src/kudu/tablet/diskrowset-test.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/diskrowset.h"

#include <algorithm>
#include <cstddef>
#include <cstdint>
#include <memory>
#include <optional>
#include <ostream>
#include <string>
#include <tuple>
#include <type_traits>
#include <unordered_set>
#include <vector>

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

#include "kudu/clock/logical_clock.h"
#include "kudu/common/common.pb.h"
#include "kudu/common/iterator.h"
#include "kudu/common/row.h"
#include "kudu/common/row_changelist.h"
#include "kudu/common/rowid.h"
#include "kudu/common/schema.h"
#include "kudu/common/timestamp.h"
#include "kudu/consensus/log_anchor_registry.h"
#include "kudu/consensus/opid.pb.h"
#include "kudu/consensus/opid_util.h"
#include "kudu/fs/block_id.h"
#include "kudu/fs/io_context.h"
#include "kudu/gutil/ref_counted.h"
#include "kudu/gutil/stringprintf.h"
#include "kudu/gutil/strings/stringpiece.h"
#include "kudu/gutil/strings/substitute.h"
#include "kudu/tablet/compaction.h"
#include "kudu/tablet/delta_key.h"
#include "kudu/tablet/delta_store.h"
#include "kudu/tablet/delta_tracker.h"
#include "kudu/tablet/diskrowset-test-base.h"
#include "kudu/tablet/mvcc.h"
#include "kudu/tablet/rowset.h"
#include "kudu/tablet/tablet-test-util.h"
#include "kudu/tablet/tablet.pb.h"
#include "kudu/tablet/tablet_mem_trackers.h"
#include "kudu/util/bloom_filter.h"
#include "kudu/util/faststring.h"
#include "kudu/util/memory/arena.h"
#include "kudu/util/random.h"
#include "kudu/util/slice.h"
#include "kudu/util/status.h"
#include "kudu/util/stopwatch.h"
#include "kudu/util/test_macros.h"
#include "kudu/util/test_util.h"

namespace kudu {
namespace tablet {
class RowSetMetadata;
}  // namespace tablet
}  // namespace kudu

DEFINE_double(update_fraction, 0.1f, "fraction of rows to update");
DECLARE_bool(cfile_lazy_open);
DECLARE_bool(crash_on_eio);
DECLARE_int32(cfile_default_block_size);
DECLARE_double(env_inject_eio);
DECLARE_double(tablet_delta_store_major_compact_min_ratio);
DECLARE_int32(tablet_delta_store_minor_compact_max);

using std::is_sorted;
using std::make_tuple;
using std::shared_ptr;
using std::string;
using std::unique_ptr;
using std::unordered_set;
using std::tuple;
using std::vector;
using strings::Substitute;

namespace kudu {
namespace tablet {

// TODO: add test which calls CopyNextRows on an iterator with no more
// rows - i think it segfaults!

// Test round-trip writing and reading back a rowset with
// multiple columns. Does not test any modifications.
TEST_F(TestRowSet, TestRowSetRoundTrip) {
  WriteTestRowSet();

  // Now open the DiskRowSet for read
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  // First iterate over all columns
  LOG_TIMING(INFO, "Iterating over all columns") {
    IterateProjection(*rs, schema_, n_rows_);
  }

  // Now iterate only over the key column
  Schema proj_key;
  ASSERT_OK(schema_.CreateProjectionByNames({ "key" }, &proj_key));

  LOG_TIMING(INFO, "Iterating over only key column") {
    IterateProjection(*rs, proj_key, n_rows_);
  }


  // Now iterate only over the non-key column
  Schema proj_val;
  ASSERT_OK(schema_.CreateProjectionByNames({ "val" }, &proj_val));
  LOG_TIMING(INFO, "Iterating over only val column") {
    IterateProjection(*rs, proj_val, n_rows_);
  }

  // Test that CheckRowPresent returns correct results
  Arena arena(64);
  ProbeStats stats;

  // 1. Check a key which comes before all keys in rowset
  {
    Schema pk = schema_.CreateKeyProjection();
    RowBuilder rb(&pk);
    rb.AddString(Slice("h"));
    RowSetKeyProbe probe(rb.row(), &arena);
    bool present;
    ASSERT_OK(rs->CheckRowPresent(probe, nullptr, &present, &stats));
    ASSERT_FALSE(present);
  }

  // 2. Check a key which comes after all keys in rowset
  {
    Schema pk = schema_.CreateKeyProjection();
    RowBuilder rb(&pk);
    rb.AddString(Slice("z"));
    RowSetKeyProbe probe(rb.row(), &arena);
    bool present;
    ASSERT_OK(rs->CheckRowPresent(probe, nullptr, &present, &stats));
    ASSERT_FALSE(present);
  }

