SERVER-65192 Implement ColumnShredder to extract paths and cells for a document

1 files changed, 595 insertions, 0 deletions

diff --git a/src/mongo/db/index/column_key_generator.cpp b/src/mongo/db/index/column_key_generator.cpp
new file mode 100644
index 00000000000..6e348179da8
--- /dev/null
+++ b/src/mongo/db/index/column_key_generator.cpp
@@ -0,0 +1,595 @@
+/**
+ *    Copyright (C) 2022-present MongoDB, Inc.
+ *
+ *    This program is free software: you can redistribute it and/or modify
+ *    it under the terms of the Server Side Public License, version 1,
+ *    as published by MongoDB, Inc.
+ *
+ *    This program is distributed in the hope that it will be useful,
+ *    but WITHOUT ANY WARRANTY; without even the implied warranty of
+ *    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ *    Server Side Public License for more details.
+ *
+ *    You should have received a copy of the Server Side Public License
+ *    along with this program. If not, see
+ *    <http://www.mongodb.com/licensing/server-side-public-license>.
+ *
+ *    As a special exception, the copyright holders give permission to link the
+ *    code of portions of this program with the OpenSSL library under certain
+ *    conditions as described in each individual source file and distribute
+ *    linked combinations including the program with the OpenSSL library. You
+ *    must comply with the Server Side Public License in all respects for
+ *    all of the code used other than as permitted herein. If you modify file(s)
+ *    with this exception, you may extend this exception to your version of the
+ *    file(s), but you are not obligated to do so. If you do not wish to do so,
+ *    delete this exception statement from your version. If you delete this
+ *    exception statement from all source files in the program, then also delete
+ *    it in the license file.
+ */
+
+#define MONGO_LOGV2_DEFAULT_COMPONENT ::mongo::logv2::LogComponent::kIndex
+
+#include "mongo/platform/basic.h"
+
+#include "mongo/db/index/column_key_generator.h"
+
+#include <iomanip>
+#include <ostream>
+#include <vector>
+
+#include "mongo/db/storage/column_store.h"
+#include "mongo/util/functional.h"
+#include "mongo/util/string_map.h"
+
+namespace mongo::column_keygen {
+namespace {
+
+/**
+ * Special handling for vals because BSONElement doesn't support ==, and we want binaryEqualValues
+ * rather than woCompare equality anyway.
+ */
+bool identicalBSONElementArrays(const std::vector<BSONElement>& lhs,
+                                const std::vector<BSONElement>& rhs) {
+    if (lhs.size() != rhs.size())
+        return false;
+    for (size_t i = 0; i < lhs.size(); i++) {
+        if (!lhs[i].binaryEqualValues(rhs[i]))
+            return false;
+    }
+    return true;
+}
+
+/**
+ * This class handles the logic of key generation for columnar indexes. It produces
+ * UnencodedCellViews, which have all of the data that should be put in the index values, but it is
+ * not responsible for encoding that data into a flat buffer. That is handled by XXX.
+ *
+ * TODO once the code for that that is written (SERVER-64766) update this to replace the XXX.
+ *
+ * "Shredding" is an informal term for taking a single BSON document and splitting it into the data
+ * for each unique path. The data at each path should be sufficient to reconstruct the object in
+ * full fidelity, possibly considering parent paths. When determining the path, array index are not
+ * considered, only field names are.
+ *
+ * This class is private to this file, and only exposed via the free-functions in the header.
+ */
+class ColumnShredder {
+public:
+    /**
+     * Option to constructor to avoid overhead when you only need to know about the paths in an
+     * object, such as for deletes.
+     */
+    enum PathsAndCells : bool { kOnlyPaths = false, kPathsAndCells = true };
+
+    explicit ColumnShredder(const BSONObj& obj, PathsAndCells pathsAndCells = kPathsAndCells)
+        : _pathsAndCells(pathsAndCells) {
+        walkObj(_paths[ColumnStore::kRowIdPath], obj, /*isRoot=*/true);
+    }
+
+    void visitPaths(function_ref<void(PathView)> cb) const {
+        for (auto&& [path, _] : _paths) {
+            cb(path);
+        }
+    }
+
+    void visitCells(function_ref<void(PathView, const UnencodedCellView&)> cb) {
+        // This function requires that the shredder was constructed to record cell values.
