×

6. sharding-jdbc源码之group by结果合并(1)

96
阿飞的博客
2018.02.01 15:47* 字数 940

阿飞Javaer,转载请注明原创出处,谢谢!

5. sharding-jdbc源码之结果合并中已经分析了OrderByStreamResultSetMerger、LimitDecoratorResultSetMerger、IteratorStreamResultSetMerger,查看源码目录下ResultSetMerger的实现类,只剩下GroupByMemoryResultSetMerger和GroupByStreamResultSetMerger两个实现类的分析,接下来根据源码对两者的实现进行剖析;

ResultSetMerge关系图.png

如何选择

GroupBy有两个ResultSetMerge的实现:GroupByMemoryResultSetMerger和GroupByStreamResultSetMerger,那么如何选择呢?在MergeEngine中有一段这样的代码:

private ResultSetMerger build() throws SQLException {
    // 如果有group by或者聚合类型(例如sum, avg等)的SQL条件,就会选择一个GroupBy***ResultSetMerger
    if (!selectStatement.getGroupByItems().isEmpty() || !selectStatement.getAggregationSelectItems().isEmpty()) {
        // isSameGroupByAndOrderByItems()源码紧随其后
        if (selectStatement.isSameGroupByAndOrderByItems()) {
            return new GroupByStreamResultSetMerger(columnLabelIndexMap, resultSets, selectStatement);
        } else {
            return new GroupByMemoryResultSetMerger(columnLabelIndexMap, resultSets, selectStatement);
        }
    }
    if (!selectStatement.getOrderByItems().isEmpty()) {
        return new OrderByStreamResultSetMerger(resultSets, selectStatement.getOrderByItems());
    }
    return new IteratorStreamResultSetMerger(resultSets);
}

// 如果只有group by条件,没有order by,那么isSameGroupByAndOrderByItems()为true,例如:`SELECT o.* FROM t_order o where o.user_id=? group by o.order_id`(因为这种sql会被改写为SELECT o.* , o.order_id AS GROUP_BY_DERIVED_0 FROM t_order_0 o where o.user_id=?  group by o.order_id  ORDER BY GROUP_BY_DERIVED_0 ASC,即group by和order by完全相同)
public boolean isSameGroupByAndOrderByItems() {
    return !getGroupByItems().isEmpty() && getGroupByItems().equals(getOrderByItems());
}

由上段源码分析可知,如果只有group by条件,那么选择GroupByStreamResultSetMerger;那么如果既有group by,又有order by,那么就会选择GroupByStreamResultSetMerger;

接下来分析GroupByStreamResultSetMerger中如何对结果进行group by聚合,假设数据源js_jdbc_0中实际表t_order_0和实际表t_order_1的数据如下:

order_id user_id status
1000 10 INIT
1002 10 INIT
1004 10 VALID
1006 10 NEW
1008 10 INIT
order_id user_id status
1001 10 NEW
1003 10 NEW
1005 10 VALID
1007 10 INIT
1009 10 INIT

GroupByStreamResultSetMerger

以执行SQLSELECT o.status, count(o.user_id) FROM t_order o where o.user_id=10 group by o.status为例,分析GroupByStreamResultSetMerger,其部分源码如下:

public final class GroupByStreamResultSetMerger extends OrderByStreamResultSetMerger {  
    ... ... 
    public GroupByStreamResultSetMerger(
            final Map<String, Integer> labelAndIndexMap, final List<ResultSet> resultSets, final SelectStatement selectStatement) throws SQLException {
        // GroupByStreamResultSetMerger的父类是OrderByStreamResultSetMerger,所以调用super()就是调用OrderByStreamResultSetMerger的构造方法
        super(resultSets, selectStatement.getOrderByItems());
        // 标签(列名)和位置索引的map关系,例如{order_id:1, status:3, user_id:2}        
        this.labelAndIndexMap = labelAndIndexMap;
        // 执行的SQL语句
        this.selectStatement = selectStatement;
        currentRow = new ArrayList<>(labelAndIndexMap.size());
        // 如果优先级队列不为空,表示where条件中有group by,将队列中第一个元素的group值赋值给currentGroupByValues,即INIT(默认升序排列,所以INIT > NEW > VALID)
        currentGroupByValues = getOrderByValuesQueue().isEmpty() ? Collections.emptyList() : new GroupByValue(getCurrentResultSet(), selectStatement.getGroupByItems()).getGroupValues();
    }
    ...
}