  // 3. Check a key which is not present, but comes between present
  // keys
  {
    Schema pk = schema_.CreateKeyProjection();
    RowBuilder rb(&pk);
    rb.AddString(Slice("hello 00000000000049x"));
    RowSetKeyProbe probe(rb.row(), &arena);
    bool present;
    ASSERT_OK(rs->CheckRowPresent(probe, nullptr, &present, &stats));
    ASSERT_FALSE(present);
  }

  // 4. Check a key which is present
  {
    char buf[256];
    FormatKey(49, buf, sizeof(buf));
    Schema pk = schema_.CreateKeyProjection();
    RowBuilder rb(&pk);
    rb.AddString(Slice(buf));
    RowSetKeyProbe probe(rb.row(), &arena);
    bool present;
    ASSERT_OK(rs->CheckRowPresent(probe, nullptr, &present, &stats));
    ASSERT_TRUE(present);
  }
}

// Test writing a rowset, and then updating some rows in it.
TEST_F(TestRowSet, TestRowSetUpdate) {
  Arena arena(64);

  WriteTestRowSet();

  // Now open the DiskRowSet for read
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  // Add an update to the delta tracker for a number of keys
  // which exist. These updates will change the value to
  // equal idx*5 (whereas in the original data, value = idx)
  unordered_set<uint32_t> updated;
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, &updated);
  ASSERT_EQ(static_cast<int>(n_rows_ * FLAGS_update_fraction),
            rs->delta_tracker_->dms_->Count());

  // Try to add a mutation for a key not in the file (but which falls
  // between two valid keys)
  faststring buf;
  RowChangeListEncoder enc(&buf);
  enc.SetToDelete();

  Timestamp timestamp(0);
  Schema proj_key = schema_.CreateKeyProjection();
  RowBuilder rb(&proj_key);
  rb.AddString(Slice("hello 00000000000049x"));
  RowSetKeyProbe probe(rb.row(), &arena);

  OperationResultPB result;
  ProbeStats stats;
  Status s = rs->MutateRow(timestamp, probe, enc.as_changelist(), op_id_, nullptr, &stats, &result);
  ASSERT_TRUE(s.IsNotFound()) << s.ToString();
  ASSERT_EQ(0, result.mutated_stores_size());

  // Now read back the value column, and verify that the updates
  // are visible.
  VerifyUpdates(*rs, updated);
}

TEST_F(TestRowSet, TestErrorDuringUpdate) {
  Arena arena(64);

  WriteTestRowSet();
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  faststring buf;
  RowChangeListEncoder enc(&buf);
  enc.SetToDelete();

  // Get a row that we expect to be in the rowset.
  Timestamp timestamp(0);
  Schema proj_key = schema_.CreateKeyProjection();
  RowBuilder rb(&proj_key);
  rb.AddString(Slice("hello 000000000000050"));
  RowSetKeyProbe probe(rb.row(), &arena);

  // But fail while reading it!
  FLAGS_crash_on_eio = false;
  FLAGS_env_inject_eio = 1.0;

  // The mutation should result in an IOError.
  OperationResultPB result;
  ProbeStats stats;
  Status s = rs->MutateRow(timestamp, probe, enc.as_changelist(), op_id_, nullptr, &stats, &result);
  LOG(INFO) << s.ToString();
  ASSERT_TRUE(s.IsIOError()) << s.ToString();
}

TEST_F(TestRowSet, TestRandomRead) {
  // Write 100 rows.
  WriteTestRowSet(100);
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  // Read un-updated row.
  NO_FATALS(VerifyRandomRead(*rs, "hello 000000000000050",
                             R"((string key="hello 000000000000050", uint32 val=50))"));

  // Update the row.
  OperationResultPB result;
  ASSERT_OK(UpdateRow(rs.get(), 50, 12345, &result));

  // Read it again -- should see the updated value.
  NO_FATALS(VerifyRandomRead(*rs, "hello 000000000000050",
                             R"((string key="hello 000000000000050", uint32 val=12345))"));

  // Try to read a row which comes before the first key.
  // This should return no rows.
  NO_FATALS(VerifyRandomRead(*rs, "aaaaa", ""));

  // Same with a row which falls between keys.
  NO_FATALS(VerifyRandomRead(*rs, "hello 000000000000050_between_keys", ""));

  // And a row which falls after the last key.
  NO_FATALS(VerifyRandomRead(*rs, "hello 000000000000101", ""));
}

// Test Delete() support within a DiskRowSet.
TEST_F(TestRowSet, TestDelete) {
  // Write and open a DiskRowSet with 2 rows.
  WriteTestRowSet(2);
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));
  MvccSnapshot snap_before_delete(mvcc_);

