Presto阅读源码Sql执行过程的记录

前言
Presto源码主要从两部分入手阅读,presto-cli与presto-main分别对应的是client端的入口与server端的入口工程。
版本如下
com.facebook.presto
presto-root
0.190-SNAPSHOT

presto-cli的main方法入口为

com.facebook.presto.cli.Presto.java

presto-main的main方法入口为

com.facebook.presto.server.PrestoServer.java

client端与server端的启动分别从这两个main中启动,今天主要说一下client端将查询请求到server端后如何执行查询的。

---

StatementResource.createQuery()方法

client会发送一个http请求到 StatementResource 类的createQuery方法

    if (isNullOrEmpty(statement)) {
            throw new WebApplicationException(Response
                    .status(Status.BAD_REQUEST)
                    .type(MediaType.TEXT_PLAIN)
                    .entity("SQL statement is empty")
                    .build());
        }

        SessionContext sessionContext = new HttpRequestSessionContext(servletRequest);

        ExchangeClient exchangeClient = exchangeClientSupplier.get(deltaMemoryInBytes -> {});    
        //进入到create方法
        Query query = Query.create(
                sessionContext,
                statement,
                queryManager,
                sessionPropertyManager,
                exchangeClient,
                responseExecutor,
                timeoutExecutor,
                blockEncodingSerde);

---

Query.create()方法

准备构造Query对象

  Query result = new Query(sessionContext, query, queryManager, sessionPropertyManager, exchangeClient, dataProcessorExecutor, timeoutExecutor, blockEncodingSerde);

执行Query构造器

QueryInfo queryInfo = queryManager.createQuery(sessionContext, query);

---

SqlQueryManager.createQuery()方法

向查询管理器中提交sql

 // start the query in the background
 queueManager.submit(statement, queryExecution, queryExecutor);

---

SqlQueryQueueManager.submit()方法

public void submit(Statement statement, QueryExecution queryExecution, Executor executor)
{
    List queues;
    try {
     //选择一个符合角色的队列集合
        queues = selectQueues(queryExecution.getSession(), executor);
    }
    catch (PrestoException e) {
        queryExecution.fail(e);
        return;
    }
    for (QueryQueue queue : queues) {
        if (!queue.reserve(queryExecution)) {
            // 如果我们无法获得许可证进入队列,拒绝查询
            // 当这个查询失败时，许可证将被释放
            queryExecution.fail(new PrestoException(QUERY_QUEUE_FULL, "Too many queued queries"));
            return;
        }
    }
    //将构造好的QueuedExecution入队
    queues.get(0).enqueue(createQueuedExecution(queryExecution, queues.subList(1, queues.size()), executor));
}

---

QueryQueue.enqueue.enqueue()方法

public void enqueue(QueuedExecution queuedExecution)
{
    queryQueueSize.incrementAndGet();
     //添加一个回调，如果它已经完成的条目出列。
     //这使得我们可以在开始前取消输入，
     //如果在启动后调用，则不起作用。
    QueueEntry entry = new QueueEntry(queuedExecution, queryQueueSize::decrementAndGet);
    queuedExecution.getCompletionFuture().addListener(entry::dequeue, MoreExecutors.directExecutor());
//异步提交QueueEntry
    asyncSemaphore.submit(entry);
}

---

AsyncSemaphore.submit()方法

public ListenableFuture submit(T task)
{
    QueuedTask queuedTask = new QueuedTask<>(task);
    //向任务队列里填充QueueEntry
    queuedTasks.add(queuedTask);
    //获取许可
    acquirePermit();
    return queuedTask.getCompletionFuture();
}

获取许可开始任务

private final Runnable runNextTask = this::runNext;
private void acquirePermit()
{
 //判断是否有剩余的许可可以发放
    if (counter.incrementAndGet() <= maxPermits) {
        //如果不是所有的许可证都已经发放，那就开始一个任务,注意runNexTask调用的为runNext方法
        submitExecutor.execute(runNextTask);
    }
}

private void runNext()
 {
      final QueuedTask queuedTask = queuedTasks.poll();
      //提交任务
      ListenableFuture future = submitTask(queuedTask.getTask());
      Futures.addCallback(future, new FutureCallback()
      {
          @Override
          public void onSuccess(Object result)
          {
              queuedTask.markCompleted();
              releasePermit();
          }

