Smaller block size and add _col2LastId to block metadata

1. The block size used to be `1 << 23` (over 8M), which is too large,
   since we always need to decompress at least one whole block, even
   when reading only few triples. It's now 100'000, which still has a
   small relatively small overall space consumption.

2. Add member `_col2LastId` to block data because we need it for the
   delta triples (#916).

src/index/CompressedRelation.cpp

@@ -367,6 +367,7 @@ void CompressedRelationWriter::addRelation(Id col0Id,
     }
     _currentBlockData._col0LastId = col0Id;
     _currentBlockData._col1LastId = col1And2Ids(col1And2Ids.numRows() - 1, 0);
+    _currentBlockData._col2LastId = col1And2Ids(col1And2Ids.numRows() - 1, 1);
     AD_CORRECTNESS_CHECK(_buffer.numColumns() == col1And2Ids.numColumns());
     auto bufferOldSize = _buffer.numRows();
     _buffer.resize(_buffer.numRows() + col1And2Ids.numRows());
@@ -396,7 +397,8 @@ void CompressedRelationWriter::writeRelationToExclusiveBlocks(
 
     _blockBuffer.push_back(CompressedBlockMetadata{
         std::move(offsets), actualNumRowsPerBlock, col0Id, col0Id, data[i][0],
-        data[i + actualNumRowsPerBlock - 1][0]});
+        data[i + actualNumRowsPerBlock - 1][0],
+        data[i + actualNumRowsPerBlock - 1][1]});
   }
 }
 

src/index/CompressedRelation.h

@@ -60,11 +60,21 @@ struct CompressedBlockMetadata {
   // For example, in the PSO permutation, col0 is the P and col1 is the S. The
   // col0 ID is not stored in the block. First and last are meant inclusively,
   // that is, they are both part of the block.
+  //
+  // NOTE: Strictly speaking, we don't need `_col0FirstId` and `_col1FirstId`.
+  // However, they are convenient to have and don't really harm with respect to
+  // space efficiency. For example, for Wikidata, we have only around 50K blocks
+  // with block size 8M and around 5M blocks with block size 80K; even the
+  // latter takes only half a GB in total.
   Id _col0FirstId;
   Id _col0LastId;
   Id _col1FirstId;
   Id _col1LastId;
 
+  // For our `DeltaTriples` (https://github.com/ad-freiburg/qlever/pull/916), we
+  // need to know the least significant `Id` of the last triple as well.
+  Id _col2LastId;
+
   // Two of these are equal if all members are equal.
   bool operator==(const CompressedBlockMetadata&) const = default;
 };
@@ -83,6 +93,7 @@ AD_SERIALIZE_FUNCTION(CompressedBlockMetadata) {
   serializer | arg._col0LastId;
   serializer | arg._col1FirstId;
   serializer | arg._col1LastId;
+  serializer | arg._col2LastId;
 }
 
 // The metadata of a whole compressed "relation", where relation refers to a
@@ -304,6 +315,7 @@ class CompressedRelationReader {
   static void decompressColumn(const std::vector<char>& compressedColumn,
                                size_t numRowsToRead, Iterator iterator);
 
+ public:
   // Read the block that is identified by the `blockMetaData` from the `file`,
   // decompress and return it.
   // If `columnIndices` is `nullopt`, then all columns of the block are read,

src/index/ConstantsIndexBuilding.h

@@ -79,6 +79,10 @@ constexpr size_t QUEUE_SIZE_BEFORE_PARALLEL_PARSING = 10;
 // time
 constexpr size_t QUEUE_SIZE_AFTER_PARALLEL_PARSING = 10;
 
-//  The uncompressed size in bytes of a block of the permutations. Currently 8MB
-//   is chosen which is well suited for zstd compression
-constexpr size_t BLOCKSIZE_COMPRESSED_METADATA = 1ul << 23u;
+// The uncompressed size in bytes of a block of the permutations.
+//
+// NOTE: This used to be `1 << 23` (over 8M), which is fairly large (we always
+// need to decompress at least one whole block, even when reading only few
+// triples). With 100K, the total space for all the `CompressedBlockMetadata` is
+// still small compared to the rest of the index.
+constexpr size_t BLOCKSIZE_COMPRESSED_METADATA = 100'000;