  // Delete one of the two rows
  OperationResultPB result;
  ASSERT_OK(DeleteRow(rs.get(), 0, &result));
  ASSERT_EQ(1, result.mutated_stores_size());
  ASSERT_EQ(0L, result.mutated_stores(0).rs_id());
  ASSERT_EQ(0L, result.mutated_stores(0).dms_id());
  MvccSnapshot snap_after_delete(mvcc_);

  vector<string> rows;
  Status s;

  RowIteratorOptions opts;
  opts.projection = &schema_;
  for (int i = 0; i < 2; i++) {
    // Reading the MVCC snapshot prior to deletion should show the row.
    opts.snap_to_include = snap_before_delete;
    ASSERT_OK(DumpRowSet(*rs, opts, &rows));
    ASSERT_EQ(2, rows.size());
    EXPECT_EQ(R"((string key="hello 000000000000000", uint32 val=0))", rows[0]);
    EXPECT_EQ(R"((string key="hello 000000000000001", uint32 val=1))", rows[1]);

    // Reading the MVCC snapshot after the deletion should hide the row.
    opts.snap_to_include = snap_after_delete;
    ASSERT_OK(DumpRowSet(*rs, opts, &rows));
    ASSERT_EQ(1, rows.size());
    EXPECT_EQ(R"((string key="hello 000000000000001", uint32 val=1))", rows[0]);

    // Trying to delete or update the same row again should fail.
    OperationResultPB result;
    s = DeleteRow(rs.get(), 0, &result);
    ASSERT_TRUE(s.IsNotFound()) << "bad status: " << s.ToString();
    ASSERT_EQ(0, result.mutated_stores_size());
    result.Clear();
    s = UpdateRow(rs.get(), 0, 12345, &result);
    ASSERT_TRUE(s.IsNotFound()) << "bad status: " << s.ToString();
    ASSERT_EQ(0, result.mutated_stores_size());

    // CheckRowPresent should return false.
    bool present;
    ASSERT_OK(CheckRowPresent(*rs, 0, &present));
    EXPECT_FALSE(present);

    if (i == 1) {
      // Flush DMS. The second pass through the loop will re-verify that the
      // externally visible state of the layer has not changed.
      // deletions now in a DeltaFile.
      ASSERT_OK(rs->FlushDeltas(nullptr));
    }
  }
}


TEST_F(TestRowSet, TestDMSFlush) {
  WriteTestRowSet();

  unordered_set<uint32_t> updated;

  // Now open the DiskRowSet for read
  {
    shared_ptr<DiskRowSet> rs;
    ASSERT_OK(OpenTestRowSet(&rs));

    // Add an update to the delta tracker for a number of keys
    // which exist. These updates will change the value to
    // equal idx*5 (whereas in the original data, value = idx)
    UpdateExistingRows(rs.get(), FLAGS_update_fraction, &updated);
    ASSERT_EQ(static_cast<int>(n_rows_ * FLAGS_update_fraction),
              rs->delta_tracker_->dms_->Count());

    ASSERT_OK(rs->FlushDeltas(nullptr));

    // Check that the DiskRowSet's DMS has not been initialized.
    ASSERT_FALSE(rs->delta_tracker_->dms_);

    // Now read back the value column, and verify that the updates
    // are visible.
    SCOPED_TRACE("before reopen");
    VerifyUpdates(*rs, updated);
  }

  LOG(INFO) << "Reopening rowset ===============";
  // Close and re-open the rowset and ensure that the updates were
  // persistent.
  {
    shared_ptr<DiskRowSet> rs;
    ASSERT_OK(OpenTestRowSet(&rs));

    // Now read back the value column, and verify that the updates
    // are visible.
    SCOPED_TRACE("after reopen");
    VerifyUpdates(*rs, updated);
  }
}

// Test that when a single row is updated multiple times, we can query the
// historical values using MVCC, even after it is flushed.
TEST_F(TestRowSet, TestFlushedUpdatesRespectMVCC) {
  Arena arena(64);
  const Slice key_slice("row");

  // Write a single row into a new DiskRowSet.
  LOG_TIMING(INFO, "Writing rowset") {
    DiskRowSetWriter drsw(rowset_meta_.get(), &schema_,
                          BloomFilterSizing::BySizeAndFPRate(32*1024, 0.01f));

    ASSERT_OK(drsw.Open());

    RowBuilder rb(&schema_);
    rb.AddString(key_slice);
    rb.AddUint32(1);
    ASSERT_OK_FAST(WriteRow(rb.data(), &drsw));
    ASSERT_OK(drsw.Finish());
  }


  // Reopen the rowset.
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  // Take a snapshot of the pre-update state.
  vector<MvccSnapshot> snaps;
  snaps.emplace_back(mvcc_);


