/**
* Process the FileSink operator to generate a MoveTask if necessary.
*
* @param nd current FileSink operator
* @param stack parent operators
* @param opProcCtx
* @param chDir whether the operator should be first output to a tmp dir and then merged to the
* final dir later
* @return the final file name to which the FileSinkOperator should store.
* @throws SemanticException
*/
private String processFS(Node nd, Stack<Node> stack, NodeProcessorCtx opProcCtx, boolean chDir)
throws SemanticException {
// Is it the dummy file sink after the mapjoin
FileSinkOperator fsOp = (FileSinkOperator) nd;
if ((fsOp.getParentOperators().size() == 1)
&& (fsOp.getParentOperators().get(0) instanceof MapJoinOperator)) {
return null;
}
GenMRProcContext ctx = (GenMRProcContext) opProcCtx;
List<FileSinkOperator> seenFSOps = ctx.getSeenFileSinkOps();
if (seenFSOps == null) {
seenFSOps = new ArrayList<FileSinkOperator>();
}
if (!seenFSOps.contains(fsOp)) {
seenFSOps.add(fsOp);
}
ctx.setSeenFileSinkOps(seenFSOps);
Task<? extends Serializable> currTask = ctx.getCurrTask();
// If the directory needs to be changed, send the new directory
String dest = null;
if (chDir) {
dest = fsOp.getConf().getDirName();
// generate the temporary file
// it must be on the same file system as the current destination
ParseContext parseCtx = ctx.getParseCtx();
Context baseCtx = parseCtx.getContext();
String tmpDir = baseCtx.getExternalTmpFileURI((new Path(dest)).toUri());
fsOp.getConf().setDirName(tmpDir);
}
Task<? extends Serializable> mvTask = null;
if (!chDir) {
mvTask = findMoveTask(ctx.getMvTask(), fsOp);
}
Operator<? extends Serializable> currTopOp = ctx.getCurrTopOp();
String currAliasId = ctx.getCurrAliasId();
HashMap<Operator<? extends Serializable>, Task<? extends Serializable>> opTaskMap =
ctx.getOpTaskMap();
List<Operator<? extends Serializable>> seenOps = ctx.getSeenOps();
List<Task<? extends Serializable>> rootTasks = ctx.getRootTasks();
// Set the move task to be dependent on the current task
if (mvTask != null) {
currTask.addDependentTask(mvTask);
}
// In case of multi-table insert, the path to alias mapping is needed for
// all the sources. Since there is no
// reducer, treat it as a plan with null reducer
// If it is a map-only job, the task needs to be processed
if (currTopOp != null) {
Task<? extends Serializable> mapTask = opTaskMap.get(null);
if (mapTask == null) {
assert (!seenOps.contains(currTopOp));
seenOps.add(currTopOp);
GenMapRedUtils.setTaskPlan(
currAliasId, currTopOp, (MapredWork) currTask.getWork(), false, ctx);
opTaskMap.put(null, currTask);
rootTasks.add(currTask);
} else {
if (!seenOps.contains(currTopOp)) {
seenOps.add(currTopOp);
GenMapRedUtils.setTaskPlan(
currAliasId, currTopOp, (MapredWork) mapTask.getWork(), false, ctx);
}
// mapTask and currTask should be merged by and join/union operator
// (e.g., GenMRUnion1j) which has multiple topOps.
assert mapTask == currTask
: "mapTask.id = " + mapTask.getId() + "; currTask.id = " + currTask.getId();
}
return dest;
}
UnionOperator currUnionOp = ctx.getCurrUnionOp();
if (currUnionOp != null) {
opTaskMap.put(null, currTask);
GenMapRedUtils.initUnionPlan(ctx, currTask, false);
return dest;
}
AbstractMapJoinOperator<? extends MapJoinDesc> currMapJoinOp = ctx.getCurrMapJoinOp();
if (currMapJoinOp != null) {
opTaskMap.put(null, currTask);
GenMRMapJoinCtx mjCtx = ctx.getMapJoinCtx(currMapJoinOp);
MapredWork plan = (MapredWork) currTask.getWork();
String taskTmpDir = mjCtx.getTaskTmpDir();
TableDesc tt_desc = mjCtx.getTTDesc();
assert plan.getPathToAliases().get(taskTmpDir) == null;
plan.getPathToAliases().put(taskTmpDir, new ArrayList<String>());
plan.getPathToAliases().get(taskTmpDir).add(taskTmpDir);
plan.getPathToPartitionInfo().put(taskTmpDir, new PartitionDesc(tt_desc, null));
plan.getAliasToWork().put(taskTmpDir, mjCtx.getRootMapJoinOp());
return dest;
}
return dest;
}
/**
* Process the FileSink operator to generate a MoveTask if necessary.