备注:OrderByStreamResultSetMerger在5. sharding-jdbc源码之结果合并这篇文章中已经分析,不再赘述;

next()方法核心源码如下:

@Override
public boolean next() throws SQLException {
    currentRow.clear();
    // 如果优先级队列为空,表示没有任何结果,那么返回false
    if (getOrderByValuesQueue().isEmpty()) {
        return false;
    }
    if (isFirstNext()) {
        super.next();
    }
    // 集合的核心逻辑在这里
    if (aggregateCurrentGroupByRowAndNext()) {
        currentGroupByValues = new GroupByValue(getCurrentResultSet(), selectStatement.getGroupByItems()).getGroupValues();
    }
    return true;
}

aggregateCurrentGroupByRowAndNext()实现如下:

private boolean aggregateCurrentGroupByRowAndNext() throws SQLException {
    boolean result = false;
    // selectStatement.getAggregationSelectItems()先得到select所有举行类型的项,例如select count(o.user_id) ***中聚合项是count(o.user_id), 然后转化成map,key就是聚合项即o.user_id,value就是集合unit实例即AccumulationAggregationUnit;即o.user_id的COUNT集合计算是通过AccumulationAggregationUnit实现的,下面有对AggregationUnitFactory的分析
    Map<AggregationSelectItem, AggregationUnit> aggregationUnitMap = Maps.toMap(selectStatement.getAggregationSelectItems(), new Function<AggregationSelectItem, AggregationUnit>() {
        
        @Override
        public AggregationUnit apply(final AggregationSelectItem input) {
            return AggregationUnitFactory.create(input.getType());
        }
    });
    // 接下来准备聚合,如何group by的值相同,则进行聚合(因为SQL可能会在多个数据源以及多个实际表上执行)
    while (currentGroupByValues.equals(new GroupByValue(getCurrentResultSet(), selectStatement.getGroupByItems()).getGroupValues())) {
        // 调用aggregate()方法进行䄦
        aggregate(aggregationUnitMap);
        cacheCurrentRow();
        // 调用next()方法,实际调用OrderByStreamResultSetMerger中的next()方法,currentResultSet会指向下一个元素;
        result = super.next();
        // 如果还有值,那么继续遍历
        if (!result) {
            break;
        }
    }
    setAggregationValueToCurrentRow(aggregationUnitMap);
    return result;
}

AggregationUnitFactory 源码如下:

public final class AggregationUnitFactory {
    
    /**
     * Create aggregation unit instance.
     * 根据这段代码可知,select中MAX和MIN这种聚合查询需要使用ComparableAggregationUnit,SUM和COUNT需要使用AccumulationAggregationUnit,AVG需要使用AverageAggregationUnit;(目前只支持这些聚合操作),
     */
    public static AggregationUnit create(final AggregationType type) {
        switch (type) {
            case MAX:
                return new ComparableAggregationUnit(false);
            case MIN:
                return new ComparableAggregationUnit(true);
            case SUM:
            case COUNT:
                return new AccumulationAggregationUnit();
            case AVG:
                return new AverageAggregationUnit();
            default:
                throw new UnsupportedOperationException(type.name());
        }
    }
}

aggregate()源码如下:

private void aggregate(final Map<AggregationSelectItem, AggregationUnit> aggregationUnitMap) throws SQLException {
    for (Entry<AggregationSelectItem, AggregationUnit> entry : aggregationUnitMap.entrySet()) {
        List<Comparable<?>> values = new ArrayList<>(2);
        if (entry.getKey().getDerivedAggregationSelectItems().isEmpty()) {
            values.add(getAggregationValue(entry.getKey()));
        } else {
            for (AggregationSelectItem each : entry.getKey().getDerivedAggregationSelectItems()) {
                values.add(getAggregationValue(each));
            }
        }
        // aggregate()的核心就是调用AggregationUnit具体实现中的merge()方法,即调用AccumulationAggregationUnit.merge()方法(后面会对AggregationUnit的各个实现进行分析)
        entry.getValue().merge(values);
    }
}