          @Override
          public void onFailure(Throwable t)
          {
              queuedTask.markFailure(t);
              releasePermit();
          }
      });
  }

private final Function> submitter;
private ListenableFuture submitTask(T task)
{
     try {
      //提交并且回调到QueryQueue
         ListenableFuture future = submitter.apply(task);
         if (future == null) {
             return Futures.immediateFailedFuture(new NullPointerException("Submitter returned a null future for task: " + task));
         }
         return future;
     }
     catch (Exception e) {
         return Futures.immediateFailedFuture(e);
     }
 }

---

QueryQueue构造
该对象调用是来自SqlQueryQueueManager.selectQueues().getOrCreateQueues

QueryQueue(Executor queryExecutor, int maxQueuedQueries, int maxConcurrentQueries)
{
    requireNonNull(queryExecutor, "queryExecutor is null");
    checkArgument(maxQueuedQueries > 0, "maxQueuedQueries must be greater than zero");
    checkArgument(maxConcurrentQueries > 0, "maxConcurrentQueries must be greater than zero");

    int permits = maxQueuedQueries + maxConcurrentQueries;
    // Check for overflow
    checkArgument(permits > 0, "maxQueuedQueries + maxConcurrentQueries must be less than or equal to %s", Integer.MAX_VALUE);

    this.queuePermits = new AtomicInteger(permits);
    this.asyncSemaphore = new AsyncSemaphore<>(maxConcurrentQueries,
            queryExecutor,
            queueEntry -> {
                QueuedExecution queuedExecution = queueEntry.dequeue();
                if (queuedExecution != null) {
                 //获取queuedExecution启动他
                    queuedExecution.start();
                    return queuedExecution.getCompletionFuture();
                }
                return Futures.immediateFuture(null);
            });
}

---

QueuedExecution.start()方法

public void start()
{
      // Only execute if the query is not already completed (e.g. cancelled)
      if (listenableFuture.isDone()) {
          return;
      }
      if (nextQueues.isEmpty()) {
          executor.execute(() -> {
              try (SetThreadName ignored = new SetThreadName("Query-%s", queryExecution.getQueryId())) {
               //继续启动,异步的执行的是SqlQueryExecution的start方法
                  queryExecution.start();
              }
          });
      }
      else {
          nextQueues.get(0).enqueue(new QueuedExecution(queryExecution, nextQueues.subList(1, nextQueues.size()), executor, listenableFuture));
      }
  }

---

SqlQueryExecution.start()方法

public void start()
{
    try (SetThreadName ignored = new SetThreadName("Query-%s", stateMachine.getQueryId())) {
        try {
            // transition to planning
            if (!stateMachine.transitionToPlanning()) {
                // query already started or finished
                return;
            }

            // 分析查询
            PlanRoot plan = analyzeQuery();

            metadata.beginQuery(getSession(), plan.getConnectors());

            // plan distribution of query
            planDistribution(plan);

            // transition to starting
            if (!stateMachine.transitionToStarting()) {
                // query already started or finished
                return;
            }

            // if query is not finished, start the scheduler, otherwise cancel it
            SqlQueryScheduler scheduler = queryScheduler.get();

            if (!stateMachine.isDone()) {
                scheduler.start();
            }
        }
        catch (Throwable e) {
            fail(e);
            throwIfInstanceOf(e, Error.class);
        }
    }
}

 private PlanRoot analyzeQuery()
 {
      try {
          //开始分析
          return doAnalyzeQuery();
      }
      catch (*Error e) {
          throw new PrestoException(NOT_SUPPORTED, "statement is too large (stack overflow during analysis)", e);
      }
  }

private PlanRoot doAnalyzeQuery()
{
    // time analysis phase
    long analysisStart = System.nanoTime();

    // analyze query
    Analyzer analyzer = new Analyzer(stateMachine.getSession(), metadata, sqlParser, accessControl, Optional.of(queryExplainer), parameters);
    //分析的关键,statment是sql解析后的对象
    Analysis analysis = analyzer.analyze(statement);

    stateMachine.setUpdateType(analysis.getUpdateType());

    // plan query
    PlanNodeIdAllocator idAllocator = new PlanNodeIdAllocator();
    LogicalPlanner logicalPlanner = new LogicalPlanner(stateMachine.getSession(), planOptimizers, idAllocator, metadata, sqlParser, costCalculator);
    Plan plan = logicalPlanner.plan(analysis);
    queryPlan.set(plan);