+        invariant(_pathsAndCells);
+
+        for (auto&& [path, cell] : _paths) {
+            cb(path, makeCellView(path, cell));
+        }
+    }
+
+    static void visitDiff(const BSONObj& oldObj,
+                          const BSONObj& newObj,
+                          function_ref<void(DiffAction, PathView, const UnencodedCellView*)> cb) {
+        auto oldShredder = ColumnShredder(oldObj);
+        auto newShredder = ColumnShredder(newObj);
+
+        for (auto&& [path, rawCellNew] : newShredder._paths) {
+            auto itOld = oldShredder._paths.find(path);
+            if (itOld == oldShredder._paths.end()) {
+                auto cell = newShredder.makeCellView(path, rawCellNew);
+                cb(kInsert, path, &cell);
+                continue;
+            }
+
+            auto&& [_, rawCellOld] = *itOld;  // For symmetry with rawCellNew.
+
+            // Need to compute sparseness prior to checking cell equality.
+            newShredder.computeIsSparse(path, &rawCellNew);
+            oldShredder.computeIsSparse(path, &rawCellOld);
+            if (rawCellNew == rawCellOld)
+                continue;  // No change for this path.
+
+
+            auto cell = newShredder.makeCellView(path, rawCellNew);
+            cb(kUpdate, path, &cell);
+        }
+
+        for (auto&& [path, _] : oldShredder._paths) {
+            if (!newShredder._paths.contains(path)) {
+                cb(kDelete, path, nullptr);
+            }
+        }
+    }
+
+
+private:
+    enum Sparseness : int8_t { kUnknown, kIsSparse, kNotSparse };
+
+    struct RawCellValue {
+        // This is where we incrementally build up the full arrayInfo for this cell. While each
+        // position is essentially a walk from the root to somewhere in the document, this is a walk
+        // that starts at the root and then passes through each value or subobject at this path
+        // taking (hopefully) the shortest distance between each value. First we build it up in
+        // "uncompressed" format, then we call compressArrayInfo() to remove redundant information.
+        std::string arrayInfoBuf;
+
+        // This is used when building up arrayInfoBuf. It records the absolute position of the last
+        // value or subobject appended to arrayInfoBuf, so that we can easily compute the relative
+        // path to the next value. Equivalently, it is the value of _currentArrayInfo at the time of
+        // the last call to appendToArrayInfo(). It never has any '|' or 'o' bytes.
+        std::string lastPosition;
+
+        std::vector<BSONElement> vals;
+        int nSeen = 0;                     // Number of times this field has been seen.
+        int nNonEmptySubobjects = 0;       // Number of values that are non-empty objects
+        Sparseness sparseness = kUnknown;  // Memoized in computeIsSparse()
+        bool childrenMustBeSparse = false;
+        bool hasDoubleNestedArrays = false;
+        bool hasDuplicateFields = false;
+
+        bool operator==(const RawCellValue& rhs) const {
+            const RawCellValue& lhs = *this;
+
+            // Sparseness must have already been resolved.
+            invariant(lhs.sparseness != kUnknown);
+            invariant(rhs.sparseness != kUnknown);
+
+            // Intentionally not checking nSeen or childrenMustBeSparse, since they are just inputs
+            // to sparseness computation.
+
+            // Ordered to test cheaper stuff before more expensive stuff.
+            return lhs.sparseness == rhs.sparseness &&                     //
+                lhs.hasDoubleNestedArrays == rhs.hasDoubleNestedArrays &&  //
+                lhs.nNonEmptySubobjects == rhs.nNonEmptySubobjects &&      //
+                lhs.hasDuplicateFields == rhs.hasDuplicateFields &&        //
+                lhs.arrayInfoBuf == rhs.arrayInfoBuf &&                    //
+                identicalBSONElementArrays(lhs.vals, rhs.vals);
+        }
+    };
+
+    UnencodedCellView makeCellView(PathView path, RawCellValue& cell) {
+        compressArrayInfo(cell.arrayInfoBuf);
+        return {
+            cell.vals,
+            cell.arrayInfoBuf,
+            cell.hasDuplicateFields,
+            /*hasSubPaths*/ cell.nNonEmptySubobjects != 0,
+            computeIsSparse(path, &cell),
+            cell.hasDoubleNestedArrays,
+        };
+    }
+
+    // Memoized, so amortized O(1) when called on all paths, even though it must check all parent
+    // paths.
