hugegraph 存取数据解析

hugegraph 是百度开源的图数据库，支持hbase，mysql，rocksdb等作为存储后端。本文以EDGE 存储，hbase为存储后端，来探索hugegraph是如何存取数据的。

存数据

序列化

首先需要序列化，hbase 使用BinarySerializer:

• keyWithIdPrefix 和indexWithIdPrefix都是false

这个后面会用到。

public class HbaseSerializer extends BinarySerializer {

    public HbaseSerializer() {
        super(false, true);
    }
}

要存到db，首先需要序列化为BackendEntry，BackendEntry 是图数据库和后端存储的传输对象，Hbase对应的是BinaryBackendEntry:

public class BinaryBackendEntry implements BackendEntry {

    private static final byte[] EMPTY_BYTES = new byte[]{};

    private final HugeType type;
    private final BinaryId id;
    private Id subId;
    private final List<BackendColumn> columns;
    private long ttl;

    public BinaryBackendEntry(HugeType type, byte[] bytes) {
        this(type, BytesBuffer.wrap(bytes).parseId(type));
    }

    public BinaryBackendEntry(HugeType type, BinaryId id) {
        this.type = type;
        this.id = id;
        this.subId = null;
        this.columns = new ArrayList<>();
        this.ttl = 0L;
    }

我们来看序列化，序列化，其实就是要将数据放到entry的column列里。

• hbase 的keyWithIdPrefix是false，因此name不包含ownerVertexId（参考下面的EdgeId，去掉ownerVertexId）

 public BackendEntry writeEdge(HugeEdge edge) {
        BinaryBackendEntry entry = newBackendEntry(edge);
        byte[] name = this.keyWithIdPrefix ?
                      this.formatEdgeName(edge) : EMPTY_BYTES;
        byte[] value = this.formatEdgeValue(edge);
        entry.column(name, value);

        if (edge.hasTtl()) {
            entry.ttl(edge.ttl());
        }

        return entry;
    }

EdgeId：

    private final Id ownerVertexId;
    private final Directions direction;
    private final Id edgeLabelId;
    private final String sortValues;
    private final Id otherVertexId;

    private final boolean directed;
    private String cache;

backend 存储

生成BackendEntry后，通过store机制，交给后端的backend存储。

EDGE的保存，对应HbaseTables.Edge:

public static class Edge extends HbaseTable {

        @Override
        public void insert(Session session, BackendEntry entry) {
            long ttl = entry.ttl();
            if (ttl == 0L) {
                session.put(this.table(), CF, entry.id().asBytes(),
                            entry.columns());
            } else {
                session.put(this.table(), CF, entry.id().asBytes(),
                            entry.columns(), ttl);
            }
        }
}

CF 是固定的f：

    protected static final byte[] CF = "f".getBytes();

session.put 对应：

 @Override
        public void put(String table, byte[] family, byte[] rowkey,
                        Collection<BackendColumn> columns) {
            Put put = new Put(rowkey);
            for (BackendColumn column : columns) {
                put.addColumn(family, column.name, column.value);
            }
            this.batch(table, put);
        }

可以看出，存储时，edgeid作为rowkey，然后把去除ownerVertexId后的edgeid作为column.name

EDGE 读取

从backend读取BackendEntry

读取就是从hbase读取result，转换为BinaryBackendEntry，再转成Edge。

读取，是scan的过程：

 /**
         * Inner scan: send scan request to HBase and get iterator
         */
        @Override
        public RowIterator scan(String table, Scan scan) {
            assert !this.hasChanges();

            try (Table htable = table(table)) {
                return new RowIterator(htable.getScanner(scan));
            } catch (IOException e) {
                throw new BackendException(e);
            }
        }

scan后，返回BackendEntryIterator

protected BackendEntryIterator newEntryIterator(Query query,
                                                    RowIterator rows) {
        return new BinaryEntryIterator<>(rows, query, (entry, row) -> {
            E.checkState(!row.isEmpty(), "Can't parse empty HBase result");
            byte[] id = row.getRow();
            if (entry == null || !Bytes.prefixWith(id, entry.id().asBytes())) {
                HugeType type = query.resultType();
                // NOTE: only support BinaryBackendEntry currently
                entry = new BinaryBackendEntry(type, id);
            }
            try {
                this.parseRowColumns(row, entry, query);
            } catch (IOException e) {
                throw new BackendException("Failed to read HBase columns", e);
            }
            return entry;
        });
    }

注意，new BinaryBackendEntry(type, id) 时，BinaryBackendEntry的id并不是rowkey，而是对rowkey做了处理：

public BinaryId parseId(HugeType type) {
        if (type.isIndex()) {
            return this.readIndexId(type);
        }
        // Parse id from bytes
        int start = this.buffer.position();
        /*
         * Since edge id in edges table doesn't prefix with leading 0x7e,
         * so readId() will return the source vertex id instead of edge id,
         * can't call: type.isEdge() ? this.readEdgeId() : this.readId();
         */
        Id id = this.readId();
        int end = this.buffer.position();
        int len = end - start;
        byte[] bytes = new byte[len];
        System.arraycopy(this.array(), start, bytes, 0, len);
        return new BinaryId(bytes, id);
    }

这里是先读取ownervertexId作为Id部分, 然后将剩余的直接放入bytes，组合成BinaryId，和序列化的时候有差别，为什么这么设计呢？原来不管是vertex还是edge，都是当成Vertex来读取的。

protected final BinaryBackendEntry newBackendEntry(HugeEdge edge) {
        BinaryId id = new BinaryId(formatEdgeName(edge),
                                   edge.idWithDirection());
        return newBackendEntry(edge.type(), id);
    }

public EdgeId directed(boolean directed) {
    return new EdgeId(this.ownerVertexId, this.direction, this.edgeLabelId,
                      this.sortValues, this.otherVertexId, directed);
}