    // extract inputs
    List inputs = new InputExtractor(metadata, stateMachine.getSession()).extractInputs(plan.getRoot());
    stateMachine.setInputs(inputs);

    // extract output
    Optional

output = new OutputExtractor().extractOutput(plan.getRoot()); stateMachine.setOutput(output); // fragment the plan SubPlan fragmentedPlan = PlanFragmenter.createSubPlans(stateMachine.getSession(), metadata, plan); stateMachine.setPlan(planFlattener.flatten(fragmentedPlan, stateMachine.getSession())); // record analysis time stateMachine.recordAnalysisTime(analysisStart); boolean explainAnalyze = analysis.getStatement() instanceof Explain && ((Explain) analysis.getStatement()).isAnalyze(); return new PlanRoot(fragmentedPlan, !explainAnalyze, extractConnectors(analysis)); }

---

Analyzer.analyze()方法
执行分析方法栈过多

public Analysis analyze(Statement statement)
{
    return analyze(statement, false);
}

public Analysis analyze(Statement statement, boolean isDescribe)
{
     Statement rewrittenStatement = StatementRewrite.rewrite(session, metadata, sqlParser, queryExplainer, statement, parameters, accessControl);
     Analysis analysis = new Analysis(rewrittenStatement, parameters, isDescribe);
     //构造阶段分析
     StatementAnalyzer analyzer = new StatementAnalyzer(analysis, metadata, sqlParser, accessControl, session);
     //又开始分析
     analyzer.analyze(rewrittenStatement, Optional.empty());
     return analysis;
 }

---

StatementAnalyzer.analyze()方法
Statement是Node的子类故Node不是一个机器节点,只是一个QueryStatement

public Scope analyze(Node node, Optional outerQueryScope)
{
     return new Visitor(outerQueryScope).process(node, Optional.empty());
 }

public Scope process(Node node, Optional scope)
{
    Scope returnScope = super.process(node, scope);
    checkState(returnScope.getOuterQueryParent().equals(outerQueryScope), "result scope should have outer query scope equal with parameter outer query scope");
    if (scope.isPresent()) {
        checkState(hasScopeAsLocalParent(returnScope, scope.get()), "return scope should have context scope as one of ancestors");
    }
    return returnScope;
}

---

AsVisitor.process

public R process(Node node, @Nullable C context)
{
    return node.accept(this, context);
}

---

Query.accept

public  R accept(AstVisitor visitor, C context)
{
    return visitor.visitQuery(this, context);
}

---

StatementAnalyzer.visitQuery()方法

protected Scope visitQuery(Query node, Optional scope)
{
    //分析Sql的查询范围,如想详细了解可进入到该方法中查看
    Scope withScope = analyzeWith(node, scope);
    //开始分析querybody
    Scope queryBodyScope = process(node.getQueryBody(), withScope);
    if (node.getOrderBy().isPresent()) {
        analyzeOrderBy(node, queryBodyScope);
    }
    else {
        analysis.setOrderByExpressions(node, emptyList());
    }

    // Input fields == Output fields
    analysis.setOutputExpressions(node, descriptorToFields(queryBodyScope));

    Scope queryScope = Scope.builder()
            .withParent(withScope)
            .withRelationType(RelationId.of(node), queryBodyScope.getRelationType())
            .build();
    analysis.setScope(node, queryScope);
    return queryScope;
}

private Scope process(Node node, Scope scope)
{
     return process(node, Optional.of(scope));
 }

public Scope process(Node node, Optional scope)
{
    //调用父类的分析处理中间过程过多,忽略中间阶段直接到具体逻辑
     Scope returnScope = super.process(node, scope);
     checkState(returnScope.getOuterQueryParent().equals(outerQueryScope), "result scope should have outer query scope equal with parameter outer query scope");
     if (scope.isPresent()) {
         checkState(hasScopeAsLocalParent(returnScope, scope.get()), "return scope should have context scope as one of ancestors");
     }
     return returnScope;
 }

访问规范sql是否合规

protected Scope visitQuerySpecification(QuerySpecification node, Optional scope)
{
    if (SystemSessionProperties.isLegacyOrderByEnabled(session)) {
        return legacyVisitQuerySpecification(node, scope);
    }