+    bool computeIsSparse(PathView path, RawCellValue* cell) {
+        if (cell->sparseness != kUnknown)
+            return cell->sparseness == kIsSparse;
+
+        auto parentPath = parentPathOf(path);
+        if (!parentPath) {
+            // Top level fields are never considered sparse.
+            cell->sparseness = kNotSparse;
+            return false;
+        }
+
+        auto parentIt = _paths.find(*parentPath);
+        invariant(parentIt != _paths.end());
+        auto& parent = parentIt->second;
+        const bool isSparse = parent.childrenMustBeSparse ||
+            cell->nSeen != parent.nNonEmptySubobjects || computeIsSparse(*parentPath, &parent);
+        cell->sparseness = isSparse ? kIsSparse : kNotSparse;
+        return isSparse;
+    }
+
+    void walkObj(RawCellValue& cell, const BSONObj& obj, bool isRoot = false) {
+        if (_pathsAndCells) {
+            cell.nNonEmptySubobjects++;
+            if (!isRoot) {
+                appendToArrayInfo(cell, 'o');
+            }
+        }
+
+        ON_BLOCK_EXIT([&, oldArrInfoSize = _currentArrayInfo.size()] {
+            if (_pathsAndCells)
+                _currentArrayInfo.resize(oldArrInfoSize);
+        });
+        if (_pathsAndCells && !isRoot)
+            _currentArrayInfo += '{';
+
+        for (auto [name, elem] : obj) {
+            // We skip fields in some edge cases such as dots in the field name. This may also throw
+            // if the field name contains invalid UTF-8 in a way that would break the index.
+            if (shouldSkipField(name))
+                continue;
+
+            ON_BLOCK_EXIT([&, oldPathSize = _currentPath.size()] {  //
+                _currentPath.resize(oldPathSize);
+            });
+            if (!isRoot)
+                _currentPath += '.';
+            _currentPath += name;
+
+            auto& subCell = _paths[_currentPath];
+            subCell.nSeen++;
+            if (_inDoubleNestedArray)
+                subCell.hasDoubleNestedArrays = true;
+            handleElem(subCell, elem);
+        }
+    }
+
+    void walkArray(RawCellValue& cell, const BSONObj& arr) {
+        DecimalCounter<unsigned> index;
+
+        for (auto elem : arr) {
+            ON_BLOCK_EXIT([&,
+                           oldArrInfoSize = _currentArrayInfo.size(),
+                           oldInDoubleNested = _inDoubleNestedArray] {
+                if (_pathsAndCells) {
+                    _currentArrayInfo.resize(oldArrInfoSize);
+                    _inDoubleNestedArray = oldInDoubleNested;
+                }
+            });
+
+            if (_pathsAndCells) {
+                _currentArrayInfo += '[';
+                if (index == 0) {
+                    // Zero is common so make it implied rather than explicit.
+                } else {
+                    // Theoretically, index should be the same as elem.fieldNameStringData(),
+                    // however, since we don't validate the array indexes, they cannot be trusted.
+                    // The logic to encode array info relies on "sane" array indexes (at the very
+                    // least they must be monotonically increasing), so we create a new index string
+                    // here.
+                    _currentArrayInfo += StringData(index);
+                }
+                ++index;
+
+                // [] doesn't start a double nesting since {a:{$eq: []}} matches {a: [[]]}
+                if (elem.type() == Array && !elem.Obj().isEmpty())
+                    _inDoubleNestedArray = true;
+            }
+
+            // Note: always same cell, since array traversal never changes path.
+            handleElem(cell, elem);
+        }
+    }
+
+    void handleElem(RawCellValue& cell, const BSONElement& elem) {
+        // Only recurse on non-empty objects and arrays. Empty objects and arrays are handled as
+        // scalars.
