oracle.kv.impl.query.compiler.Distributer Maven / Gradle / Ivy
Go to download
NoSQL Database Server - supplies build and runtime support for the server (store) side of the Oracle NoSQL Database.
/*-
* Copyright (C) 2011, 2018 Oracle and/or its affiliates. All rights reserved.
*
* This file was distributed by Oracle as part of a version of Oracle NoSQL
* Database made available at:
*
* http://www.oracle.com/technetwork/database/database-technologies/nosqldb/downloads/index.html
*
* Please see the LICENSE file included in the top-level directory of the
* appropriate version of Oracle NoSQL Database for a copy of the license and
* additional information.
*/
package oracle.kv.impl.query.compiler;
import java.util.ArrayList;
import oracle.kv.impl.api.query.PreparedStatementImpl.DistributionKind;
import oracle.kv.impl.query.QueryStateException;
import oracle.kv.impl.query.compiler.FunctionLib.FuncCode;
import oracle.kv.impl.query.compiler.Expr.ExprKind;
/**
* Distributer is responsible for creating ExprReceive exprs in the exprs graph
* and placing them at their appropriate position within the graph.
*
* It implements an algorithm which traverses the exprs graph looking for
* ExprBaseTable nodes. For each ExprBaseTable it finds, it creates an
* ExprReceive and places it right above the associated ExprBaseTable. It
* then tries to pull-up the ExprReceive as far as it can go.
*
* For now, a query can have at most one ExprBaseTable, which if present, is a
* child of the single ExprSFW in the query. As a result, the work done by the
* Distributer is very simple.
*/
class Distributer extends ExprVisitor {
QueryControlBlock theQCB;
FunctionLib theFuncLib = CompilerAPI.getFuncLib();
private final ExprWalker theWalker;
Distributer(QueryControlBlock qcb) {
theQCB = qcb;
theWalker = new ExprWalker(this, false/*allocateChildrenIter*/);
}
void distributeQuery() {
//System.out.println(theQCB.getRootExpr().display());
theWalker.walk(theQCB.getRootExpr());
}
@Override
void exit(ExprUpdateRow upd) {
ExprReceive recv = (ExprReceive)upd.getInput();
upd.setArg(0, recv.getInput(), false/*destroy*/);
upd.replace(recv, false);
recv.setInput(upd, false/*destroy*/);
assert(recv.getDistributionKind() == DistributionKind.SINGLE_PARTITION);
}
@Override
void exit(ExprBaseTable e) {
ExprReceive recv = new ExprReceive(theQCB, theQCB.getInitSctx());
e.replace(recv, false);
recv.setInput(e, false/*destroy*/);
recv.setEliminateIndexDups(e.getEliminateIndexDups());
recv.setIsUpdate(e.getIsUpdate());
}
@Override
boolean enter(ExprSFW sfw) {
theWalker.walk(sfw.getDomainExpr(0));
if (sfw.getDomainExpr(0).getKind() != ExprKind.RECEIVE) {
return false;
}
/*
* Pull the receive expr above the SFW expr
*/
ExprReceive rcv = (ExprReceive)sfw.getDomainExpr(0);
sfw.setDomainExpr(0, rcv.getInput(), false/*destroy*/);
sfw.replace(rcv, false);
rcv.setInput(sfw, false/*destroy*/);
Expr offset = sfw.getOffset();
Expr limit = sfw.getLimit();
boolean hasSort = (sfw.getNumSortExprs() != 0);
boolean hasGroupBy = sfw.hasGroupBy();
boolean hasOffset = (offset != null);
boolean hasLimit = (limit != null);
boolean eliminateIndexDups = rcv.getEliminateIndexDups();
boolean isSinglePartition =
(rcv.getDistributionKind() == DistributionKind.SINGLE_PARTITION);
theQCB.setHasSort(hasSort);
theQCB.setHasGroupBy(hasGroupBy);
theQCB.setHasGroupByExpr(sfw.getNumGroupByExprs() > 0);
theQCB.setGroupByExprCompleteShardKey(sfw.getGroupByExprCompleteShardKey());
/* TODO: remove this when dup elim is always done at the server */
if (sfw.getNumFroms() > 1) {
eliminateIndexDups = false;
rcv.setEliminateIndexDups(false);
}
/*
* If it is a single-partition query, sfw will be sent as-is to the
* server and there is no need to add anything in the client-side plan.
*/
if (isSinglePartition) {
return false;
}
if (!hasSort &&
!hasGroupBy &&
!eliminateIndexDups &&
!(hasOffset || hasLimit)) {
return false;
}
/*
* We are here because either (a) the SFW has order/group-by, or (b)
* it has offset-limit and it's not a single-partition query, or (c)
* we must do elimination of index duplicates.
*
* In all cases, we create a new SFW expr on top of the RCV expr,
* using the RCV expr as the FROM expr. This new SFW will be executed
* at the client, whereas the original SFW will be sent to the server.
*
* If the SFW expr has sort do the following:
* - Add the sort exprs in the SELECT clause of the server SFW, if not
* there already.
* - Add to the receive expr the positions of the sort exprs within
* the above SELECT clause; also add the sort specs.
* - Add to the client SELECT clause the fields that correspond to the
* fields of the original SFW (before the addition of the sort exprs).