  // Update the single row multiple times, taking an MVCC snapshot
  // after each update.
  faststring update_buf;
  RowChangeListEncoder update(&update_buf);
  for (uint32_t i = 2; i <= 5; i++) {
    {
      arena.Reset();
      ScopedOp op(&mvcc_, clock_.Now());
      op.StartApplying();
      update.Reset();
      update.AddColumnUpdate(schema_.column(1), schema_.column_id(1), &i);
      Schema proj_key = schema_.CreateKeyProjection();
      RowBuilder rb(&proj_key);
      rb.AddString(key_slice);
      RowSetKeyProbe probe(rb.row(), &arena);
      OperationResultPB result;
      ProbeStats stats;
      ASSERT_OK_FAST(rs->MutateRow(op.timestamp(),
                                   probe,
                                   RowChangeList(update_buf),
                                   op_id_,
                                   nullptr,
                                   &stats,
                                   &result));
      ASSERT_EQ(1, result.mutated_stores_size());
      ASSERT_EQ(0L, result.mutated_stores(0).rs_id());
      ASSERT_EQ(0L, result.mutated_stores(0).dms_id());
      op.FinishApplying();
    }
    snaps.emplace_back(mvcc_);
  }

  // Ensure that MVCC is respected by reading the value at each of the stored
  // snapshots.
  ASSERT_EQ(5, snaps.size());
  for (int i = 0; i < 5; i++) {
    SCOPED_TRACE(i);
    RowIteratorOptions opts;
    opts.projection = &schema_;
    opts.snap_to_include = snaps[i];
    unique_ptr<RowwiseIterator> iter;
    ASSERT_OK(rs->NewRowIterator(opts, &iter));
    string data = InitAndDumpIterator(iter.get());
    EXPECT_EQ(StringPrintf(R"((string key="row", uint32 val=%d))", i + 1), data);
  }

  // Flush deltas to disk and ensure that the historical versions are still
  // accessible.
  ASSERT_OK(rs->FlushDeltas(nullptr));

  for (int i = 0; i < 5; i++) {
    SCOPED_TRACE(i);
    RowIteratorOptions opts;
    opts.projection = &schema_;
    opts.snap_to_include = snaps[i];
    unique_ptr<RowwiseIterator> iter;
    ASSERT_OK(rs->NewRowIterator(opts, &iter));
    string data = InitAndDumpIterator(iter.get());
    EXPECT_EQ(StringPrintf(R"((string key="row", uint32 val=%d))", i + 1), data);
  }
}

// Similar to TestDMSFlush above, except does not actually verify
// the results (since the verification step is expensive). Additionally,
// loops the "read" side of the benchmark a number of times, so that
// the speed of applying deltas during read can be micro-benchmarked.
//
// This is most usefully run with an invocation like:
// ./rowset-test --gtest_filter=\*Performance --roundtrip_num_rows=1000000
//    --n_read_passes=1000 --update_fraction=0.01
TEST_F(TestRowSet, TestDeltaApplicationPerformance) {
  WriteTestRowSet();

  // Now open the DiskRowSet for read
  {
    shared_ptr<DiskRowSet> rs;
    ASSERT_OK(OpenTestRowSet(&rs));

    BenchmarkIterationPerformance(*rs.get(),
      StringPrintf("Reading %zd rows prior to updates %d times",
                   n_rows_, FLAGS_n_read_passes));

    UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);

    BenchmarkIterationPerformance(*rs.get(),
      StringPrintf("Reading %zd rows with %.2f%% updates %d times (updates in DMS)",
                   n_rows_, FLAGS_update_fraction * 100.0f,
                   FLAGS_n_read_passes));
    ASSERT_OK(rs->FlushDeltas(nullptr));

    BenchmarkIterationPerformance(*rs.get(),
      StringPrintf("Reading %zd rows with %.2f%% updates %d times (updates on disk)",
                   n_rows_, FLAGS_update_fraction * 100.0f,
                   FLAGS_n_read_passes));
  }
}

TEST_F(TestRowSet, TestRollingDiskRowSetWriter) {
  // Set small block size so that we can roll frequently. Otherwise
  // we couldn't output such small files.
  google::FlagSaver saver;
  FLAGS_cfile_default_block_size = 4096;

  RollingDiskRowSetWriter writer(tablet()->metadata(), schema_,
                                 BloomFilterSizing::BySizeAndFPRate(32*1024, 0.01f),
                                 64 * 1024); // roll every 64KB
  DoWriteTestRowSet(FLAGS_roundtrip_num_rows, &writer);

  // Should have rolled 4 times.
  vector<shared_ptr<RowSetMetadata> > metas;
  writer.GetWrittenRowSetMetadata(&metas);
  EXPECT_EQ(4, metas.size());
  int64_t count = 0;
  for (const shared_ptr<RowSetMetadata>& meta : metas) {
    ASSERT_TRUE(meta->HasDataForColumnIdForTests(schema_.column_id(0)));
    count += meta->live_row_count();
  }
  ASSERT_EQ(FLAGS_roundtrip_num_rows, count);
}

