree;
  available_rs_tree.Reset(available_rowsets);
  for (const auto& rse : available_rs_tree.key_endpoints()) {
    RowSet* rs = rse.rowset_;
    const Slice& next_key = rse.slice_;
    double interval_width = WidthByDataSize(prev_key, next_key, active);

    // For each active rowset, update the cdf value at the max key and the
    // running total of weighted heights. They will be divided by the
    // appropriate denominators at the end.
    for (const auto& rs_rsi : active) {
      RowSetInfo& cdf_rs = *rs_rsi.second;
      cdf_rs.cdf_max_key_ += interval_width;
    }
    weighted_height_sum += active.size() * interval_width;

    // Move sliding window
    total_width += interval_width;
    prev_key = next_key;

    // Add/remove current RowSetInfo
    if (rse.endpoint_ == RowSetTree::START) {
      info_by_min_key_tmp.push_back(RowSetInfo(rs, total_width));
      // Store reference from vector. This is safe b/c of reserve() above.
      EmplaceOrDie(&active, rs, &info_by_min_key_tmp.back());
    } else if (rse.endpoint_ == RowSetTree::STOP) {
      // If the current rowset is not in the active set, then the rowset tree
      // is inconsistent: an interval STOPs before it STARTs.
      RowSetInfo* cdf_rs = FindOrDie(active, rs);
      CHECK_EQ(cdf_rs->rowset(), rs) << "Inconsistent key interval tree.";
      CHECK_NOTNULL(EraseKeyReturnValuePtr(&active, rs));
      info_by_max_key_tmp.push_back(*cdf_rs);
    } else {
      LOG(FATAL) << "Undefined RowSet endpoint type.\n"
                 << "\tExpected either RowSetTree::START=" << RowSetTree::START
                 << " or RowSetTree::STOP=" << RowSetTree::STOP << ".\n"
                 << "\tRecieved:\n"
                 << "\t\tRowSet=" << rs->ToString() << "\n"
                 << "\t\tKey=" << KUDU_REDACT(next_key.ToDebugString()) << "\n"
                 << "\t\tEndpointType=" << rse.endpoint_;
    }
  }

  CheckCollectOrderedCorrectness(info_by_min_key_tmp,
                                 info_by_max_key_tmp,
                                 total_width);
  FinalizeCDFVector(total_width, &info_by_min_key_tmp);
  FinalizeCDFVector(total_width, &info_by_max_key_tmp);

  if (rowset_total_height && rowset_total_width) {
    *rowset_total_height = weighted_height_sum;
    *rowset_total_width = total_width;
  }

  if (info_by_min_key && info_by_max_key) {
    *info_by_min_key = std::move(info_by_min_key_tmp);
    *info_by_max_key = std::move(info_by_max_key_tmp);
  }
}

void RowSetInfo::SplitKeyRange(const RowSetTree& tree,
                               Slice start_key,
                               Slice stop_key,
                               const std::vector<ColumnId>& col_ids,
                               uint64_t target_chunk_size,
                               vector<KeyRange>* ranges) {
  // check start_key greater than stop_key
  CHECK(stop_key.empty() || start_key <= stop_key);

  // The split process works as follows:
  // For each sorted endpoint, first we identify whether it is a
  // start or stop endpoint.
  //
  // At a start point, the associated rowset is added to the
  // "active" rowset mapping.
  //
  // At a stop point, the rowset is removed from the "active" map.
  // Note that the "active" map allows access to the incomplete
  // RowSetInfo that the RowSet maps to.
  //
  // The algorithm keeps track of its state - a "sliding window"
  // across the keyspace - by maintaining the previous key and current
  // value of the total data size traversed over the intervals.
  vector<RowSetInfo> active_rsi;
  active_rsi.reserve(tree.all_rowsets().size());
  unordered_map<RowSet*, RowSetInfo*> active;
  uint64_t chunk_size = 0;
  Slice last_bound = start_key;
  Slice prev = start_key;
  Slice next;

  for (const auto& rse : tree.key_endpoints()) {
    RowSet* rs = rse.rowset_;
    next = rse.slice_;

    if (prev < next) {
      // reset next when next greater than stop_key
      if (!stop_key.empty() && next > stop_key) {
        next = stop_key;
      }