序列化的时候是EdgeId。

BackendEntryIterator迭代器支持对结果进行merge, 上面代码里的!Bytes.prefixWith(id, entry.id().asBytes())) 就是对比是否是同一个ownervertex，如果是同一个，则放到同一个BackendEntry的Columns里。

     public BinaryEntryIterator(BackendIterator<Elem> results, Query query,
                               BiFunction<BackendEntry, Elem, BackendEntry> m)

    @Override
    protected final boolean fetch() {
        assert this.current == null;
        if (this.next != null) {
            this.current = this.next;
            this.next = null;
        }

        while (this.results.hasNext()) {
            Elem elem = this.results.next();
            BackendEntry merged = this.merger.apply(this.current, elem);
            E.checkState(merged != null, "Error when merging entry");
            if (this.current == null) {
                // The first time to read
                this.current = merged;
            } else if (merged == this.current) {
                // The next entry belongs to the current entry
                assert this.current != null;
                if (this.sizeOf(this.current) >= INLINE_BATCH_SIZE) {
                    break;
                }
            } else {
                // New entry
                assert this.next == null;
                this.next = merged;
                break;
            }

            // When limit exceed, stop fetching
            if (this.reachLimit(this.fetched() - 1)) {
                // Need remove last one because fetched limit + 1 records
                this.removeLastRecord();
                this.results.close();
                break;
            }
        }

        return this.current != null;
    }

从BackendEntry转换为edge

然后再来看读取数据readVertex，前面说了，就算是edge，其实也是当vertex来读取的：

 @Override
    public HugeVertex readVertex(HugeGraph graph, BackendEntry bytesEntry) {
        if (bytesEntry == null) {
            return null;
        }
        BinaryBackendEntry entry = this.convertEntry(bytesEntry);

        // Parse id
        Id id = entry.id().origin();
        Id vid = id.edge() ? ((EdgeId) id).ownerVertexId() : id;
        HugeVertex vertex = new HugeVertex(graph, vid, VertexLabel.NONE);

        // Parse all properties and edges of a Vertex
        for (BackendColumn col : entry.columns()) {
            if (entry.type().isEdge()) {
                // NOTE: the entry id type is vertex even if entry type is edge
                // Parse vertex edges
                this.parseColumn(col, vertex);
            } else {
                assert entry.type().isVertex();
                // Parse vertex properties
                assert entry.columnsSize() == 1 : entry.columnsSize();
                this.parseVertex(col.value, vertex);
            }
        }

        return vertex;
    }

逻辑：

• 先读取ownervertexid，生成HugeVertex，这个时候只知道id，不知道vertexlabel，所以设置为VertexLabel.NONE
• 然后，读取BackendColumn，一个edge，一个Column（name是edgeid去除ownervertexid后的部分，value是边数据）

读取是在parseColumn:

protected void parseColumn(BackendColumn col, HugeVertex vertex) {
        BytesBuffer buffer = BytesBuffer.wrap(col.name);
        Id id = this.keyWithIdPrefix ? buffer.readId() : vertex.id();
        E.checkState(buffer.remaining() > 0, "Missing column type");
        byte type = buffer.read();
        // Parse property
        if (type == HugeType.PROPERTY.code()) {
            Id pkeyId = buffer.readId();
            this.parseProperty(pkeyId, BytesBuffer.wrap(col.value), vertex);
        }
        // Parse edge
        else if (type == HugeType.EDGE_IN.code() ||
                 type == HugeType.EDGE_OUT.code()) {
            this.parseEdge(col, vertex, vertex.graph());
        }
        // Parse system property
        else if (type == HugeType.SYS_PROPERTY.code()) {
            // pass
        }
        // Invalid entry
        else {
            E.checkState(false, "Invalid entry(%s) with unknown type(%s): 0x%s",
                         id, type & 0xff, Bytes.toHex(col.name));
        }
    }

从``col.name`读取type，如果是edge，则parseEdge:

protected void parseEdge(BackendColumn col, HugeVertex vertex,
                             HugeGraph graph) {
        // owner-vertex + dir + edge-label + sort-values + other-vertex

        BytesBuffer buffer = BytesBuffer.wrap(col.name);
        if (this.keyWithIdPrefix) {
            // Consume owner-vertex id
            buffer.readId();
        }
        byte type = buffer.read();
        Id labelId = buffer.readId();
        String sortValues = buffer.readStringWithEnding();
        Id otherVertexId = buffer.readId();

        boolean direction = EdgeId.isOutDirectionFromCode(type);
        EdgeLabel edgeLabel = graph.edgeLabelOrNone(labelId);

        // Construct edge
        HugeEdge edge = HugeEdge.constructEdge(vertex, direction, edgeLabel,
                                               sortValues, otherVertexId);

        // Parse edge-id + edge-properties
        buffer = BytesBuffer.wrap(col.value);

        //Id id = buffer.readId();

        // Parse edge properties
        this.parseProperties(buffer, edge);

        // Parse edge expired time if needed
        if (edge.hasTtl()) {
            this.parseExpiredTime(buffer, edge);
        }
    }

从col.name依次读取出type,labelId,sortValues和otherVertexId：

        byte type = buffer.read();
        Id labelId = buffer.readId();
        String sortValues = buffer.readStringWithEnding();
        Id otherVertexId = buffer.readId();

然后根据labelid找到 EdgeLabel edgeLabel = graph.edgeLabelOrNone(labelId);

创建edge, 解析边属性parseProperties

最后读取Ttl, 处理结果的时候，会过滤过期数据。