执行过程图解

这一块的代码逻辑稍微有点复杂,下面通过示意图分解执行过程,让sharding-jdbc执行group by整个过程更加清晰:
step1. SQL执行
首先在两个实际表t_order_0t_order_1中分别执行SQL:SELECT o.status, count(o.user_id) FROM t_order o where o.user_id=10 group by o.statust_order_0t_order_1分别得到如下的结果:

status count(o.user_id)
INIT 3
NEW 1
VALID 1
status count(o.user_id)
INIT 2
NEW 2
VALID 1

step2. 执行super(***)
即在GroupByStreamResultSetMerger中调用OrderByStreamResultSetMerger的构造方法super(resultSets, selectStatement.getOrderByItems());,从而得到优先级队列,如下图所示的第一张图,优先级中包含两个元素[(INIT, 3), (INIT 2)]:

powered by afei.png
  1. 先聚合计算(INIT,3)和(INIT,2),由于NEW和INIT不相等,进行下一轮聚合计算;
  2. 再聚合计算(NEW,1)和(NEW,2),由于VALID和NEW不相等,进行下一轮聚合计算;
  3. 再聚合计算(VALID,1)和(VALID,1),两者的next()为false,聚合计算完成;

step3. aggregationUnitMap
通过转换得到aggregationUnitMap,key就是count(user_id),value就是COUNT聚合计算的AggregationUnit实现,即AccumulationAggregationUnit;

由于select语句中只有COUNT(o.user_id涉及到聚合运行,所以这个map的size为1,且key是count(user_id);如果SQL是SELECT o.status, count(o.user_id), max(order_id) FROM t_order o where o.user_id=? group by o.status,那么aggregationUnitMap的size为2,且第一个entry的key是count(user_id),value是AccumulationAggregationUnit;第二个entry的key是max(order_id),value是ComparableAggregationUnit;

step4. 循环遍历并merge
核心代码如下,即将(INIT, 3)和(INIT, 2)通过调用AccumulationAggregationUnit中的merge方法,从而得到(INIT, 5)。同样的原因调用AccumulationAggregationUnit中的merge方法merge(NEW, 1)和(NEW, 2),从而得到(NEW, 3);merge(VALID, 1)和(VALID, 1),从而得到(VALID, 2)。所以,最终的结果就是[(INIT, 5), (NEW, 3), (VALID, 2)]

while (currentGroupByValues.equals(new GroupByValue(getCurrentResultSet(), selectStatement.getGroupByItems()).getGroupValues())) {
    aggregate(aggregationUnitMap);
    cacheCurrentRow();
    result = super.next();
    if (!result) {
        break;
    }
}

AggregationUnit

AggregationUnit即聚合计算接口,总计有三个实现类AccumulationAggregationUnit,ComparableAggregationUnit和AverageAggregationUnit,接下来分别对其简单介绍;

AccumulationAggregationUnit

实现源码如下,SUN和COUNT两个聚合计算都是用这个AggregationUnit实现,核心实现就是累加:

@Override
public void merge(final List<Comparable<?>> values) {
    if (null == values || null == values.get(0)) {
        return;
    }
    if (null == result) {
        result = new BigDecimal("0");
    }
    // 核心实现代码:累加
    result = result.add(new BigDecimal(values.get(0).toString()));
    log.trace("Accumulation result: {}", result.toString());
}

ComparableAggregationUnit

实现源码如下,MAX和MIN两个聚合计算都是用这个AggregationUnit实现,核心实现就是比较:

@Override
public void merge(final List<Comparable<?>> values) {
    if (null == values || null == values.get(0)) {
        return;
    }
    if (null == result) {
        result = values.get(0);
        log.trace("Comparable result: {}", result);
        return;
    }
    // 新的值与旧的值比较大小
    int comparedValue = ((Comparable) values.get(0)).compareTo(result);
    // 升序和降序比较方式不同(max聚合计算时asc为false,min聚合计算时asc为true),min聚合计算时找一个更小的值(asc && comparedValue < 0),max聚合计算时找一个更大的值(!asc && comparedValue > 0)
    if (asc && comparedValue < 0 || !asc && comparedValue > 0) {
        result = values.get(0);
        log.trace("Comparable result: {}", result);
    }
}

AverageAggregationUnit

实现源码如下,AVG聚合计算就是用的这个AggregationUnit实现,核心实现是将AVG转化后的SUM/COUNT,累加得到总SUM和总COUNT相除就是最终的AVG结果;

@Override
public void merge(final List<Comparable<?>> values) {
    if (null == values || null == values.get(0) || null == values.get(1)) {
        return;
    }
    if (null == count) {
        count = new BigDecimal("0");
    }
    if (null == sum) {
        sum = new BigDecimal("0");
    }
    // COUNT累加 
    count = count.add(new BigDecimal(values.get(0).toString()));
    // SUM累加
    sum = sum.add(new BigDecimal(values.get(1).toString()));
    log.trace("AVG result COUNT: {} SUM: {}", count, sum);
}
sharding-jdbc
Web note ad 1