*
* If the SFW expr has grouping do the following:
* - Add to the client SELECT clause the fields that correspond to the
* grouping exprs of the original SFW.
* - Add to the client SELECT clause appropriate aggregate functions
* to re-aggregate the partial aggregates coming from the RNs.
* - Make the receive expr a sorting one by adding as the sort
* positions the positions of the grouping exprs within the SELECT
* clause.
*
* If the SFW expr has offset-limit and the query is not single-
* partition, do the following:
* - Add the offset and limit to the client SFW. The offset-limit will
* essentially be done at the client.
* - Remove the offset from the server SFW,
* - Change the limit in the server SFW to be the sum of the original
* limit plus the original offset. This is the maximum number of
* results that any RN will need to compute and send to the client.
*
* If dup elimination is needed:
* - Add in the SELECT clause of the server SFW exprs to retrieve the
* prim key column, if not there already.
* - Add to the receive expr the positions of the prim key exprs within
* the above SELECT clause.
* - Add to the client SELECT clause the fields that correspond to the
* fields of the original SFW.
*/
int numFields = sfw.getNumFields();
boolean constructsRecord = sfw.getConstructsRecord();
if (hasSort) {
int[] sortExprPositions = sfw.addSortExprsToSelect();
rcv.addSort(sortExprPositions, sfw.getSortSpecs());
sfw.setDoNullOnEmpty(false);
}
if (eliminateIndexDups) {
int[] primKeyPositions = sfw.addPrimKeyToSelect();
rcv.addPrimKeyPositions(primKeyPositions);
}
/*
* If we didn't add any exprs in the server-side sfw and there is no
* grouping/offset/limit, we are done (no need to add a client-side
* SFW).
*/
if (numFields == sfw.getNumFields() &&
!hasGroupBy &&
!hasOffset &&
!hasLimit) {
return false;
}
ExprSFW clientSFW =
new ExprSFW(theQCB, sfw.getSctx(), sfw.getLocation());
rcv.replace(clientSFW, false);
ExprVar fromVar =
clientSFW.createFromVar(rcv, theQCB.createInternalVarName("from"));
ArrayList fieldExprs = new ArrayList(numFields);
ArrayList fieldNames = new ArrayList(numFields);
if (sfw.getConstructsRecord()) {
for (int i = 0; i < numFields; ++i) {
Expr clientFieldExpr =
new ExprFieldStep(theQCB,
sfw.getSctx(),
sfw.getFieldExpr(i).getLocation(),
fromVar,
sfw.getFieldName(i));
if (hasGroupBy && !sfw.isGroupingField(i)) {
ExprFuncCall aggrExpr = (ExprFuncCall)sfw.getFieldExpr(i);
clientFieldExpr = getRegroupingExpr(aggrExpr,
clientFieldExpr);
}
fieldExprs.add(clientFieldExpr);
fieldNames.add(sfw.getFieldName(i));
}
} else {
assert(numFields == 1);
Expr clientFieldExpr = fromVar;
if (hasGroupBy && !sfw.isGroupingField(0)) {
ExprFuncCall aggrExpr = (ExprFuncCall)sfw.getFieldExpr(0);
clientFieldExpr = getRegroupingExpr(aggrExpr,
clientFieldExpr);
}
fieldExprs.add(clientFieldExpr);
fieldNames.add(sfw.getFieldName(0));
}
clientSFW.addSelectClause(fieldNames, fieldExprs);
clientSFW.setConstructsRecord(constructsRecord);
if (hasGroupBy) {
int numGBExprs = sfw.getNumGroupByExprs();
clientSFW.setNumGroupByExprs(numGBExprs);
if (numGBExprs > 0) {
int[] sortPositions = new int[numGBExprs];
SortSpec[] sortSpecs = new SortSpec[numGBExprs];
for (int i = 0; i < numGBExprs; ++i) {
sortPositions[i] = i;
sortSpecs[i] = new SortSpec(false, false);
}
rcv.addSort(sortPositions, sortSpecs);
}
}
if (hasOffset || hasLimit) {
if (hasLimit && hasOffset) {
sfw.removeOffset(false/*destroy*/);
Expr newLimit = FuncArithOp.createArithExpr(offset, limit, "+");
sfw.setLimit(newLimit, false/*destroy*/);
} else if (hasOffset) {
sfw.removeOffset(false/*destroy*/);
}
clientSFW.addOffsetLimit(offset, limit);
}
return false;
}
private Expr getRegroupingExpr(
ExprFuncCall aggrExpr,
Expr inputExpr) {
Function aggrFunc = aggrExpr.getFunction(null);
switch (aggrFunc.getCode()) {
case FN_COUNT:
case FN_COUNT_STAR:
case FN_COUNT_NUMBERS:
case FN_SUM:
aggrFunc = theFuncLib.getFunc(FuncCode.FN_SUM);
break;
case FN_MIN:
aggrFunc = theFuncLib.getFunc(FuncCode.FN_MIN);
break;
case FN_MAX:
aggrFunc = theFuncLib.getFunc(FuncCode.FN_MAX);
break;
default:
throw new QueryStateException(
"Unknown aggregate function: " + aggrFunc.getCode());
}
return ExprFuncCall.create(theQCB,
aggrExpr.getSctx(),
aggrExpr.getLocation(),
aggrFunc,
inputExpr);
}
}
© 2015 - 2025 Weber Informatics LLC | Privacy Policy