TEST_F(TestRowSet, TestMakeDeltaIteratorMergerUnlocked) {
  WriteTestRowSet();

  // Now open the DiskRowSet for read
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));
  DeltaTracker *dt = rs->delta_tracker();
  int num_stores = dt->redo_delta_stores_.size();
  vector<shared_ptr<DeltaStore> > compacted_stores;
  vector<BlockId> compacted_blocks;
  unique_ptr<DeltaIterator> merge_iter;
  ASSERT_OK(dt->MakeDeltaIteratorMergerUnlocked(nullptr, 0, num_stores - 1, &schema_,
                                                &compacted_stores,
                                                &compacted_blocks, &merge_iter));
  vector<string> results;
  ASSERT_OK(DebugDumpDeltaIterator(REDO, merge_iter.get(), schema_,
                                          ITERATE_OVER_ALL_ROWS,
                                          &results));
  for (const string &str : results) {
    VLOG(1) << str;
  }
  ASSERT_EQ(compacted_stores.size(), num_stores);
  ASSERT_EQ(compacted_blocks.size(), num_stores);
  ASSERT_TRUE(is_sorted(results.begin(), results.end()));
}

void BetweenZeroAndOne(double to_check) {
  ASSERT_LT(0, to_check);
  ASSERT_GT(1, to_check);
}

TEST_F(TestRowSet, TestCompactStores) {
  // With this setting, we want major compactions to basically always have a score.
  FLAGS_tablet_delta_store_major_compact_min_ratio = 0.0001;
  // With this setting, the perf improvement will be 0 until we have two files, at which point
  // it will be the expected ratio, then with three files we get the maximum improvement.
  FLAGS_tablet_delta_store_minor_compact_max = 3;
  // Turning this off so that we can call DeltaStoresCompactionPerfImprovementScore without having
  // to open the files after creating them.
  FLAGS_cfile_lazy_open = false;


  WriteTestRowSet();
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));
  ASSERT_EQ(0, rs->DeltaStoresCompactionPerfImprovementScore(RowSet::MINOR_DELTA_COMPACTION));
  ASSERT_EQ(0, rs->DeltaStoresCompactionPerfImprovementScore(RowSet::MAJOR_DELTA_COMPACTION));

  // Write a first delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));
  // One file isn't enough for minor compactions, but a major compaction can run.
  ASSERT_EQ(0, rs->DeltaStoresCompactionPerfImprovementScore(RowSet::MINOR_DELTA_COMPACTION));
  NO_FATALS(BetweenZeroAndOne(rs->DeltaStoresCompactionPerfImprovementScore(
      RowSet::MAJOR_DELTA_COMPACTION)));

  // Write a second delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));
  // Two files is enough for all delta compactions.
  NO_FATALS(BetweenZeroAndOne(rs->DeltaStoresCompactionPerfImprovementScore(
      RowSet::MINOR_DELTA_COMPACTION)));
  NO_FATALS(BetweenZeroAndOne(rs->DeltaStoresCompactionPerfImprovementScore(
      RowSet::MAJOR_DELTA_COMPACTION)));

  // Write a third delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));
  // We're hitting the max for minor compactions but not for major compactions.
  ASSERT_EQ(1, rs->DeltaStoresCompactionPerfImprovementScore(RowSet::MINOR_DELTA_COMPACTION));
  NO_FATALS(BetweenZeroAndOne(rs->DeltaStoresCompactionPerfImprovementScore(
      RowSet::MAJOR_DELTA_COMPACTION)));

  // Compact the deltafiles
  DeltaTracker *dt = rs->delta_tracker();
  int num_stores = dt->redo_delta_stores_.size();
  VLOG(1) << "Number of stores before compaction: " << num_stores;
  ASSERT_EQ(num_stores, 3);
  ASSERT_OK(dt->CompactStores(nullptr, 0, num_stores - 1));
  num_stores = dt->redo_delta_stores_.size();
  VLOG(1) << "Number of stores after compaction: " << num_stores;
  ASSERT_EQ(1,  num_stores);
  // Back to one store, can't minor compact.
  ASSERT_EQ(0, rs->DeltaStoresCompactionPerfImprovementScore(RowSet::MINOR_DELTA_COMPACTION));
  NO_FATALS(BetweenZeroAndOne(rs->DeltaStoresCompactionPerfImprovementScore(
      RowSet::MAJOR_DELTA_COMPACTION)));

