Skip to content
This repository has been archived by the owner on Jul 23, 2024. It is now read-only.

Commit

Permalink
HAWQ-1660. Optimize parquet scan when bloom filter enabled.
Browse files Browse the repository at this point in the history
  • Loading branch information
Kuien Liu committed Sep 20, 2018
1 parent c6c7988 commit 2320cf0
Show file tree
Hide file tree
Showing 2 changed files with 155 additions and 55 deletions.
200 changes: 145 additions & 55 deletions src/backend/cdb/cdbparquetrowgroup.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -211,6 +211,33 @@ ParquetRowGroupReader_ScanNextTuple(
TupleTableSlot *slot)
{
Assert(slot);

int natts = slot->tts_tupleDescriptor->natts;
Assert(natts <= tupDesc->natts);

Datum *values = slot_get_values(slot);
bool *nulls = slot_get_isnull(slot);

bool useBloomFilter = false;
int joinKeyCount = 0;
int *joinKeySet = NULL;
if (rfState != NULL && rfState->hasRuntimeFilter && !rfState->stopRuntimeFilter)
{
useBloomFilter = true;

joinKeyCount = list_length(rfState->joinkeys);
Assert(joinKeyCount <= natts);
joinKeySet = palloc(sizeof(int) * joinKeyCount);

ListCell *hk;
int i = 0;
foreach(hk, rfState->joinkeys)
{
AttrNumber attrno = (AttrNumber) lfirst(hk);
joinKeySet[i++] = attrno -1;
}
}

while (rowGroupReader->rowRead < rowGroupReader->rowCount)
{

Expand All @@ -219,12 +246,9 @@ ParquetRowGroupReader_ScanNextTuple(
*/
rowGroupReader->rowRead++;

int natts = slot->tts_tupleDescriptor->natts;
Assert(natts <= tupDesc->natts);

Datum *values = slot_get_values(slot);
bool *nulls = slot_get_isnull(slot);

/*
* Step 1: fetch those columns as hash join keys
*/
int colReaderIndex = 0;
for (int i = 0; i < natts; i++)
{
Expand All @@ -233,78 +257,50 @@ ParquetRowGroupReader_ScanNextTuple(
nulls[i] = true;
continue;
}
ParquetColumnReader *nextReader =
&rowGroupReader->columnReaders[colReaderIndex];
int hawqTypeID = tupDesc->attrs[i]->atttypid;

if (hawqAttrToParquetColNum[i] == 1)
bool isJoinKeyColumn = false;
for (int j = 0; j < joinKeyCount; j++)
{
ParquetColumnReader_readValue(nextReader, &values[i], &nulls[i],
hawqTypeID);
}
else
{
/*
* Because there are some memory reused inside the whole column reader, so need
* to switch the context from PerTupleContext to rowgroup->context
*/
MemoryContext oldContext = MemoryContextSwitchTo(
rowGroupReader->memoryContext);

switch (hawqTypeID) {
case HAWQ_TYPE_POINT:
ParquetColumnReader_readPoint(nextReader, &values[i], &nulls[i]);
break;
case HAWQ_TYPE_PATH:
ParquetColumnReader_readPATH(nextReader, &values[i], &nulls[i]);
break;
case HAWQ_TYPE_LSEG:
ParquetColumnReader_readLSEG(nextReader, &values[i], &nulls[i]);
break;
case HAWQ_TYPE_BOX:
ParquetColumnReader_readBOX(nextReader, &values[i], &nulls[i]);
break;
case HAWQ_TYPE_CIRCLE:
ParquetColumnReader_readCIRCLE(nextReader, &values[i], &nulls[i]);
break;
case HAWQ_TYPE_POLYGON:
ParquetColumnReader_readPOLYGON(nextReader, &values[i], &nulls[i]);
break;
default:
/* TODO array type */
/* TODO UDT */
Insist(false);
if (joinKeySet[j] == i)
{
isJoinKeyColumn = true;
break;
}
}

MemoryContextSwitchTo(oldContext);
if (isJoinKeyColumn)
{
ParquetColumnReader *nextReader =
&rowGroupReader->columnReaders[colReaderIndex];
int hawqTypeID = tupDesc->attrs[i]->atttypid;