+        if (elem.type() == Object) {
+            if (auto obj = elem.Obj(); !obj.isEmpty())
+                return walkObj(cell, obj);
+        } else if (elem.type() == Array) {
+            if (auto obj = elem.Obj(); !obj.isEmpty())
+                return walkArray(cell, obj);
+        }
+
+        if (_pathsAndCells) {
+            // If we get here, then this walk will not have any children. This means that there is
+            // at least one path where all children (if any) will be missing structural information,
+            // so they will need to consult the parent path.
+            cell.childrenMustBeSparse = true;
+
+            if (_inDoubleNestedArray)
+                cell.hasDoubleNestedArrays = true;
+
+            cell.vals.push_back(elem);
+            appendToArrayInfo(cell, '|');
+        }
+    }
+
+    void appendToArrayInfo(RawCellValue& rcd, char finalByte) {
+        dassert(finalByte == '|' || finalByte == 'o');
+
+        auto foundDuplicateField = [&] {
+            rcd.arrayInfoBuf.clear();
+            rcd.hasDuplicateFields = true;
+        };
+
+        if (rcd.hasDuplicateFields) {
+            // arrayInfo should be left empty in this case.
+            invariant(rcd.arrayInfoBuf.empty());
+            return;
+        }
+
+        ON_BLOCK_EXIT([&] { rcd.lastPosition = _currentArrayInfo; });
+
+        if (rcd.arrayInfoBuf.empty()) {
+            // The first time we get a position for this path, just record it verbatim. The first
+            // is special because we are essentially recording the absolute path to this value,
+            // while every other time we append the path relative to the prior value.
+            invariant(rcd.lastPosition.empty());
+            rcd.arrayInfoBuf.reserve(_currentArrayInfo.size() + 1);
+            rcd.arrayInfoBuf += _currentArrayInfo;
+            rcd.arrayInfoBuf += finalByte;
+            return;
+        }
+
+        // Make better names for symmetry (and to prevent accidental modifications):
+        StringData oldPosition = rcd.lastPosition;
+        StringData newPosition = _currentArrayInfo;
+
+        if (MONGO_unlikely(oldPosition.empty() || newPosition.empty())) {
+            // This can only happen if there is a duplicate field at the top level.
+            return foundDuplicateField();
+        }
+
+        auto [oldIt, newIt] = std::mismatch(oldPosition.begin(),  //
+                                            oldPosition.end(),
+                                            newPosition.begin(),
+                                            newPosition.end());
+        if (MONGO_unlikely(newIt == newPosition.end())) {
+            // This can only happen if there is a duplicate field in an array, because otherwise
+            // the raw array infos must differ by an index in some array.
+            return foundDuplicateField();
+        }
+
+        // Walk back to start of differing elem. Important to use newIt here because if they are
+        // in the same array, oldIt may have an implicitly encoded 0 index, while newIt must
+        // have a higher index.
+        while (*newIt != '[') {
+            if (MONGO_unlikely(!(*newIt >= '0' && *newIt <= '9'))) {
+                // Non-index difference can only happen if there are duplicate fields in an array.
+                return foundDuplicateField();
+            }
+            invariant(newIt > newPosition.begin());
+            dassert(oldIt > oldPosition.begin());  // oldIt and newIt are at same index.
+            --newIt;
+            --oldIt;
+        }
+        invariant(oldIt < oldPosition.end());
+        if (MONGO_unlikely(*oldIt != '[')) {
+            // This is another type of non-index difference.
+            return foundDuplicateField();
+        }
+
+        // Close out arrays past the first mismatch in LIFO order.
+        for (auto revOldIt = oldPosition.end() - 1; revOldIt != oldIt; --revOldIt) {
+            if (*revOldIt == '[') {
+                rcd.arrayInfoBuf += ']';
+            } else {
+                dassert((*revOldIt >= '0' && *revOldIt <= '9') || *revOldIt == '{');
+            }
+        }
+
+        // Now process the mismatch. It must be a difference in array index (checked above).
+        dassert(*oldIt == '[' && *newIt == '[');
+        ++oldIt;
+        ++newIt;
+        const auto oldIx = ColumnStore::readArrInfoNumber(&oldIt, oldPosition.end());
+        const auto newIx = ColumnStore::readArrInfoNumber(&newIt, newPosition.end());
+
+        if (MONGO_unlikely(newIx <= oldIx)) {
+            // If this element is at a same or lower index, we must have hit a duplicate field
+            // above and restarted the array indexing.