    // TODO: extract candidate names from SELECT, WHERE, HAVING, GROUP BY and ORDER BY expressions
    // to pass down to analyzeFrom
    //开始分析NodeBody
    Scope sourceScope = analyzeFrom(node, scope);

    node.getWhere().ifPresent(where -> analyzeWhere(node, sourceScope, where));

    List outputExpressions = analyzeSelect(node, sourceScope);
    List> groupByExpressions = analyzeGroupBy(node, sourceScope, outputExpressions);
    analyzeHaving(node, sourceScope);

    Scope outputScope = computeAndAssignOutputScope(node, scope, sourceScope);

    List orderByExpressions = emptyList();
    Optional orderByScope = Optional.empty();
    if (node.getOrderBy().isPresent()) {
        orderByScope = Optional.of(computeAndAssignOrderByScope(node.getOrderBy().get(), sourceScope, outputScope));
        orderByExpressions = analyzeOrderBy(node, orderByScope.get(), outputExpressions);
    }
    else {
        analysis.setOrderByExpressions(node, emptyList());
    }

    List sourceExpressions = new ArrayList<>();
    sourceExpressions.addAll(outputExpressions);
    node.getHaving().ifPresent(sourceExpressions::add);

    analyzeGroupingOperations(node, sourceExpressions, orderByExpressions);
    List aggregations = analyzeAggregations(node, sourceScope, orderByScope, groupByExpressions, sourceExpressions, orderByExpressions);
    analyzeWindowFunctions(node, outputExpressions, orderByExpressions);

    if (!groupByExpressions.isEmpty() && node.getOrderBy().isPresent()) {
        // Create a different scope for ORDER BY expressions when aggregation is present.
        // This is because planner requires scope in order to resolve names against fields.
        // Original ORDER BY scope "sees" FROM query fields. However, during planning
        // and when aggregation is present, ORDER BY expressions should only be resolvable against
        // output scope, group by expressions and aggregation expressions.
        computeAndAssignOrderByScopeWithAggregation(node.getOrderBy().get(), sourceScope, outputScope, aggregations, groupByExpressions, analysis.getGroupingOperations(node));
    }

    return outputScope;
}

private Scope analyzeFrom(QuerySpecification node, Optional scope)
{
    if (node.getFrom().isPresent()) {
        //继续向里面进入忽略中间的进入到Table类
        return process(node.getFrom().get(), scope);
    }

    return createScope(scope);
}

---

com.facebook.presto.sql.tree.Table.accept()方法 Table继承自QueryBody

public  R accept(AstVisitor visitor, C context)
{
     //访问表
     return visitor.visitTable(this, context);
 }

---

StatementAnalyzer.visitTable()方法

//由于这个方法过长,只留了关键方法,这里是获取metadata获取到table的句柄
Optional tableHandle = metadata.getTableHandle(session, name);
if (!tableHandle.isPresent()) {
     if (!metadata.getCatalogHandle(session, name.getCatalogName()).isPresent()) {
         throw new SemanticException(MISSING_CATALOG, table, "Catalog %s does not exist", name.getCatalogName());
     }
     if (!metadata.schemaExists(session, new CatalogSchemaName(name.getCatalogName(), name.getSchemaName()))) {
         throw new SemanticException(MISSING_SCHEMA, table, "Schema %s does not exist", name.getSchemaName());
     }
     throw new SemanticException(MISSING_TABLE, table, "Table %s does not exist", name);
 }

---

MetadataManager.getTableHandle()方法
获取表句柄

public Optional getTableHandle(Session session, QualifiedObjectName table)
{
    requireNonNull(table, "table is null");

    Optional catalog = getOptionalCatalogMetadata(session, table.getCatalogName());
    if (catalog.isPresent()) {
        CatalogMetadata catalogMetadata = catalog.get();
        ConnectorId connectorId = catalogMetadata.getConnectorId(table);
        ConnectorMetadata metadata = catalogMetadata.getMetadataFor(connectorId);
        //创建对应的Connector
        ConnectorTableHandle tableHandle = metadata.getTableHandle(session.toConnectorSession(connectorId), table.asSchemaTableName());
        if (tableHandle != null) {
            return Optional.of(new TableHandle(connectorId, tableHandle));
        }
    }
    return Optional.empty();
}