  // Verify that the resulting deltafile is valid
  vector<shared_ptr<DeltaStore> > compacted_stores;
  vector<BlockId> compacted_blocks;
  unique_ptr<DeltaIterator> merge_iter;
  ASSERT_OK(dt->MakeDeltaIteratorMergerUnlocked(nullptr, 0, num_stores - 1, &schema_,
                                                &compacted_stores,
                                                &compacted_blocks, &merge_iter));
  vector<string> results;
  ASSERT_OK(DebugDumpDeltaIterator(REDO, merge_iter.get(), schema_,
                                   ITERATE_OVER_ALL_ROWS,
                                   &results));
  for (const string &str : results) {
    VLOG(1) << str;
  }
  ASSERT_TRUE(is_sorted(results.begin(), results.end()));
}

TEST_F(TestRowSet, TestGCAncientStores) {
  // Disable lazy open so that major delta compactions don't require manual REDO initialization.
  FLAGS_cfile_lazy_open = false;

  // Write some base data.
  // Note: Our test methods here don't write UNDO delete deltas.
  WriteTestRowSet();
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));
  DeltaTracker *dt = rs->delta_tracker();
  ASSERT_EQ(0, dt->CountUndoDeltaStores());
  ASSERT_EQ(0, dt->CountRedoDeltaStores());

  // Write and flush a new REDO delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));
  ASSERT_EQ(0, dt->CountUndoDeltaStores());
  ASSERT_EQ(1, dt->CountRedoDeltaStores());

  // Convert the REDO delta to an UNDO delta.
  ASSERT_OK(rs->MajorCompactDeltaStores(nullptr, HistoryGcOpts::Disabled()));
  ASSERT_EQ(1, dt->CountUndoDeltaStores()); // From doing the major delta compaction.
  ASSERT_EQ(0, dt->CountRedoDeltaStores());

  // Delete all the UNDO deltas. There shouldn't be any delta stores left.
  int64_t blocks_deleted;
  int64_t bytes_deleted;
  ASSERT_OK(dt->DeleteAncientUndoDeltas(clock_.Now(), nullptr, &blocks_deleted, &bytes_deleted));
  ASSERT_GT(blocks_deleted, 0);
  ASSERT_GT(bytes_deleted, 0);
  ASSERT_EQ(0, dt->CountUndoDeltaStores());
  ASSERT_EQ(0, dt->CountRedoDeltaStores());
}

TEST_F(TestRowSet, TestDiskSizeEstimation) {
  // Force the files to be opened so the stats are read.
  FLAGS_cfile_lazy_open = false;

  // Write a rowset.
  WriteTestRowSet();
  shared_ptr<DiskRowSet> rs;
  ASSERT_OK(OpenTestRowSet(&rs));

  // Write a first delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));

  // The rowset consists of the cfile set and REDO deltas, so the rowset's
  // on-disk size and the sum of the cfile set and REDO sizes should equal.
  DiskRowSetSpace drss;
  rs->GetDiskRowSetSpaceUsage(&drss);
  ASSERT_GT(drss.redo_deltas_size, 0);
  ASSERT_EQ(rs->OnDiskSize(), drss.CFileSetOnDiskSize() + drss.redo_deltas_size);

  // Convert the REDO delta to an UNDO delta.
  ASSERT_OK(rs->MajorCompactDeltaStores(nullptr, HistoryGcOpts::Disabled()));

  // REDO size should be zero, but there should be UNDOs, so the on-disk size
  // of the rowset should be the sum of the cfile set and UNDO sizes.
  rs->GetDiskRowSetSpaceUsage(&drss);
  ASSERT_GT(drss.undo_deltas_size, 0);
  ASSERT_EQ(rs->OnDiskSize(), drss.CFileSetOnDiskSize() + drss.undo_deltas_size);

  // Write a second delta file.
  UpdateExistingRows(rs.get(), FLAGS_update_fraction, nullptr);
  ASSERT_OK(rs->FlushDeltas(nullptr));

  // Now there's base data, REDOs, and UNDOs.
  rs->GetDiskRowSetSpaceUsage(&drss);
  ASSERT_GT(drss.undo_deltas_size, 0);
  ASSERT_GT(drss.redo_deltas_size, 0);
}

class DiffScanRowSetTest : public KuduRowSetTest,
                           public ::testing::WithParamInterface<tuple<bool, bool>> {
 public:
  DiffScanRowSetTest()
      : KuduRowSetTest(CreateTestSchema()),
        op_id_(consensus::MaximumOpId()),
        clock_(Timestamp::kInitialTimestamp) {
  }

 protected:
  static Schema CreateTestSchema() {
    SchemaBuilder builder;
    CHECK_OK(builder.AddKeyColumn("key", UINT32));
    CHECK_OK(builder.AddColumn("val1", UINT32));
    CHECK_OK(builder.AddColumn("val2", UINT32));
    return builder.BuildWithoutIds();
  }

  consensus::OpId op_id_;
  clock::LogicalClock clock_;
  MvccManager mvcc_;
};