ParquetRowGroupReader_ScanOneAttribute(
rowGroupReader, hawqAttrToParquetColNum[i],
nextReader, &values[i], &nulls[i], hawqTypeID);
}

colReaderIndex += hawqAttrToParquetColNum[i];
}

if (rfState != NULL && rfState->hasRuntimeFilter
&& !rfState->stopRuntimeFilter)
/*
* Step 2: skip following columns decoding if bloomfilter is mismatched
*/
if (useBloomFilter)
{
Assert(rfState->bloomfilter != NULL);
uint32_t hashkey = 0;
ListCell *hk;
int i = 0;
foreach(hk, rfState->joinkeys)
for (int i = 0; i < joinKeyCount; i++)
{
AttrNumber attrno = (AttrNumber) lfirst(hk);
Datum keyval;
uint32 hkey;

/* rotate hashkey left 1 bit at each step */
hashkey = (hashkey << 1) | ((hashkey & 0x80000000) ? 1 : 0);
keyval = values[attrno - 1];
keyval = values[joinKeySet[i]];

/* Evaluate expression */
hkey = DatumGetUInt32(
FunctionCall1(&rfState->hashfunctions[i], keyval));
hashkey ^= hkey;
i++;
}

if (!FindBloomFilter(rfState->bloomfilter, hashkey))
Expand All @@ -313,6 +309,43 @@ ParquetRowGroupReader_ScanNextTuple(
}
}

/*
* Step 3: fetch those columns not in hash join keys
*/
colReaderIndex = 0;
for (int i = 0; i < natts; i++)
{
// it is not expensive to do twice
if (projs[i] == false)
{
nulls[i] = true;
continue;
}

bool isJoinKeyColumn = false;
for (int j = 0; j < joinKeyCount; j++)
{
if (joinKeySet[j] == i)
{
isJoinKeyColumn = true;
break;
}
}

if (!isJoinKeyColumn)
{
ParquetColumnReader *nextReader =
&rowGroupReader->columnReaders[colReaderIndex];
int hawqTypeID = tupDesc->attrs[i]->atttypid;

ParquetRowGroupReader_ScanOneAttribute(
rowGroupReader, hawqAttrToParquetColNum[i],
nextReader, &values[i], &nulls[i], hawqTypeID);
}

colReaderIndex += hawqAttrToParquetColNum[i];
}

/*construct tuple, and return back*/
TupSetVirtualTupleNValid(slot, natts);
return true;
Expand All @@ -322,6 +355,63 @@ ParquetRowGroupReader_ScanNextTuple(
return false;
}

/*
* Get one attribute of a tuple from current row group into slot.
*
* Similar to ParquetColumnReader_readValue() but consider more hawq types.
*/
void
ParquetRowGroupReader_ScanOneAttribute(
ParquetRowGroupReader *rowGroupReader,
int colChildNum, // hawqAttrToParquetColNum
ParquetColumnReader *columnReader,
Datum *value,
bool *null,
int hawqTypeID)
{
if (colChildNum == 1)
{
ParquetColumnReader_readValue(columnReader, value, null, hawqTypeID);
}
else
{
/*
* Because there are some memory reused inside the whole column reader, so need
* to switch the context from PerTupleContext to rowgroup->context
*/
MemoryContext oldContext = MemoryContextSwitchTo(
rowGroupReader->memoryContext);

switch (hawqTypeID) {
case HAWQ_TYPE_POINT:
ParquetColumnReader_readPoint(columnReader, value, null);
break;
case HAWQ_TYPE_PATH:
ParquetColumnReader_readPATH(columnReader, value, null);
break;
case HAWQ_TYPE_LSEG:
ParquetColumnReader_readLSEG(columnReader, value, null);
break;
case HAWQ_TYPE_BOX:
ParquetColumnReader_readBOX(columnReader, value, null);
break;
case HAWQ_TYPE_CIRCLE:
ParquetColumnReader_readCIRCLE(columnReader, value, null);
break;
case HAWQ_TYPE_POLYGON:
ParquetColumnReader_readPOLYGON(columnReader, value, null);
break;
default:
/* TODO array type */
/* TODO UDT */
Insist(false);
break;
}

MemoryContextSwitchTo(oldContext);
}
}

/**
* finish scanning row group, but keeping the structure palloced
*/
Expand Down
10 changes: 10 additions & 0 deletions src/include/cdb/cdbparquetrowgroup.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -75,6 +75,16 @@ ParquetRowGroupReader_ScanNextTuple(
RuntimeFilterState *rfState,
TupleTableSlot *slot);

/* Get one attribute of a tuple from current row group*/
void
ParquetRowGroupReader_ScanOneAttribute(
ParquetRowGroupReader *rowGroupReader,
int colChildNum,
ParquetColumnReader *columnReader,
Datum *value,
bool *null,
int hawqTypeID);

/* Finish scanning current row group*/
void
ParquetRowGroupReader_FinishedScanRowGroup(
Expand Down

0 comments on commit 2320cf0

Please sign in to comment.