+            return foundDuplicateField();
+        }
+        const auto delta = newIx - oldIx;
+        const auto skips = delta - 1;
+        if (skips == 0) {
+            // Nothing. Increment by one (skipping zero) is implicit.
+        } else {
+            rcd.arrayInfoBuf += '+';
+            rcd.arrayInfoBuf += ItoA(skips);
+        }
+
+        // Now put the rest of the new array info into the output buffer.
+        rcd.arrayInfoBuf += StringData(newIt, newPosition.end());
+        rcd.arrayInfoBuf += finalByte;
+    }
+
+    static void compressArrayInfo(std::string& arrayInfo) {
+        // NOTE: all operations in this function either shrink the array info or keep it the same
+        // size, so they are able to work in-place in the arrayInfo buffer.
+
+        // Logic below assumes arrayInfo is null terminated as a simplification.
+        dassert(strlen(arrayInfo.data()) == arrayInfo.size());
+
+        const char* in = arrayInfo.data();
+        char* out = arrayInfo.data();
+        bool anyArrays = false;
+
+        // Remove all '{' immediately before a '|' or 'o', and check for arrays
+        char* lastNonBrace = nullptr;
+        while (auto c = *in++) {
+            if (c == '[') {
+                anyArrays = true;
+            } else if ((c == '|' || c == 'o') && lastNonBrace) {
+                // Rewind output to just past last non-brace, since the last set of opens are
+                // encoded implicitly.
+                dassert(lastNonBrace + 1 <= out);
+                out = lastNonBrace + 1;
+            }
+
+            if (c != '{')
+                lastNonBrace = out;
+
+            *out++ = c;
+        }
+
+        // If there were no arrays, we don't need any array info.
+        if (!anyArrays) {
+            arrayInfo.clear();
+            return;
+        }
+
+        invariant(size_t(out - arrayInfo.data()) <= arrayInfo.size());
+
+        // Remove final run of '|' since end implies an infinite sequence of them.
+        while (out - arrayInfo.data() >= 1 && out[-1] == '|') {
+            out--;
+        }
+        arrayInfo.resize(out - arrayInfo.data());  // Reestablishes null termination.
+
+        // Now do a final pass to RLE remaining runs of '|' or 'o'. It may be possible to integrate
+        // this with the first loop above, but it would be a bit more complicated because we need
+        // to have removed any implicit runs of '{' prior to each '|' or 'o' before looking for
+        // runs.
+        out = arrayInfo.data();
+        in = arrayInfo.data();
+        while (auto c = *in++) {
+            *out++ = c;
+            if (c != '|' && c != 'o')
+                continue;
+
+            size_t repeats = 0;
+            while (*in == c) {
+                repeats++;
+                in++;
+            }
+
+            if (repeats) {
+                auto buf = ItoA(repeats);  // Must be on own line so the buffer outlives numStr.
+                auto numStr = StringData(buf);
+
+                // We know there is room because a number > 0 takes up no more space than the number
+                // of '|' that it is replacing. repeats == 1 is the worst case because it isn't
+                // saving any space. However we still replace here since it should make decoding
+                // slightly faster.
+                dassert(numStr.size() <= repeats);
+                for (char c : numStr) {
+                    dassert(out < in);  // Note: `in` has already been advanced.
+                    *out++ = c;
+                }
+            }
+        }
+
+        invariant(size_t(out - arrayInfo.data()) <= arrayInfo.size());
+        arrayInfo.resize(out - arrayInfo.data());
+    }
+
+    static bool shouldSkipField(PathView fieldName) {
+        // TODO consider skipping when the nesting depths exceed our documented limits of 100
+        // documents. This would allow decoding to use fixed-size bitsets without needing to check
+        // for overflow.
+
+        // WARNING: do not include the field name in error messages because if we throw from this
+        // function then the field name isn't valid utf-8, and could poison the whole error message!
+
+        // We only care about \xFF at the beginning, even though utf-8 bans it everywhere.