// Tests the Cartesian product of two boolean parameters:
// 1. Whether to include deleted rows in the scan.
// 2. Whether to include the "is deleted" virtual column in the scan's projection.
INSTANTIATE_TEST_SUITE_P(RowIteratorOptionsPermutations, DiffScanRowSetTest,
                         ::testing::Combine(::testing::Bool(),
                                            ::testing::Bool()));

// Tests diff scans on a diskrowset. The rowset is generated with only a handful
// of rows, but we randomly flush/compact between each update operation so that
// the test operates on a variety of different on-disk and in-memory layouts.
TEST_P(DiffScanRowSetTest, TestFuzz) {
  fs::IOContext test_context;
  scoped_refptr<log::LogAnchorRegistry> log_anchor_registry(new log::LogAnchorRegistry());

  // Create and open a DRS with four rows.
  shared_ptr<DiskRowSet> rs;
  {
    DiskRowSetWriter drsw(rowset_meta_.get(), &schema_,
                          BloomFilterSizing::BySizeAndFPRate(32 * 1024, 0.01f));
    ASSERT_OK(drsw.Open());

    RowBuilder rb(&schema_);
    for (int i = 0; i < 4; i++) {
      rb.Reset();
      rb.AddUint32(i);
      rb.AddUint32(i * 2);
      rb.AddUint32(i * 3);
      ASSERT_OK(WriteRow(rb.data(), &drsw));
    }
    ASSERT_OK(drsw.Finish());
    ASSERT_OK(DiskRowSet::Open(rowset_meta_, log_anchor_registry.get(),
                               TabletMemTrackers(), &test_context, &rs));
  }

  // Run the diff scan test. Scan time boundaries are given by 'ts1_val' and
  // 'ts2_val'. The expected results are given by 'expected_rows'.
  using RowTuple = tuple<uint32_t, uint32_t, uint32_t, bool>;
  auto run_test = [&](uint64_t ts1_val,
                      uint64_t ts2_val,
                      vector<RowTuple> expected_rows) {
    bool include_deleted_rows = std::get<0>(GetParam());
    bool add_vc_is_deleted = std::get<1>(GetParam());

    // Create a projection of the schema, adding the IS_DELETED virtual
    // column if desired.
    vector<ColumnSchema> col_schemas;
    vector<ColumnId> col_ids;
    for (int i = 0; i < schema_.num_columns(); i++) {
      col_schemas.push_back(schema_.column(i));
      col_ids.push_back(schema_.column_id(i));
    }
    if (add_vc_is_deleted) {
      bool read_default = false;
      col_schemas.emplace_back("is_deleted", IS_DELETED, /*is_nullable=*/ false,
                               /*is_immutable=*/ false, &read_default);
      col_ids.emplace_back(schema_.max_col_id() + 1);
    }
    Schema projection(col_schemas, col_ids, 1);

    // Set up the iterator.
    Timestamp ts1(ts1_val);
    Timestamp ts2(ts2_val);
    SCOPED_TRACE(Substitute("Scanning at $0,$1 with schema $2",
                            ts1.ToString(), ts2.ToString(), projection.ToString()));
    RowIteratorOptions opts;
    opts.include_deleted_rows = include_deleted_rows;
    opts.projection = &projection;
    opts.snap_to_exclude = MvccSnapshot(ts1);
    opts.snap_to_include = MvccSnapshot(ts2);
    unique_ptr<RowwiseIterator> iter;
    ASSERT_OK(rs->NewRowIterator(opts, &iter));
    ASSERT_OK(iter->Init(nullptr));

    // Scan the data out of the iterator and test it.
    vector<string> lines;
    ASSERT_OK(IterateToStringList(iter.get(), &lines));
    if (!include_deleted_rows) {
      vector<RowTuple> without_deleted_rows;
      for (const auto& e : expected_rows) {
        if (!std::get<3>(e)) {
          without_deleted_rows.emplace_back(e);
        }
      }
      expected_rows = without_deleted_rows;
    }
    ASSERT_EQ(expected_rows.size(), lines.size()) << lines;
    for (int i = 0; i < expected_rows.size(); i++) {
      string expected_is_deleted = add_vc_is_deleted ? Substitute(
          ", is_deleted is_deleted=$0", std::get<3>(expected_rows[i])) : "";
      string expected_line = Substitute("(uint32 key=$0, "
                                        "uint32 val1=$1, "
                                        "uint32 val2=$2$3)",
                                        std::get<0>(expected_rows[i]),
                                        std::get<1>(expected_rows[i]),
                                        std::get<2>(expected_rows[i]),
                                        expected_is_deleted);
      ASSERT_EQ(expected_line, lines[i]);
    }
  };