      uint64_t interval_size = 0;
      if (col_ids.empty()) {
        interval_size = WidthByDataSize(prev, next, active);
      } else {
        interval_size = WidthByDataSize(prev, next, active, col_ids);
      }

      if (chunk_size != 0 && chunk_size + interval_size / 2 >= target_chunk_size) {
        // Select the interval closest to the target chunk size
        ranges->push_back(KeyRange(
            last_bound.ToString(), prev.ToString(), chunk_size));
        last_bound = prev;
        chunk_size = 0;
      }
      chunk_size += interval_size;
      prev = next;
    }

    if (!stop_key.empty() && prev >= stop_key) {
      break;
    }

    // Add/remove current RowSetInfo
    if (rse.endpoint_ == RowSetTree::START) {
      // Store reference from vector. This is safe b/c of reserve() above.
      active_rsi.push_back(RowSetInfo(rs, 0));
      active.insert(std::make_pair(rs, &active_rsi.back()));
    } else if (rse.endpoint_ == RowSetTree::STOP) {
      CHECK_EQ(active.erase(rs), 1);
    } else {
      LOG(FATAL) << "Undefined RowSet endpoint type.\n"
                 << "\tExpected either RowSetTree::START=" << RowSetTree::START
                 << " or RowSetTree::STOP=" << RowSetTree::STOP << ".\n"
                 << "\tRecieved:\n"
                 << "\t\tRowSet=" << rs->ToString() << "\n"
                 << "\t\tKey=" << KUDU_REDACT(next.ToDebugString()) << "\n"
                 << "\t\tEndpointType=" << rse.endpoint_;
    }
  }
  if (stop_key.empty() || last_bound < stop_key) {
    ranges->emplace_back(last_bound.ToString(), stop_key.ToString(), chunk_size);
  }
}

RowSetInfo::RowSetInfo(RowSet* rs, double init_cdf)
    : cdf_min_key_(init_cdf),
      cdf_max_key_(init_cdf),
      value_(0.0),
      density_(0.0),
      extra_(new ExtraData()) {
  extra_->rowset = rs;
  extra_->base_and_redos_size_bytes = rs->OnDiskBaseDataSizeWithRedos();
  extra_->size_bytes = rs->OnDiskSize();
  extra_->has_bounds = rs->GetBounds(&extra_->min_key, &extra_->max_key).ok();
  base_and_redos_size_mb_ =
      std::max(implicit_cast<int>(extra_->base_and_redos_size_bytes / 1024 / 1024),
                                  kMinSizeMb);
}

uint64_t RowSetInfo::size_bytes(const ColumnId& col_id) const {
  return extra_->rowset->OnDiskBaseDataColumnSize(col_id);
}

void RowSetInfo::FinalizeCDFVector(double quot, vector<RowSetInfo>* vec) {
  if (quot == 0) return;
  for (RowSetInfo& cdf_rs : *vec) {
    CHECK_GT(cdf_rs.base_and_redos_size_mb_, 0)
        << "Expected file size to be at least 1MB "
        << "for RowSet " << cdf_rs.rowset()->ToString()
        << ", was " << cdf_rs.base_and_redos_size_bytes()
        << " bytes.";
    cdf_rs.cdf_min_key_ /= quot;
    cdf_rs.cdf_max_key_ /= quot;
    cdf_rs.value_ = ComputeRowsetValue(cdf_rs.width(), cdf_rs.extra_->size_bytes);
    cdf_rs.density_ = cdf_rs.value_ / cdf_rs.base_and_redos_size_mb_;
  }
}

string RowSetInfo::ToString() const {
  string ret;
  ret.append(rowset()->ToString());
  StringAppendF(&ret, "(% 3dM) [%.04f, %.04f]", base_and_redos_size_mb_,
                cdf_min_key_, cdf_max_key_);
  if (extra_->has_bounds) {
    ret.append(" [").append(KUDU_REDACT(Slice(extra_->min_key).ToDebugString()));
    ret.append(",").append(KUDU_REDACT(Slice(extra_->max_key).ToDebugString()));
    ret.append("]");
  }
  return ret;
}

bool RowSetInfo::Intersects(const RowSetInfo &other) const {
  if (other.cdf_min_key() > cdf_max_key()) return false;
  if (other.cdf_max_key() < cdf_min_key()) return false;
  return true;
}

} // namespace tablet
} // namespace kudu