*
* @param fsOp current FileSink operator
* @param stack parent operators
* @param opProcCtx
* @param chDir whether the operator should be first output to a tmp dir and then merged to the
* final dir later
* @return the final file name to which the FileSinkOperator should store.
* @throws SemanticException
*/
private String processFS(
FileSinkOperator fsOp, Stack<Node> stack, NodeProcessorCtx opProcCtx, boolean chDir)
throws SemanticException {
GenMRProcContext ctx = (GenMRProcContext) opProcCtx;
List<FileSinkOperator> seenFSOps = ctx.getSeenFileSinkOps();
if (seenFSOps == null) {
seenFSOps = new ArrayList<FileSinkOperator>();
}
if (!seenFSOps.contains(fsOp)) {
seenFSOps.add(fsOp);
}
ctx.setSeenFileSinkOps(seenFSOps);
Task<? extends Serializable> currTask = ctx.getCurrTask();
// If the directory needs to be changed, send the new directory
String dest = null;
if (chDir) {
dest = fsOp.getConf().getFinalDirName();
// generate the temporary file
// it must be on the same file system as the current destination
ParseContext parseCtx = ctx.getParseCtx();
Context baseCtx = parseCtx.getContext();
String tmpDir = baseCtx.getExternalTmpFileURI((new Path(dest)).toUri());
FileSinkDesc fileSinkDesc = fsOp.getConf();
// Change all the linked file sink descriptors
if (fileSinkDesc.isLinkedFileSink()) {
for (FileSinkDesc fsConf : fileSinkDesc.getLinkedFileSinkDesc()) {
String fileName = Utilities.getFileNameFromDirName(fsConf.getDirName());
fsConf.setParentDir(tmpDir);
fsConf.setDirName(tmpDir + Path.SEPARATOR + fileName);
}
} else {
fileSinkDesc.setDirName(tmpDir);
}
}
Task<MoveWork> mvTask = null;
if (!chDir) {
mvTask = findMoveTask(ctx.getMvTask(), fsOp);
}
Operator<? extends OperatorDesc> currTopOp = ctx.getCurrTopOp();
String currAliasId = ctx.getCurrAliasId();
HashMap<Operator<? extends OperatorDesc>, Task<? extends Serializable>> opTaskMap =
ctx.getOpTaskMap();
List<Operator<? extends OperatorDesc>> seenOps = ctx.getSeenOps();
List<Task<? extends Serializable>> rootTasks = ctx.getRootTasks();
// Set the move task to be dependent on the current task
if (mvTask != null) {
addDependentMoveTasks(ctx, mvTask, currTask);
}
// In case of multi-table insert, the path to alias mapping is needed for
// all the sources. Since there is no
// reducer, treat it as a plan with null reducer
// If it is a map-only job, the task needs to be processed
if (currTopOp != null) {
Task<? extends Serializable> mapTask = opTaskMap.get(null);
if (mapTask == null) {
if (!seenOps.contains(currTopOp)) {
seenOps.add(currTopOp);
GenMapRedUtils.setTaskPlan(
currAliasId, currTopOp, (MapredWork) currTask.getWork(), false, ctx);
}
opTaskMap.put(null, currTask);
if (!rootTasks.contains(currTask)
&& (currTask.getParentTasks() == null || currTask.getParentTasks().isEmpty())) {
rootTasks.add(currTask);
}
} else {
if (!seenOps.contains(currTopOp)) {
seenOps.add(currTopOp);
GenMapRedUtils.setTaskPlan(
currAliasId, currTopOp, (MapredWork) mapTask.getWork(), false, ctx);
} else {
UnionOperator currUnionOp = ctx.getCurrUnionOp();
if (currUnionOp != null) {
opTaskMap.put(null, currTask);
ctx.setCurrTopOp(null);
GenMapRedUtils.initUnionPlan(ctx, currUnionOp, currTask, false);
return dest;
}
}
// mapTask and currTask should be merged by and join/union operator
// (e.g., GenMRUnion1) which has multiple topOps.
// assert mapTask == currTask : "mapTask.id = " + mapTask.getId()
// + "; currTask.id = " + currTask.getId();
}
return dest;
}
UnionOperator currUnionOp = ctx.getCurrUnionOp();
if (currUnionOp != null) {
opTaskMap.put(null, currTask);
GenMapRedUtils.initUnionPlan(ctx, currUnionOp, currTask, false);
return dest;
}
return dest;
}