  // Mutate a particular column in the diskrowset given by 'row_idx' and
  // 'col_idx'. If 'val' is unset, deletes the row.
  auto mutate_row = [&](rowid_t row_idx,
                        size_t col_idx,
                        std::optional<uint32_t> val) {
    // Build the mutation.
    faststring buf;
    RowChangeListEncoder enc(&buf);
    if (val) {
      enc.AddColumnUpdate(schema_.column(col_idx),
                          schema_.column_id(col_idx),
                          &(*val));
    } else {
      enc.SetToDelete();
    }

    // Build the row key.
    Schema proj_key = schema_.CreateKeyProjection();
    RowBuilder rb(&proj_key);
    rb.AddUint32(row_idx);
    Arena arena(64);
    RowSetKeyProbe probe(rb.row(), &arena);

    // Apply the mutation.
    ScopedOp op(&mvcc_, clock_.Now());
    op.StartApplying();
    ProbeStats stats;
    OperationResultPB result;
    ASSERT_OK(rs->MutateRow(op.timestamp(), probe, enc.as_changelist(), op_id_,
                            &test_context, &stats, &result));
    op.FinishApplying();
  };

  Random prng(SeedRandom());

  // Possibly apply a randomly chosen delta flush or compaction operation.
  auto maybe_flush_compact = [&]() {
    // 25% for a minor delta compaction.
    // 25% for a major delta compaction.
    // 50% for a DMS flush.
    int r = prng.Uniform(4);
    if (r == 0) {
      ASSERT_OK(rs->MinorCompactDeltaStores(&test_context));
    } else if (r == 1) {
      ASSERT_OK(rs->MajorCompactDeltaStores(&test_context, HistoryGcOpts::Disabled()));
    } else {
      ASSERT_OK(rs->FlushDeltas(&test_context));
    }
  };

  // Update the rows in the diskrowset.
  NO_FATALS(maybe_flush_compact());
  NO_FATALS(mutate_row(1, 2, 1000)); // ts 1
  NO_FATALS(maybe_flush_compact());
  NO_FATALS(mutate_row(1, 1, 200)); // ts 2
  NO_FATALS(maybe_flush_compact());
  NO_FATALS(mutate_row(2, 1, 300)); // ts 3
  NO_FATALS(maybe_flush_compact());
  NO_FATALS(mutate_row(3, 1, 400)); // ts 4
  NO_FATALS(maybe_flush_compact());
  NO_FATALS(mutate_row(3, 1, std::nullopt)); // ts 5
  NO_FATALS(maybe_flush_compact());

  // Run the diff scan test on every permutation of time bounds.
  //
  // Note: a regular DRS would include an UNDO DELETE for every INSERT by virtue
  // of having been flushed from an MRS. That's not the case here, and as a
  // result, the insertions themselves aren't associated with any timestamps and
  // thus are not captured by these diff scans.

  // The time bounds specify an empty range so nothing is captured.
  for (int i = 0; i < 7; i++) {
    NO_FATALS(run_test(i, i, {}));
  }

  NO_FATALS(run_test(0, 1, {}));
  NO_FATALS(run_test(1, 2, { make_tuple(1, 2, 1000, false) }));
  NO_FATALS(run_test(2, 3, { make_tuple(1, 200, 1000, false) }));
  NO_FATALS(run_test(3, 4, { make_tuple(2, 300, 6, false) }));
  NO_FATALS(run_test(4, 5, { make_tuple(3, 400, 9, false) }));
  NO_FATALS(run_test(5, 6, { make_tuple(3, 400, 9, true) }));

  NO_FATALS(run_test(0, 2, { make_tuple(1, 2, 1000, false) }));
  NO_FATALS(run_test(1, 3, { make_tuple(1, 200, 1000, false) }));
  NO_FATALS(run_test(2, 4, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false) }));
  NO_FATALS(run_test(3, 5, { make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, false) }));
  NO_FATALS(run_test(4, 6, { make_tuple(3, 400, 9, true) }));

  NO_FATALS(run_test(0, 3, { make_tuple(1, 200, 1000, false) }));
  NO_FATALS(run_test(1, 4, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false) }));
  NO_FATALS(run_test(2, 5, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, false) }));
  NO_FATALS(run_test(3, 6, { make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, true) }));

  NO_FATALS(run_test(0, 4, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false) }));
  NO_FATALS(run_test(1, 5, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, false) }));
  NO_FATALS(run_test(2, 6, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, true) }));

  NO_FATALS(run_test(0, 5, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, false) }));
  NO_FATALS(run_test(1, 6, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, true) }));

  NO_FATALS(run_test(0, 6, { make_tuple(1, 200, 1000, false),
                             make_tuple(2, 300, 6, false),
                             make_tuple(3, 400, 9, true) }));
}

} // namespace tablet
} // namespace kudu