+        static_assert(ColumnStore::kRowIdPath == "\xFF"_sd);
+        uassert(6519200,
+                "Field name contains '\\xFF' which isn't valid in UTF-8",
+                fieldName[0] != '\xFF');  // We know field names are nul-terminated so this is safe.
+
+        // Fields with dots are not illegal, but we skip them because a) the query optimizer can't
+        // correctly track dependencies with them, and b) the current storage format uses dots as a
+        // separator and everything (including the array info encoder) would get confused with
+        // documents like {a: {b:1}, a.b: 2} because that breaks some of the assumptions.
+        return fieldName.find('.') != std::string::npos;
+    }
+
+    static boost::optional<PathView> parentPathOf(PathView path) {
+        auto sep = path.rfind('.');
+        if (sep == std::string::npos)
+            return {};
+
+        return path.substr(0, sep);
+    }
+
+    // This is the same as StringMap<V> but with node_hash_map. We need the reference stability
+    // guarantee because we have recursive functions that both insert into _paths and hold a
+    // reference to one of its values. It seems better to use a node-based hash table than to redo
+    // the lookup every time we want to touch a cell.
+    template <typename V>
+    using NodeStringMap = absl::node_hash_map<std::string, V, StringMapHasher, StringMapEq>;
+
+    PathValue _currentPath;
+    std::string _currentArrayInfo;  // Describes a walk from root to current pos. No '|' or 'o'.
+    NodeStringMap<RawCellValue> _paths;
+    bool _inDoubleNestedArray = false;
+
+    const PathsAndCells _pathsAndCells = kPathsAndCells;
+};
+}  // namespace
+
+void visitCellsForInsert(const BSONObj& obj,
+                         function_ref<void(PathView, const UnencodedCellView&)> cb) {
+    ColumnShredder(obj).visitCells(cb);
+}
+
+void visitPathsForDelete(const BSONObj& obj, function_ref<void(PathView)> cb) {
+    ColumnShredder(obj, ColumnShredder::kOnlyPaths).visitPaths(cb);
+}
+
+void visitDiffForUpdate(const BSONObj& oldObj,
+                        const BSONObj& newObj,
+                        function_ref<void(DiffAction, PathView, const UnencodedCellView*)> cb) {
+    ColumnShredder::visitDiff(oldObj, newObj, cb);
+}
+
+bool operator==(const UnencodedCellView& lhs, const UnencodedCellView& rhs) {
+    if (lhs.hasDuplicateFields || rhs.hasDuplicateFields) {
+        // As a special case, if either is true, we only care about comparing this field, since all
+        // other fields are suspect. This simplifies testing, because we don't need to guess what
+        // the output will be (and lock it into a correctness test!) for the case with duplicate
+        // fields.
+        return lhs.hasDuplicateFields == rhs.hasDuplicateFields;
+    }
+
+    return identicalBSONElementArrays(lhs.vals, rhs.vals) &&  //
+        lhs.arrayInfo == rhs.arrayInfo &&                     //
+        lhs.hasSubPaths == rhs.hasSubPaths &&                 //
+        lhs.isSparse == rhs.isSparse &&                       //
+        lhs.hasDoubleNestedArrays == rhs.hasDoubleNestedArrays;
+}
+
+std::ostream& operator<<(std::ostream& os, const UnencodedCellView& cell) {
+    if (cell.hasDuplicateFields) {
+        // As a special case just output this info, since other fields don't matter.
+        return os << "{duplicateFields: 1}";
+    }
+
+    os << "{vals: [";
+    for (auto&& elem : cell.vals) {
+        if (&elem != &cell.vals.front())
+            os << ", ";
+        os << elem.toString(/*includeFieldName*/ false);
+    }
+    os << "], arrayInfo: '" << cell.arrayInfo;
+    os << "', hasSubPaths: " << cell.hasSubPaths;
+    os << ", isSparse: " << cell.isSparse;
+    os << ", hasDoubleNestedArrays: " << cell.hasDoubleNestedArrays;
+    os << '}';
+
+    return os;
+}
+std::ostream& operator<<(std::ostream& os, const UnencodedCellView* cell) {
+    return cell ? (os << *cell) : (os << "(no cell)");
+}
+}  // namespace mongo::column_keygen